Towards Archival Engagement With AI: An exploratory study using historical photographs

Towards Archival Engagement With AI: An exploratory study using historical photographs | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Towards Archival Engagement With AI: An exploratory study using historical photographs Farzaneh Talebhaghighi This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8615665/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract As archives and cultural heritage institutions experiment with immersive and semi-immersive technologies, generative artificial intelligence is increasingly proposed as a means of expanding access to archival materials. This exploratory case study evaluates three commercial artificial intelligence tools: Skybox, LTX Studio, and DeeVid by applying them to thirty-three historical photographs from the Arizona Historical Society’s Tucson 250+: Where We Live, What We Do, and Who We Are digital exhibition. Using a rubric-based evaluation framework, the study assesses fidelity, coherence, authenticity, engagement, and representational accuracy. The findings reveal systematic differences across platforms: Skybox frequently produces historically inaccurate outputs in scenes involving people and identity markers; LTX Studio demonstrates greater visual consistency but introduces subtle representational distortions; and DeeVid generates stable but limited transformations. The results indicate that generative tools currently require substantial human oversight to align with archival standards and ethical responsibilities. Also, the results point to the importance of conceptually informed and logic-based approaches in the development of artificial intelligence tools for archival applications. Archival practice generative artificial intelligence immersive technologies cultural heritage archival photographs representation and bias AI ethics metadata and description semi-immersive media digital archives Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Introduction As museums and archives experiment with new technologies to support learning and visitor engagement, immersive and semi-immersive media have become increasingly prominent. Research in museum and heritage contexts suggests that interactive and immersive environments can deepen user engagement, emotional involvement, and learning outcomes (Spadoni et al., 2023; Xu & Pan, 2024). However, archival studies also show the practical constraints faced by institutions, including limited staff time, budgets, and technological capacity (Chanakira et al., 2023; Shehade & Stylianou-Lambert, 2020). These tensions form the backdrop for this project, which examines whether AI-driven tools might help reduce production burdens while still enabling audiences to interact with archival photographs in more compelling ways. This study draws on archival photographs from three collections included in the Arizona Historical Society (AHS)’s “Tucson 250+: Where We Live, What We Do, and Who We Are” online exhibition [1] , to evaluate how contemporary AI tools interpret and transform historical visual materials. My original intention was to generate fully interactive VR scenes from these photographs. However, early experiments revealed significant limitations in current AI capacities. These constraints required a methodological pivot: rather than creating full VR environments, the project shifted toward using historical photos to produce images with enhanced immersive affect, such as 360-degree scenes, image-to-video transformations, and colorized or augmented representations. These approaches still afford semi-immersive engagement but are more technically feasible and more closely aligned with the capabilities of existing AI systems. This pivot reflects a core insight that emerged through practice: while immersive technologies hold promise for cultural heritage engagement, the process of generating VR-ready content remains expensive, time-consuming, and technically complex (GadAllah, 2020; Shehade & Stylianou-Lambert, 2020). AI may help streamline parts of production, yet generative models also introduce new risks, such as hallucination, cultural bias, inaccuracies, and the erasure or distortion of marginalized identities (Hazan, 2023; Kotsiubivska et al., 2024). Therefore, a second goal of the current study is to understand how AI tools behave when applied to archival materials. This kind of analysis is essential for increasing dialogue between archivists and developers on the risks and opportunities of AI in archives, libraries, and museums. By systematically testing AI-generated outputs across 33 archival images, scoring the results with a rubric, and analyzing patterns in the outputs produced by specific prompt strategies, this research examines how contemporary AI can support archival interpretation, what distortions and biases emerge, and what this means for the future of AI-assisted cultural heritage access. The goal is not simply to evaluate the tools but to understand how AI intersects with archival description, metadata, visual accuracy, and the ethical considerations embedded in archival practice. In doing so, the project contributes to broader conversations in archival theory about the constructed nature of records (Harris, 2007; Ketelaar, 2000, 2010), the interpretive power embedded in description (Zhang, 2012), and the responsibilities institutions bear when mediating historical memory through emerging technologies. Background Immersive and AI driven systems have become central to recent work in cultural heritage, archives, and museums. Large scale reviews and case studies show that AR and VR have been used to increase engagement, learning, and interpretive richness across museums, archives, and heritage sites, often through interactive storytelling, spatial reconstructions, and adaptive content delivery (Boboc et al., 2022 ; Colavizza et al., 2022 ; Kotsiubivska et al., 2024 ; Shinde et al., 2024 ; Xu & Pan, 2024 ). These projects usually rely on custom VR pipelines, game engines, and institution specific infrastructure, which makes them powerful but also costly and difficult to replicate in small or under-resourced archives. Within immersive cultural heritage specifically, recent work has begun to integrate artificial intelligence as part of the interpretive layer, rather than as the generator of the visual scene. For example, Farella’s dissertation on immersive exploration of cultural heritage sites uses an AI-driven query framework inside a photogrammetry-based VR environment so that visitors can ask context-dependent questions and receive tailored explanations (Farella, 2025 ). Similarly, AI has been used to generate narration, branching stories, or gameplay elements inside museum experiences, augmenting engagement and learning without altering the underlying historical images or 3D models (Hettmann et al., 2023 ). This strand of work shows how AI can enrich meaning in immersive spaces, but it does not focus on how AI might be used to fabricate the spatial environment itself. A second cluster of research addresses user acceptance, immersion quality, and the risks and benefits of adopting digital heritage technologies. Studies on the adoption of digital intangible cultural heritage, for example, highlight how willingness to use immersive tools is shaped by performance expectations, social influence, hedonic motivation, and perceived immersion, but also by concerns about distraction, discomfort, and authenticity (Ye et al., 2025 ). In parallel, Pope’s work with archaeologists documents excitement about VR for storytelling and preservation alongside deep worries about accuracy, cultural sensitivity, and the danger that audiences may treat virtual reconstructions as historically true (Pope, 2025 ). From an archival perspective, this aligns with broader critiques of generative AI in museums and heritage that warn about issues of authorship, bias, and “doppelgänger” versions of culture that can be mistaken for the real thing (Hazan, 2023 ). Other projects demonstrate how immersive technologies can be integrated into archival practice without generative reconstruction. Colegrove and Mikel’s “Radical Inclusion” project uses consumer grade 360 degree video to document ephemeral events and fold them into archival workflows, with attention to metadata and sustainable access (Colegrove & Mikel, 2018 ). Here, immersive media functions as a recording strategy rather than a speculative reimagining. My study departs from this documentation-oriented model by examining what happens when AI is used to transform archival photographs into semi-immersive 3D scenes, 360 style images, and short AI-generated videos. The literature shows that immersive heritage systems and AI-enhanced interpretation can support engagement and learning, and that users’ adoption of such systems depends strongly on immersion quality, comfort, and perceived authenticity. At the same time, there is very little empirical work that systematically evaluates off-the-shelf generative tools, especially in small archival settings that rely on subscription-based platforms rather than custom pipelines. Existing studies rarely analyze patterns of AI error across different content types, or connect those patterns to archival concerns about representation, authority, and the future use of AI-generated content in public access systems. This gap motivates the present study, which focuses on identifying “AI logic” patterns and mismatches in the outputs of three contemporary tools and considers what these patterns might mean for archivists who may want to experiment with AI-assisted immersive access in practice. Research Questions This exploratory study focuses on examining how contemporary artificial intelligence tools transform archival photographs into semi-immersive outputs such as 360-degree views, image-based videos, and enhanced image representations. The broader goal is to evaluate the strengths, limitations, costs, technical barriers, and potential uses of these tools within archival practice. A secondary, longer-term vision for this research is to assess the feasibility of embedding AI-assisted immersive features directly into archival websites, enabling users to generate VR-like experiences from archival materials on demand. Guided by these aims, the study addresses the following research questions: What kinds of errors, confusions, or mismatches appear in AI-generated outputs? How do these errors correspond to specific content types or platforms? What strengths, alignments, or meaningful enhancements appear in AI-generated outputs? What are the costs, time requirements, learning curves, and technical barriers of using these tools with archival photo collections? Which AI-generated outputs best support engagement with history? For each AI tool, what combinations of images and prompts contribute to successful outcomes, and under what conditions do outputs degrade or fail? These questions allow the study to evaluate AI tools not only as creative engines, but also as potential components of future archival access systems, where ethical representation, accuracy, and public engagement must remain central. The final question originally planned for this study, “How can archivists balance imagination, annotation, and AI augmentation with fidelity to historical evidence?”, will be addressed in the researcher’s next project, which will involve qualitative interviews with archivists and cultural heritage professionals. Purpose and Significance The purpose of this study is to evaluate how three contemporary AI tools (Skybox, LTX Studio, and DeeVid) interpret and transform archival photographs into semi-immersive outputs such as 360-degree views, image-based videos, and enhanced representations. By systematically testing these tools on images from the Arizona Historical Society (AHS)’s Tucson 250 + collections, the study identifies patterns of accuracy, distortion, bias, and creative augmentation, as well as the technical and conceptual limits of each platform. The significance of this work lies in its direct relevance to archival practice. As archives begin to experiment with AI-assisted access and interpretation, they require practical evidence about what these tools can and cannot reliably produce. This study contributes insight into fidelity, historical alignment, prompt sensitivity, distortion, and tool behavior across image types, offering archivists an early roadmap for safe experimentation. The findings also highlight the need for future, archivist-centered models designed around archival values, authenticity, and community engagement. Methodology This project unfolded as a practice-based, exploratory case study shaped by both the affordances and the limitations of working with AI tools that were never designed for archival use. From the beginning, my intention was to experiment with AI platforms that promised 3D, video, or immersive scene generation and evaluate whether they could support a lightweight, semi-immersive access layer for archival photographs. The process did not follow a rigid experimental protocol; instead, it evolved iteratively as I confronted real technical barriers, unexpected failures, and surprising tool behaviors. I began by selecting images from the Arizona Historical Society’s (AHS) Tucson 250 + exhibition, ensuring representation across its three subcollections: Where We Live, What We Do, and Who We Are. My goal was to choose images that reflected the thematic diversity of the exhibition: landscapes, street scenes, daily life, buildings, people, and group activities. In the end, to ensure appropriate distribution and coverage of all themes in the exhibition, I selected 11 images from each subcollection, resulting in a total of 33 archival photographs used for the study. Exclusion criteria were adopted to avoid extremely low-resolution images that prevented any meaningful visual interpretation, and duplicates or highly redundant views when a near equivalent was already selected. This project began with a broad exploratory phase designed to map the current landscape of AI tools capable of transforming archival photographs into semi-immersive media. Rather than assuming that a single platform could meet the needs of this case study, I intentionally approached the process with an open survey mindset. The goal was to understand what kinds of outputs were even possible with existing tools, how they differed in behavior, and where their technical limits appeared in practice. In the first stage, I experimented with a wide range of AI platforms offering image enhancement, animation, 3D reconstruction, and scene generation features. These included: Meshy 2 , Runway 3 , Luma AI 4 , Polycam 5 , Tripo 3D 6 , DepthR 7 , Gemini Nano 8 , OpenArt 9 , Pikart 10 , Copilot Image Generator 11 . Each tool was tested with the same guiding question: Can this platform meaningfully transform an archival photograph into a 3D model, a semi-immersive scene, or a dynamic video suitable for public engagement? Throughout this exploratory phase, I evaluated whether the tools could recreate spatial environments, preserve historical details, or even respond coherently to archival descriptions. Most of the tools tested provided interesting prototypes but were not sufficiently consistent, historically faithful, or adapted to the image quality of archival material. Some platforms produced visually appealing results but disregarded historical details. Others could generate 3D reconstructions only when given high-resolution, object-centered images that differed significantly from the archival photographs used in this study. This phase of research clarified that the challenge of my project lay not only generating an immersive scene, but in doing so in a way that respected the constraints of archival images and descriptions. Ultimately, after testing this wider landscape of tools, I selected Skybox 12 , LTX Studio 13 , and DeeVid 14 for further evaluation. Skybox AI by Blockade Labs is an AI-powered tool that generates immersive, 360-degree panoramic images (skyboxes) from simple text prompts. LTX Studio is Lightricks' first entirely AI-powered platform, enabling professionals to visualize and develop any creative concept and DeeVid turns still images into animated video clips, whether you're using a single image or a series of images. All these three platforms offered the most stable semi-immersive outputs, supported text and/or image input, demonstrated clear differences in logic and behavior, and represented three distinct technical approaches (360-generation, video generation, and scene representation). These three tools became the focus of the larger case study and the foundation for the pattern analysis discussed in later sections. Before beginning formal testing, I spent time learning each platform, experimenting with their settings, subscriptions, and limitations. I subscribed to all three tools on a monthly basis, knowing that some features were locked behind paid access. The outputs generated by the three platforms differed according to their respective input structures and technical configurations. LTX Studio produced still-image outputs based on a combination of textual prompts and a single input image, supported by a built-in prompt optimizer. For each photograph, LTX Studio generated three colorized or enhanced image variants derived from the same image and text prompt. These outputs were produced using the FLUX configuration, with no additional stylistic parameters applied, a fixed 16:9 aspect ratio, you can change it to vertical manually, and no specified camera angle or location. DeeVid generated video outputs from single images paired with textual prompts. Two output configurations were tested: Fast V2.0, which produced 5-second videos at 512p resolution without sound effects, and Master V2.0, which produced 8-second videos at 720p resolution with sound effects enabled. SkyBox generated immersive outputs in the form of 360-degree environmental views, based on textual prompts or mixed text–image inputs, and included a prompt optimization feature. These outputs were generated using model 3, with an image size of 8K and a landscape depth map. Outputs were organized by source image, tool, date, and prompt type to allow controlled comparison in an Excel sheet. Prompting, “an input into the AI system to obtain specific results” (“Prompts for AI,” n.d.), became one of the most challenging and revealing parts of the methodology. For each platform, I began by using as a prompt the archival descriptions exactly as written by AHS. When results were poor, I tested two additional variations: an AI-generated prompt (using ChatGPT to create a 600-character Skybox prompt from the archival metadata) and a researcher-crafted prompt where I wrote my own description. Throughout the process, I documented not only the outputs, but also the tool behaviors: crashes, “high demand” messages, failed renders, missing textures, and unexplained changes in model behavior following software updates. These observations became just as important as the outputs themselves, because they revealed how unpredictable these systems currently are for archival use. To evaluate the results, I built a rubric with multiple dimensions: fidelity, completeness, coherence, emotional impact, authenticity, engagement, ease/comfort, and an overall qualitative notes section. I scored each output individually, and only after completing all evaluations, I calculated averages and looked for patterns. I also created a second table scored independently from the archival originals, focusing instead on how each tool handled specific content categories: race, gender, age, animals, buildings, environment, objects, vegetation, and people. Finally, once all outputs were scored, I began analyzing patterns across prompts, image types, and tools. I looked for repeated failures, repeated successes, and the specific words that led to either outcome. These patterns became the basis for my findings and shaped the larger implications for archivists who may want to experiment with AI-assisted access in the future. Findings Across all the outputs generated from 33 archival photographs, clear patterns emerged in how each tool interpreted archival materials, how prompts influenced results, and where distortions or meaningful enhancements appeared. The findings reflect both the numerical scores (1–5, 1 for the lowest score and 5 for the highest score) from the rubric tables and the qualitative insights that surfaced through repeated experimentation. Together, these patterns illustrate not only what the tools can do, but also how their internal logic diverges from archival values and constraints. Table 1 Rubric Table number 1 Results (scores from 1 to 5) Fidelity Completeness Coherence Emotional Impact Authenticity Engagement Ease & Comfort AVG_ARC 1.42 1.85 1.91 2.00 1.45 2.06 5.00 AVG_AI 1.96 2.08 2.16 2.32 1.88 2.28 5.00 AVG_RES 1.94 2.12 2.06 2.12 1.76 2.24 5.00 AVG_DeeVid 4.61 4.58 4.45 4.33 4.58 4.36 5.00 AVG_LTX 3.67 3.88 3.91 4.09 3.82 4.06 5.00 AVG_ARC stands for average score of outputs from SKYBOX with an archival description. AVG_AI stands for average score of outputs from SKYBOX with an AI description. AVG_RES stands for average score of outputs from SKYBOX with the researcher description. AVG_ DeeVid stands for average score of outputs from DeeVid. platform. AVG_LTX stands for average score of outputs from LTX Studio platform. The two evaluation tables reveal distinct but complementary patterns about how each tool performs when generating outputs from archival photographs and textual descriptions. When the focus is on compatibility with the original archival record (Table 1 ), measured through fidelity, completeness, coherence, emotional impact, authenticity, engagement, and ease of use, a clear hierarchy emerges. Skybox performs the weakest across all three prompt types (archival description, AI-generated description, and researcher-written description), particularly in fidelity, authenticity, and completeness. Although ease-of-use scores are high for all tools, Skybox’s inability to recreate historically grounded scenes remains consistent, and the differences between its three prompt types are minimal. In other words, changing the description does not dramatically improve Skybox’s reliability. By contrast, DeeVid and LTX consistently achieve higher scores, with DeeVid showing the strongest overall performance in reconstructing or visually extending the original archival scenes, followed closely by LTX. Table 2 Rubric Table number 2 Results (scores from 1 to 5) landscape People building objects animal race gender age AVG_ARC 1.93 1.76 2.00 1.79 2.00 4.00 3.30 3.63 AVG_AI 2.43 1.75 2.53 1.92 3.00 2.00 3.00 2.83 AVG_RES 2.07 1.92 2.17 2.00 2.75 2.80 3.13 3.50 AVG_DeeVid 4.90 4.27 4.82 4.62 4.75 5.00 4.92 4.80 AVG_LTX 4.27 3.41 4.05 3.62 4.00 4.63 4.25 4.80 The second table shifts the focus from historical accuracy to general performance across content types: landscapes, people, buildings, objects, animals, race, gender, and age. Here again the overall pattern aligns with the first table. Skybox underperforms in nearly all categories involving people, bodies, identity, and social scenes, reflecting its difficulty with differentiation in human subjects. Even when prompts avoided identity terms, Skybox frequently produced distorted or semantically incorrect representations, confirming the lack of internal logic previously observed in qualitative results. LTX and DeeVid both score substantially higher across all categories; however, one LTX anomaly reveals an important limitation: LTX ignored text in the archival descriptions that specified racial identity in the subjects and sometimes introduced an additional person who did not exist in the archival record. This case shows that even high-performing tools are not immune to representational bias. Consistent with earlier observations, DeeVid remains the most semantically reliable, performing strongly across all categories, though its output style is limited to camera motion rather than full reconstruction. Overall, the patterns across both tables show that Skybox’s weaknesses are systematic and rooted in how it interprets (or fails to interpret) descriptive prompts, whereas LTX and DeeVid show higher stability and accuracy but still present risks in racially sensitive contexts. These findings also highlight a methodological insight: textual descriptions of the past, whether through archival metadata, AI-generated prompts, or researcher prompts, are a central driver of how AI tools interpret or misinterpret history. The language given to these systems shapes the internal logic they apply, for better or for worse, and this dependency is visible across every category of the evaluation. Note Black-and-white photograph by Pereira Studio, from the PC 1000 Tucson General Photo Collection, Arizona Historical Society. https://cdm15812.contentdm.oclc.org/digital/collection/p15812coll15/id/14 . Test number 10. Analysis This section analyzes patterns of success and failure across the three AI platforms tested in this study—Skybox, LTX Studio, and DeeVid—by examining how each tool transformed archival photographs from the Arizona Historical Society’s Tucson 250 + collections into semi-immersive outputs. Drawing on rubric scores and qualitative evaluation, the analysis focuses on recurring visual, semantic, and ethical issues, particularly around representation, historical accuracy, and internal coherence. Rather than treating failures as isolated anomalies, this section identifies systematic patterns linking tool performance to image content, prompt language, and the presence of people, events, or identity markers. Specific archival photographs are referenced by title to ground the analysis in concrete examples. 1. Skybox Nearly all major failures in the dataset originate from Skybox. These failures are not random; instead, they show strong lexical, conceptual, and visual similarities across images and prompts. Skybox consistently performs poorly when the archival photograph depicts people, social groups, or historically specific events, particularly when identity markers such as race, ethnicity, age, number of items or gender are present. Human-centered scenes are especially problematic. For example, when Skybox was applied to Dunbar Junior High class (1949), a black-and-white group portrait documenting a segregated school in Tucson, the output significantly distorted facial features and failed to preserve the structure and seriousness of the original class photograph. Similarly, in Chinese vegetable seller on Meyer Street, Tucson (1904), Skybox ignored the archival description identifying the subject as Chinese and introduced incoherent objects and visual noise into the scene. Across such cases, Skybox frequently altered racial identity, age, and even gender. These distortions indicate that the problem is not merely aesthetic but semantic and ethical: the model fails to preserve identity markers that are central to archival meaning and historical accountability. Skybox also struggles with event-driven or complex street scenes. In images such as Alianza Hispano Americana parade or pageant, Tucson, Arizona (circa 1930s), which depicts a costumed procession outside a downtown café, the model removed or altered key elements, introduced objects that never existed, and produced chaotic compositions lacking internal logic. Similar failures occurred with historically documented events such as the Congress Hotel Fire and Fox Theatre-related scenes, where Skybox generated vague spatial layouts and visually implausible structures. Another consistent weakness involves text, numbers, and signage. In scenes containing storefront numbers, marquees, or signage, such as the “FOX” marquee or visible numbers on buildings and vehicles, Skybox failed to render legible or accurate text. This inability to reproduce written and numerical information results in outputs that are historically misleading, particularly for urban archival photographs where signage conveys temporal and cultural specificity. Interior scenes with structured spatial rhythms further expose Skybox’s limitations. In images depicting hotel lobbies, porch interiors, jury boxes, or classroom interiors, furniture placement, shadows, and perspective often became inconsistent or distorted. The model appeared unable to maintain coherence in enclosed environments with repeated architectural elements. Across the entire dataset, Skybox produced only one high-scoring result with an average above 4: Bird’s-eye view of Tucson (1947) 15 , a wide aerial photograph looking south from A Mountain. This image contains no people, minimal architectural detail, and a clear horizon line. The success of this case underscores a consistent pattern: Skybox performs best when the scene depicts a wide-open environment with no human subjects and limited semantic complexity. Taken together, these results suggest that Skybox generates what might be described as “possible pasts” rather than historically grounded reconstructions. While its outputs may be visually atmospheric, they lack the internal logic required for archival recreation. Skybox is therefore unsuitable for representing historically specific people, events, or urban scenes, though it may still be useful for early-stage prototyping of abstract or environmental backdrops. 2. LTX In contrast to Skybox, LTX Studio demonstrates substantially higher stability and coherence across the dataset. No LTX outputs scored below an average of 2, and most fall between 3 and 5. High-scoring examples include Chinese vegetable seller on Meyer Street, Tucson (1904), Five Black children, Club Filarmónico Tucsonense, Dillinger gang, and multiple views of adobe houses, railyards, railroad stations, and water reservoirs. LTX handles buildings, landscapes, and relatively stable environments particularly well, producing visually consistent and spatially coherent outputs. LTX also performs moderately well with groups of people. However, faces often appear painterly or softened, especially when the source image is low quality. While this stylization does not always undermine coherence, it does reduce photographic fidelity. Importantly, LTX does not exhibit strong lexical triggers for failure. Instead, its performance appears more sensitive to technical factors such as image resolution, lighting, and perspective complexity than to specific prompt wording. Despite its overall reliability, LTX presents a significant representational risk. In at least one case, even when the textual prompt explicitly identified the subjects as Black, the output altered their racial appearance and introduced an additional character not present in the original archival photograph. This example demonstrates that racial distortion can occur even in tools that otherwise appear consistent and technically competent. Such distortions are especially concerning in an archival context, where the preservation of identity is fundamental. Additional patterns observed across the dataset reinforce these findings. When source images were low quality, LTX often shifted toward a watercolor-like aesthetic. In some cases, the model introduced small but meaningful visual artifacts, such as altering shadows into facial hair or adding objects to clothing. Prompts containing words like “historical” tended to produce monochrome outputs, while vertical images were sometimes cropped unless the aspect ratio was manually adjusted. Overall, LTX shows clear internal logic and visual consistency, making it the most reliable of the three tools for semi-immersive archival experimentation. However, its occasional failures around racial representation highlights the need for careful human oversight, even when rubric scores are high. 3. DeeVid DeeVid displays the most uniformly high performance across the dataset, with average scores consistently between 4 and 5. The sole major failure occurred with Dunbar Junior High class (1949), which received an average score of 1.57. This failure may reflect difficulty handling group portraits involving many faces, or an inability to convey the seriousness and structure of a formal class photograph through animated camera movement. Aside from this case, DeeVid performed strongly across a wide range of subjects, including parade scenes, fire scenes, canyons, railroad stations, Black families, the Chinese vegetable seller photograph, auto parties, adobe buildings, railyards, and hotel interiors. Unlike Skybox, DeeVid rarely introduced invented objects or altered identities. Prompt wording appears to have minimal impact on DeeVid’s success. Its outputs are largely limited to zooming, panning, or subtle motion applied to the original image, rather than reconstructing the scene in depth. This constraint explains both its stability and its limitation: DeeVid does not offer dramatic reinterpretation or spatial speculation, but it also avoids many of the distortions seen in other tools. As a result, DeeVid is reliable but conceptually narrow. 4. Cross-Tool Lexical Patterns (Success vs Failure) Keywords associated with failure (mainly Skybox) group, crowd, class, parade, band, people swimming, Male/Female/Children descriptors, African American / Mexican American / Chinese, street scene, sign, marquee, numbers, fire, smoke, interior lobby, shotgun, guns, horses, clutter. Keywords associated with success (Skybox + LTX + DeeVid) aerial view / bird’s-eye view, panoramic, landscape, desert plain, mountains, adobe houses, railroad station, railyards, canyon, river, vegetation, early Tucson / 1880s / 1940s (when not mixed with people), simple architectural description. Overall, these findings show that while AI tools can create engaging or visually rich outputs, they do so in ways that are often disconnected from archival evidence, descriptive metadata, or the interpretive needs of cultural heritage institutions. Their strengths lie in ambiance, mood, and environmental immersion, while their weaknesses are concentrated in representation, accuracy, and internal consistency, which are key values in archival practice. Discussion The results of this case study reveal a clear and consistent pattern: contemporary generative AI tools offer visually impressive possibilities for transforming archival photographs into semi-immersive media, but they do so with uneven accuracy, unpredictable logic, and ethically significant distortions. Across both rubric tables, Skybox in particular shows systematic weaknesses, especially in scenes involving people, race, age, signage, or complex environments. One possible explanation for AI platforms like Skybox’s poor performance in these contexts is that the platform relies primarily on text-driven generation without an explicit reasoning or constraint mechanism to enforce historical or semantic consistency. As a result, Skybox may respond more effectively to broad, generalized descriptions than to the detailed, specific, and metadata-rich language typical of archival description, where precision rather than atmosphere is the primary goal. LTX performs more reliably but still ignores explicit identifiers of age, race, and gender, and it adds invented details. DeeVid is the most consistent but is limited in expressive depth. These findings resonate strongly with the emerging AI and cultural heritage scholarship from 2023–2025, which repeatedly warns that generative systems excel at creating plausible “atmospheres” but struggle with authenticity, representation, and contextual fidelity. Recent literature demonstrates enthusiasm for immersive technologies while simultaneously cautioning against design shortcuts and historical distortions. Studies such as Park et al. ( 2025 ) show that generative systems often default to culturally dominant or Eurocentric representations, even when prompts specify otherwise. This aligns closely with my findings: in several cases, both Skybox and LTX tended to ignore racial identifiers specified in the archival description. These distortions mirror the “cultural flattening” issues Park identifies (Park et al., 2025 ). Similarly, the concerns raised by Riter et al. ( 2025 ) about authenticity, interpretive integrity, and representational ethics in AI-generated archival content directly map onto the risks identified in this project. My rubric results show that even high-performing tools like LTX produce subtle but meaningful misinterpretations, and Skybox’s low scores across people, buildings, and events demonstrate what Riter et al. describe as “algorithmic misrepresentation,” where models generate aesthetically coherent but historically ungrounded scenes. These parallels validate the importance of the systematic error analysis in my workflow (Riter et al., 2025 ). My results suggest that AI-generated reconstructions must be carefully evaluated before being presented to the public. Immersive experiences may appear engaging but can erode historical trust if they manifestly misrepresent historical detail. Ye et al. ( 2025 ) show that users’ willingness to engage with digital heritage depends on perceived authenticity and immersion, and caution that immersion without credibility leads to distraction or misinterpretation rather than learning. To avoid this, my research suggests that instead of reducing labor, much human oversight is still needed when incorporating AI into archival workflows. As Wu (2025) highlights, AI can support intangible cultural heritage when cultural safeguards and expert review are integrated into the workflow. My findings reinforce this argument. The archival photographs used in this study document Tucson’s diverse communities. The fact that AI models repeatedly altered race, gender, and age directly contradicts Wu’s emphasis on cultural integrity (Wu, 2025). Without human oversight, AI-generated outputs risk misrepresenting or erasing marginalized histories. A major takeaway from 2019–2025 VR and AR heritage research is that immersive technologies support emotional engagement, spatial understanding, and knowledge retention when the environments are coherent and contextually accurate (Baradaran Rahimi et al., 2022 ; Chanakira et al., 2023 ; Gong et al., 2024 ; Wang, 2024 ). My results diverge sharply from this expectation in the case of AI-generated immersive spaces, such as Skybox. Although Skybox produces high emotional scores because of its 360° format, its lack of authenticity and coherence undermines its potential for learning. This tension evokes Harrington’s ( 2020 ) warning that perceived learning may increase even when actual learning does not. The risk of immersive outputs is that users may trust inaccurate AI scenes simply because they are immersive (Harrington, 2020 ). Similarly, Shehade & Stylianou-Lambert ( 2020 ) report that VR must align with historical evidence to avoid misleading museum visitors. My findings confirm that generative AI dramatically increases this risk: unlike manually designed VR environments, AI-generated scenes may contain invented details that the user cannot detect. This expands the ethical problem beyond what the VR literature has traditionally addressed (Shehade & Stylianou-Lambert, 2020 ). Archival theory offers important context for interpreting these results. Harris ( 2007 ) and Ketelaar ( 2000 ) argue that archives are constructed, interpretive spaces shaped by power, mediation, and selective representation. My outputs demonstrate how generative AI extends this mediation further, by inventing new content, altering identity markers, or introducing stylistic assumptions from its training data. This reinforces the need to treat AI-generated VR environments as interpretive objects, not neutral reconstructions. The pattern analysis in this study shows that descriptive language (archival metadata, researcher-written descriptions, and AI-generated summaries) directly affects generative outcomes. Yakel’s writing on description and access (2003) underscores the centrality of metadata in shaping how users understand archival materials. The models do not “understand” history; they translate the descriptiveness of metadata into visual speculation. This supports the argument that archival description must be carefully evaluated when used as generative input (Yakel & Deboreh A., 2003 ). An important pattern that emerges from this analysis is the difficulty users would likely face in recognizing representational distortions within AI-generated immersive outputs. In several Skybox and LTX cases, changes to racial identity, misinterpretation of signage, or the introduction of invented architectural elements were subtle enough to remain visually plausible, particularly when presented in immersive or animated formats. This suggests the presence of an “AI logic gap” in which the relationship between descriptive input and visual output is not legible to viewers. Even when prompts specify historically grounded details, the internal reasoning of the model remains opaque, making it difficult for users to assess whether an output aligns with archival evidence or diverges from it. This pattern aligns with, and extends, recent findings in the literature on AI bias and interpretability. Zhou et al. ( 2025 ) show that even knowledgeable users struggle to evaluate whether AI debiasing mechanisms are effective (Zhou et al., 2025 ). The present study suggests a parallel challenge in immersive archival contexts: viewers may be unable to identify when generative systems have altered or erased historically significant details. Related work by Park et al. ( 2025 ), Pope ( 2025 ), and Riter et al. ( 2025 ) similarly demonstrates that generative systems behave inconsistently across prompts and often rely on culturally dominant patterns embedded in training data. The prompt-level analysis in this study empirically supports these observations, particularly in cases where keywords related to race, ethnicity, gender, or community correlated with degraded or distorted outputs in Skybox and, to a lesser extent, LTX. Rather than resolving representational risk, prompt specificity alone appears insufficient to constrain generative behavior in ways that align with archival values. Conclusion This case study demonstrates both the promise and the limits of contemporary generative AI tools for archival access, interpretation, and semi-immersive engagement. By systematically comparing three commercial AI plaforms for immersive image generation across all outputs derived from 33 archival photographs, I observed clear patterns in how these systems reconstruct (or misconstruct) historical material. While all three platforms are highly accessible and capable of producing visually compelling results, their internal logic diverges sharply from archival values of authenticity, representation, and evidential reliability. The consistent weaknesses in Skybox, the selective but significant representational distortions in LTX, and the limited interpretive depth of DeeVid together highlight a core tension: AI can generate striking atmospheres, but it cannot yet reliably recreate the cultural, social, and historical specificity embedded in archival records which it confirms our need to train an archive-specific AI model to produce immersive environments. These findings reinforce concerns raised in recent scholarship on AI, XR, and cultural heritage, particularly around bias, opacity, and the risk of producing immersive misinformation. Across tools and prompt types, outputs involving people, race, age, signage, and complex events were the most vulnerable to distortion. Even high-quality prompts and rich archival descriptions could not overcome the “AI logic gap” that emerged repeatedly in this study. At the same time, the study also shows that AI can support early-stage prototyping, brainstorming, and environmental visualization when used carefully and within its limits, especially for landscapes, buildings, and non-identity-based scenes. Ultimately, this research offers a grounded, practice-oriented foundation for archivists, technologists, and cultural heritage professionals who may experiment with AI-assisted immersive access in the future. It confirms that generative AI is not yet suitable for public-facing historical reconstruction, particularly when representation and community memory are at stake. However, with stronger cultural safeguards, domain-specific training, and transparent workflows, these technologies may eventually support new forms of engagement with archival materials. This case study therefore represents both a caution and a roadmap: AI can enrich archival interpretation, but only when paired with human judgment, archival expertise, and a commitment to historical integrity. Based on this experience and conclusion following paragraphs are some recommendations for Archivists and the Arizona Historical Society and AI Developers. For archivists and institutions like the Arizona Historical Society, these findings emphasize the importance of strengthening descriptive metadata practices with AI use-cases in mind. Although the goal is not to fictionalize records, archivists may need to supplement traditional DACS (Describing Archives: A Content Standard) descriptions with additional structured detail like lighting, spatial context, number of figures, vegetation, building types, and environmental cues to help generative tools produce more responsible results. Developing controlled vocabularies for AI-friendly description could reduce distortions and preserve historical integrity. Institutions should also begin conversations about ethical guidelines for AI-assisted reconstructions: what counts as acceptable variance? How should demographic representation be monitored? And where should reconstructions be clearly labeled as interpretive rather than evidentiary? Finally, cultural heritage organizations may consider partnerships with computer scientists to train dedicated models on regionally significant datasets, ensuring that future semi-immersive reconstructions reflect local history, identities, and values rather than generic or biased AI defaults. For AI developers, these results reveal a deep need to incorporate historical reasoning, demographic consistency, and representational ethics into model design. Tools intended for use in archives or museums must adopt stricter logic engines: if a prompt specifies “Black students,” “a Chinese vegetable seller,” or “1950s Tucson,” the model should not alter race, add new figures, change architecture, or introduce anachronisms. Developers should integrate metadata schemas (DACS, Dublin Core, EAD) into training pipelines and allow models to interpret descriptive fields as constraints rather than stylistic suggestions. Fine-tuning on local historical datasets, combined with bias detection modules, can reduce misrepresentation and prevent the erasure or distortion of marginalized communities. Transparency is also crucial: users must be able to see why the model made specific choices, which internal references it weighted, and where uncertainties occur. Ultimately, AI for heritage needs to move from entertainment-driven generation to evidence-sensitive reconstruction, meaningfully grounded in historical logic, community identity, and archival ethics. Limitation This study is constrained by several limitations that should be acknowledged. First, the tools themselves displayed inconsistent performance over time, partly due to system updates that altered outputs without notice. Skybox changed download speeds, generation style, and sometimes produced unexpectedly better faces one week and worse ones the next, making reproducibility difficult. DeeVid experienced repeated “fail” states during rendering, especially when adding sound effects, which significantly slowed output generation and reduced the number of samples possible; even when successful, sound-enabled videos occasionally altered facial structure in the output, reducing authenticity. LTX temporarily went offline after an update, preventing generation for several hours. The dataset also includes cases where source images were undated, low quality, or lacking descriptive detail, reflecting typical archival conditions but limiting the tool’s ability to perform consistently. Finally, the study focuses on semi-immersive outputs rather than fully interactive VR environments, and the findings apply only to the tools and subscriptions available during the research period. Despite these limitations, the failures themselves are analytically valuable, illustrating how fragile and unpredictable generative systems can be when applied to historically sensitive material. Declarations Funding This research received no external funding. All expenses related to this project were covered by the author. Competing Interests The author declares no competing financial or non-financial interests related to this work. Data availability No datasets were generated or analyzed during the current study. Author Contribution The author conducted all aspects of the research, including study design, data collection, analysis, interpretation, and manuscript preparation, and approved the final version for publication. This is a solo author research. Acknowledgement The author would like to thank Marie Saldaña, Assistant Professor in the College of Information at the University of Arizona, for her guidance, insightful feedback, and ongoing support throughout the Directed Research course in fall 2025 that informed this study. References Baradaran Rahimi, F., Boyd, J. E., Eiserman, J. R., Levy, R. M., & Kim, B. (2022). Museum beyond physical walls: An exploration of virtual reality-enhanced experience in an exhibition-like space. Virtual Reality , 26 (4), 1471–1488. https://doi.org/10.1007/s10055-022-00643-5 Boboc, R. G., Băutu, E., Gîrbacia, F., Popovici, N., & Popovici, D.-M. (2022). Augmented Reality in Cultural Heritage: An Overview of the Last Decade of Applications. Applied Sciences , 12 (19), 9859. https://doi.org/10.3390/app12199859 Chanakira, T. A., Happonen, A., & Hasheela, V. (2023). An Applied Study of the Impact of Augmented Reality on Museum Experience: A Focus on Museums in Namibia. In Prof. G. Bucci (Ed.), Research Highlights in Science and Technology Vol. 8 (pp. 104–132). B P International (a part of SCIENCEDOMAIN International). https://doi.org/10.9734/bpi/rhst/v8/5976C Colavizza, G., Blanke, T., Jeurgens, C., & Noordegraaf, J. (2022). Archives and AI: An Overview of Current Debates and Future Perspectives. Journal on Computing and Cultural Heritage , 15 (1), 1–15. https://doi.org/10.1145/3479010 Colegrove, P. T., & Mikel, M. (2018). Radical inclusion: Immersive 360-degree video capture, dissemination, and use of emerging technology in support of traditional archival roles at a university library. Proceedings of the Association for Information Science and Technology , 55 (1), 779–780. https://doi.org/10.1002/pra2.2018.14505501113 Effective Prompts for AI: The Essentials. (n.d.). [Edu]. MIT Sloan Teaching & Learning Technologies . Retrieved January 11, 2026, from https://mitsloanedtech.mit.edu/ai/basics/effective-prompts/#What_is_a_Prompt Farella, M. (2025). Digital solutions for immersive virtual reality exploration of cultural heritage sites [Doctoral thesis, University of Palermo, University of Catania, University of Messina, University of Palermo]. https://tesidottorato.depositolegale.it/static/PDF/web/viewer.jsp GadAllah, S. I. I. (2020). Using Modern Technologies in the Museums’ Exhibitions: The Grand Egyptian Museum as a Case Study. International Journal of Heritage, Tourism and Hospitality , 14 (Issue 3 (Special Issue)), 347–359. https://doi.org/10.21608/ijhth.2020.299549 Gong, Q., Zou, N., Yang, W., Zheng, Q., & Chen, P. (2024). User experience model and design strategies for virtual reality-based cultural heritage exhibition. Virtual Reality , 28 (2), 69. https://doi.org/10.1007/s10055-024-00942-z Harrington, M. C. R. (2020). Connecting User Experience to Learning in an Evaluation of an Immersive, Interactive, Multimodal Augmented Reality Virtual Diorama in a Natural History Museum & the Importance of Story. 2020 6th International Conference of the Immersive Learning Research Network (iLRN) , 70–78. https://doi.org/10.23919/iLRN47897.2020.9155202 Harris, V. S. (2007). Archives and justice: A South African perspective . Society of American Archivists. Hazan, S. (2023). The Dance of the Doppelgängers: AI and the cultural heritage community . Proceedings of EVA London 2023. https://doi.org/10.14236/ewic/EVA2023.13 Hettmann, W., Wölfel, M., Butz, M., Torner, K., & Finken, J. (2023). Engaging Museum Visitors with AI-Generated Narration and Gameplay. In A. L. Brooks (Ed.), ArtsIT, Interactivity and Game Creation (Vol. 479, pp. 201–214). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-28993-4_15 Ketelaar, E. (2000). Archivistics Research Saving the Profession. The American Archivist , 63 (2), 322–340. https://doi.org/10.17723/aarc.63.2.0238574511vmv576 Ketelaar, E. (2010). Ten years of archival science. Archival Science , 10 (4), 345–352. https://doi.org/10.1007/s10502-011-9137-2 Kotsiubivska, K., Tymoshenko, O., & Vasylevsky, A. (2024). Artificial Intelligence Tools for Preservation and Popularization of Cultural Heritage. Digital Platform: Information Technologies in Sociocultural Sphere , 7 (2), 275–282. https://doi.org/10.31866/2617-796X.7.2.2024.317736 Park, H., Patel, J., Wise, N., Sim, U., Wang, Y., & Williams-Pierce, C. (2025). Stereotypes, Storytelling, and the (Un)Reliability of Generative AI. Proceedings of the Association for Information Science and Technology , 62 (1), 1628–1630. https://doi.org/10.1002/pra2.1489 Pope, A. (2025). Time Out of Mind: Investigating the Technical and Ethical Impacts of Virtual Reality Tools in Archaeological Studies. Proceedings of the Association for Information Science and Technology , 62 (1), 1643–1646. https://doi.org/10.1002/pra2.1494 Riter, R. B., Mehra, B., Poole, A. H., Lischer-Katz, Z., Tribelhorn, S., & Wagner, T. (2025). Representation and Authenticity in AI Generated, Curated, and Mediated Archives. Proceedings of the Association for Information Science and Technology , 62 (1), 1655–1657. https://doi.org/10.1002/pra2.1498 Shehade, M., & Stylianou-Lambert, T. (2020). Virtual Reality in Museums: Exploring the Experiences of Museum Professionals. Applied Sciences , 10 (11), 4031. https://doi.org/10.3390/app10114031 Shinde, G., Kirstein, T., Ghosh, S., & Franks, P. C. (2024). AI in Archival Science—A Systematic Review (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2410.09086 Spadoni, E., Carulli, M., Ferrise, F., & Bordegoni, M. (2023). Impact of Multisensory XR Technologies on Museum Exhibition Visits. In M. Antona & C. Stephanidis (Eds.), Universal Access in Human-Computer Interaction (Vol. 14021, pp. 120–132). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-35897-5_9 Wang, X. (2024). [DC] User Exploratory Learning in a Virtual Reality Museum. 2024 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW) , 1152–1153. https://doi.org/10.1109/VRW62533.2024.00370 Xu, J., & Pan, Y. (2024). The Future Museum: Integrating Augmented Reality (AR) and Virtual-text with AI-enhanced Information Systems. Journal of Information Systems Engineering and Management , 10 (1), 25826. https://doi.org/10.55267/iadt.07.15439 Yakel, E., & Deboreh A., T. (2003). AI: Archival Intelligence and User Expertise. The American Archivist , 66 (1), 51–78. https://www.jstor.org/stable/40294217 Ye, X., Ruan, Y., Xia, S., & Gu, L. (2025). Adoption of digital intangible cultural heritage: A configurational study integrating UTAUT2 and immersion theory. Humanities and Social Sciences Communications , 12 (1), 23. https://doi.org/10.1057/s41599-024-04222-8 Zhang, J. (2012). Archival Representation in the Digital Age. Journal of Archival Organization , 10 (1), 45–68. https://doi.org/10.1080/15332748.2012.677671 Zhou, K. Z., Cao, J., Yuan, X., Weissglass, D. E., Kilhoffer, Z., Sanfilippo, M. R., & Tong, X. (2025). “I’m not confident in debiasing AI systems since I know too little”: Designing and Evaluating Hands-on Gender Bias Tutorials for AI Practitioners and Learners. Proceedings of the Association for Information Science and Technology , 62 (1), 835–847. https://doi.org/10.1002/pra2.1301 Footnotes https://arizonahistoricalsociety.org/tucson250/ https://www.meshy.ai/ https://runwayml.com/ https://lumalabs.ai/ https://poly.cam/ https://www.tripo3d.ai/ https://depth-r.com/ https://gemini.google.com/app https://openart.ai/ https://pika.art/login https://copilot.microsoft.com/imagine https://skybox.blockadelabs.com/ https://ltx.studio/ https://deevid.ai/ Bird’s-eye view of Tucson. (1947). [Black-and-white photograph]. PC 1000 Tucson General Photograph Collection. https://azhsarchives.contentdm.oclc.org/digital/collection/p15812coll14/id/32 . Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8615665","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":583245249,"identity":"2b622d17-82e6-47dc-bc98-9a15f43d1d02","order_by":0,"name":"Farzaneh Talebhaghighi","email":"data:image/png;base64,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","orcid":"","institution":"University of Arizona","correspondingAuthor":true,"prefix":"","firstName":"Farzaneh","middleName":"","lastName":"Talebhaghighi","suffix":""}],"badges":[],"createdAt":"2026-01-16 07:49:42","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8615665/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8615665/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102427028,"identity":"a354ed3c-54a8-4776-99a5-b31ec5f7079f","added_by":"auto","created_at":"2026-02-11 14:45:25","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":168960,"visible":true,"origin":"","legend":"\u003cp\u003eAverage evaluator ratings (scale 1–5) for Rubric 1 across seven assessments\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/026bba1ee6688df56215dfc1.jpg"},{"id":102745704,"identity":"885214c1-5395-49ab-93f0-3ce24abe6925","added_by":"auto","created_at":"2026-02-16 08:53:29","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":183829,"visible":true,"origin":"","legend":"\u003cp\u003eAverage evaluator ratings (scale 1–5) for Rubric 2 across eight items\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/920698f734e348bc2fba80a4.jpg"},{"id":102745513,"identity":"491bf761-c111-4da3-bba4-87a552883cab","added_by":"auto","created_at":"2026-02-16 08:51:23","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":296293,"visible":true,"origin":"","legend":"\u003cp\u003eDunbar Junior High class (1949). Note. Black-and-white photograph from the PC 1000 Tucson General Photo Collection, Arizona Historical Society. https://cdm15812.contentdm.oclc.org/digital/collection/p15812coll15/id/12. Test number27\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/d30d86cc100c7467d75ece12.jpg"},{"id":102427022,"identity":"9e3fdcab-8188-4673-924a-05c5ef43e409","added_by":"auto","created_at":"2026-02-11 14:45:24","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":276111,"visible":true,"origin":"","legend":"\u003cp\u003eExample of adding a new character to the original image by LTX\u003c/p\u003e","description":"","filename":"Picture4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/f05e5c362ca02e0df995683d.jpg"},{"id":102427027,"identity":"fc623f78-e2bf-4766-b485-ea44583ca6a4","added_by":"auto","created_at":"2026-02-11 14:45:25","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":462973,"visible":true,"origin":"","legend":"\u003cp\u003eCocooned B-29s at Davis–Monthan Air Force Base. Note. Black-and-white photograph by Western Ways, from the MS 1255 Western Ways Manuscript and Photograph Collection, Arizona Historical Society. https://cdm15812.contentdm.oclc.org/digital/collection/p15812coll16/id/52. Test number 5.\u003c/p\u003e","description":"","filename":"Picture5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/78e1b5bea5124fcda47d4c4e.jpg"},{"id":102427020,"identity":"b4347a79-7716-4973-b7f8-eab7f3da5471","added_by":"auto","created_at":"2026-02-11 14:45:24","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":152285,"visible":true,"origin":"","legend":"\u003cp\u003eSkybox output with an AI description. Note the change in camera angle and vague geometry of the airplanes\u003c/p\u003e","description":"","filename":"Picture6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/d6b89d4c47fd9b77dee4cc15.jpg"},{"id":102427023,"identity":"7bb81aee-d614-4e50-8673-744949d1d924","added_by":"auto","created_at":"2026-02-11 14:45:25","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":324029,"visible":true,"origin":"","legend":"\u003cp\u003eAlianza Hispano Americana parade or pageant, Tucson, Arizona (circa 1930s).\u003cbr\u003e\n Note.Black-and-white photograph by Pereira Studio, from the PC 1000 Tucson General Photo Collection, Arizona Historical Society. https://cdm15812.contentdm.oclc.org/digital/collection/p15812coll15/id/14. Test number 10.\u003c/p\u003e","description":"","filename":"Picture7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/50681dcf145784f470710c46.jpg"},{"id":102427029,"identity":"1f2d827f-5aea-446d-95af-700393bf348b","added_by":"auto","created_at":"2026-02-11 14:45:25","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":213932,"visible":true,"origin":"","legend":"\u003cp\u003eResult from Sybox, based on archival description. Note creation of arbitrary objects.\u003c/p\u003e","description":"","filename":"Picture8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/f34d0c0bf1ca55de46ebdef4.jpg"},{"id":102745831,"identity":"4e5806fa-ad47-45a3-b6eb-25bbdf0cc408","added_by":"auto","created_at":"2026-02-16 08:54:14","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":205479,"visible":true,"origin":"","legend":"\u003cp\u003eChinese vegetable seller on Meyer Street, Tucson (1904). Note. Black-and-white photograph, creator unknown, from the PC 1000 Tucson General Photo Collection, Arizona Historical Society. https://cdm15812.contentdm.oclc.org/digital/collection/p15812coll15/id/36. Test number 21.\u003c/p\u003e","description":"","filename":"Picture9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/8cabb603e11a917dcf90cc69.jpg"},{"id":102427030,"identity":"e4e448e9-53b5-475c-8464-01f8bbd16043","added_by":"auto","created_at":"2026-02-11 14:45:25","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":178122,"visible":true,"origin":"","legend":"\u003cp\u003eExample of Skybox creating a scene based on researcher description. Skybox has ignored the description which identifies the individual as Chinese and adds incoherent objects.\u003c/p\u003e","description":"","filename":"Picture10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/f86ca3660b40cf3100c16f2c.jpg"},{"id":104782655,"identity":"1e5216ee-6012-4ad2-8db4-9394e497e661","added_by":"auto","created_at":"2026-03-17 07:57:39","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3093305,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8615665/v1/757c3950-1feb-496b-920a-22a1bd4c3c25.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Towards Archival Engagement With AI: An exploratory study using historical photographs","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAs museums and archives experiment with new technologies to support learning and visitor engagement, immersive and semi-immersive media have become increasingly prominent. Research in museum and heritage contexts suggests that interactive and immersive environments can deepen user engagement, emotional involvement, and learning outcomes (Spadoni et al., 2023; Xu \u0026amp; Pan, 2024). However, archival studies also show the practical constraints faced by institutions, including limited staff time, budgets, and technological capacity (Chanakira et al., 2023; Shehade \u0026amp; Stylianou-Lambert, 2020). These tensions form the backdrop for this project, which examines whether AI-driven tools might help reduce production burdens while still enabling audiences to interact with archival photographs in more compelling ways.\u003c/p\u003e\n\u003cp\u003eThis study draws on archival photographs from three collections included in the Arizona Historical Society (AHS)’s “Tucson 250+: Where We Live, What We Do, and Who We Are” online exhibition\u003csup\u003e[1]\u003c/sup\u003e, to evaluate how contemporary AI tools interpret and transform historical visual materials. My original intention was to generate fully interactive VR scenes from these photographs. However, early experiments revealed significant limitations in current AI capacities. These constraints required a methodological pivot: rather than creating full VR environments, the project shifted toward using historical photos to produce images with enhanced immersive affect, such as 360-degree scenes, image-to-video transformations, and colorized or augmented representations. These approaches still afford semi-immersive engagement but are more technically feasible and more closely aligned with the capabilities of existing AI systems.\u003c/p\u003e\n\u003cp\u003eThis pivot reflects a core insight that emerged through practice: while immersive technologies hold promise for cultural heritage engagement, the process of generating VR-ready content remains expensive, time-consuming, and technically complex (GadAllah, 2020; Shehade \u0026amp; Stylianou-Lambert, 2020). AI may help streamline parts of production, yet generative models also introduce new risks, such as hallucination, cultural bias, inaccuracies, and the erasure or distortion of marginalized identities (Hazan, 2023; Kotsiubivska et al., 2024). Therefore, a second goal of the current study is to understand how AI tools behave when applied to archival materials. This kind of analysis is essential for increasing dialogue between archivists and developers on the risks and opportunities of AI in archives, libraries, and museums.\u003c/p\u003e\n\u003cp\u003eBy systematically testing AI-generated outputs across 33 archival images, scoring the results with a rubric, and analyzing patterns in the outputs produced by specific prompt strategies, this research examines how contemporary AI can support archival interpretation, what distortions and biases emerge, and what this means for the future of AI-assisted cultural heritage access. The goal is not simply to evaluate the tools but to understand how AI intersects with archival description, metadata, visual accuracy, and the ethical considerations embedded in archival practice. In doing so, the project contributes to broader conversations in archival theory about the constructed nature of records (Harris, 2007; Ketelaar, 2000, 2010), the interpretive power embedded in description (Zhang, 2012), and the responsibilities institutions bear when mediating historical memory through emerging technologies.\u003c/p\u003e\n"},{"header":"Background","content":"\u003cp\u003eImmersive and AI driven systems have become central to recent work in cultural heritage, archives, and museums. Large scale reviews and case studies show that AR and VR have been used to increase engagement, learning, and interpretive richness across museums, archives, and heritage sites, often through interactive storytelling, spatial reconstructions, and adaptive content delivery (Boboc et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Colavizza et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Kotsiubivska et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Shinde et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Xu \u0026amp; Pan, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). These projects usually rely on custom VR pipelines, game engines, and institution specific infrastructure, which makes them powerful but also costly and difficult to replicate in small or under-resourced archives.\u003c/p\u003e \u003cp\u003eWithin immersive cultural heritage specifically, recent work has begun to integrate artificial intelligence as part of the interpretive layer, rather than as the generator of the visual scene. For example, Farella\u0026rsquo;s dissertation on immersive exploration of cultural heritage sites uses an AI-driven query framework inside a photogrammetry-based VR environment so that visitors can ask context-dependent questions and receive tailored explanations (Farella, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Similarly, AI has been used to generate narration, branching stories, or gameplay elements inside museum experiences, augmenting engagement and learning without altering the underlying historical images or 3D models (Hettmann et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This strand of work shows how AI can enrich meaning in immersive spaces, but it does not focus on how AI might be used to fabricate the spatial environment itself.\u003c/p\u003e \u003cp\u003eA second cluster of research addresses user acceptance, immersion quality, and the risks and benefits of adopting digital heritage technologies. Studies on the adoption of digital intangible cultural heritage, for example, highlight how willingness to use immersive tools is shaped by performance expectations, social influence, hedonic motivation, and perceived immersion, but also by concerns about distraction, discomfort, and authenticity (Ye et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). In parallel, Pope\u0026rsquo;s work with archaeologists documents excitement about VR for storytelling and preservation alongside deep worries about accuracy, cultural sensitivity, and the danger that audiences may treat virtual reconstructions as historically true (Pope, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). From an archival perspective, this aligns with broader critiques of generative AI in museums and heritage that warn about issues of authorship, bias, and \u0026ldquo;doppelg\u0026auml;nger\u0026rdquo; versions of culture that can be mistaken for the real thing (Hazan, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Other projects demonstrate how immersive technologies can be integrated into archival practice without generative reconstruction. Colegrove and Mikel\u0026rsquo;s \u0026ldquo;Radical Inclusion\u0026rdquo; project uses consumer grade 360 degree video to document ephemeral events and fold them into archival workflows, with attention to metadata and sustainable access (Colegrove \u0026amp; Mikel, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Here, immersive media functions as a recording strategy rather than a speculative reimagining.\u003c/p\u003e \u003cp\u003eMy study departs from this documentation-oriented model by examining what happens when AI is used to transform archival photographs into semi-immersive 3D scenes, 360 style images, and short AI-generated videos. The literature shows that immersive heritage systems and AI-enhanced interpretation can support engagement and learning, and that users\u0026rsquo; adoption of such systems depends strongly on immersion quality, comfort, and perceived authenticity. At the same time, there is very little empirical work that systematically evaluates off-the-shelf generative tools, especially in small archival settings that rely on subscription-based platforms rather than custom pipelines. Existing studies rarely analyze patterns of AI error across different content types, or connect those patterns to archival concerns about representation, authority, and the future use of AI-generated content in public access systems. This gap motivates the present study, which focuses on identifying \u0026ldquo;AI logic\u0026rdquo; patterns and mismatches in the outputs of three contemporary tools and considers what these patterns might mean for archivists who may want to experiment with AI-assisted immersive access in practice.\u003c/p\u003e "},{"header":"Research Questions","content":" \u003cp\u003eThis exploratory study focuses on examining how contemporary artificial intelligence tools transform archival photographs into semi-immersive outputs such as 360-degree views, image-based videos, and enhanced image representations. The broader goal is to evaluate the strengths, limitations, costs, technical barriers, and potential uses of these tools within archival practice. A secondary, longer-term vision for this research is to assess the feasibility of embedding AI-assisted immersive features directly into archival websites, enabling users to generate VR-like experiences from archival materials on demand. Guided by these aims, the study addresses the following research questions:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eWhat kinds of errors, confusions, or mismatches appear in AI-generated outputs?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eHow do these errors correspond to specific content types or platforms?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eWhat strengths, alignments, or meaningful enhancements appear in AI-generated outputs?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eWhat are the costs, time requirements, learning curves, and technical barriers of using these tools with archival photo collections?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eWhich AI-generated outputs best support engagement with history?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eFor each AI tool, what combinations of images and prompts contribute to successful outcomes, and under what conditions do outputs degrade or fail?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThese questions allow the study to evaluate AI tools not only as creative engines, but also as potential components of future archival access systems, where ethical representation, accuracy, and public engagement must remain central. The final question originally planned for this study, \u0026ldquo;How can archivists balance imagination, annotation, and AI augmentation with fidelity to historical evidence?\u0026rdquo;, will be addressed in the researcher\u0026rsquo;s \u003cem\u003enext\u003c/em\u003e project, which will involve qualitative interviews with archivists and cultural heritage professionals.\u003c/p\u003e \u003cp\u003ePurpose and Significance\u003c/p\u003e \u003cp\u003eThe purpose of this study is to evaluate how three contemporary AI tools (Skybox, LTX Studio, and DeeVid) interpret and transform archival photographs into semi-immersive outputs such as 360-degree views, image-based videos, and enhanced representations. By systematically testing these tools on images from the Arizona Historical Society (AHS)\u0026rsquo;s Tucson 250\u0026thinsp;+\u0026thinsp;collections, the study identifies patterns of accuracy, distortion, bias, and creative augmentation, as well as the technical and conceptual limits of each platform.\u003c/p\u003e \u003cp\u003eThe significance of this work lies in its direct relevance to archival practice. As archives begin to experiment with AI-assisted access and interpretation, they require practical evidence about what these tools can and cannot reliably produce. This study contributes insight into fidelity, historical alignment, prompt sensitivity, distortion, and tool behavior across image types, offering archivists an early roadmap for safe experimentation. The findings also highlight the need for future, archivist-centered models designed around archival values, authenticity, and community engagement.\u003c/p\u003e "},{"header":"Methodology","content":" \u003cp\u003eThis project unfolded as a practice-based, exploratory case study shaped by both the affordances and the limitations of working with AI tools that were never designed for archival use. From the beginning, my intention was to experiment with AI platforms that promised 3D, video, or immersive scene generation and evaluate whether they could support a lightweight, semi-immersive access layer for archival photographs. The process did not follow a rigid experimental protocol; instead, it evolved iteratively as I confronted real technical barriers, unexpected failures, and surprising tool behaviors.\u003c/p\u003e \u003cp\u003eI began by selecting images from the Arizona Historical Society\u0026rsquo;s (AHS) Tucson 250\u0026thinsp;+\u0026thinsp;exhibition, ensuring representation across its three subcollections: Where We Live, What We Do, and Who We Are. My goal was to choose images that reflected the thematic diversity of the exhibition: landscapes, street scenes, daily life, buildings, people, and group activities. In the end, to ensure appropriate distribution and coverage of all themes in the exhibition, I selected 11 images from each subcollection, resulting in a total of 33 archival photographs used for the study. Exclusion criteria were adopted to avoid extremely low-resolution images that prevented any meaningful visual interpretation, and duplicates or highly redundant views when a near equivalent was already selected.\u003c/p\u003e \u003cp\u003eThis project began with a broad exploratory phase designed to map the current landscape of AI tools capable of transforming archival photographs into semi-immersive media. Rather than assuming that a single platform could meet the needs of this case study, I intentionally approached the process with an open survey mindset. The goal was to understand what kinds of outputs were even possible with existing tools, how they differed in behavior, and where their technical limits appeared in practice.\u003c/p\u003e \u003cp\u003eIn the first stage, I experimented with a wide range of AI platforms offering image enhancement, animation, 3D reconstruction, and scene generation features. These included: Meshy\u003csup\u003e2\u003c/sup\u003e, Runway\u003csup\u003e3\u003c/sup\u003e, Luma AI\u003csup\u003e4\u003c/sup\u003e, Polycam\u003csup\u003e5\u003c/sup\u003e, Tripo 3D\u003csup\u003e6\u003c/sup\u003e, DepthR\u003csup\u003e7\u003c/sup\u003e, Gemini Nano\u003csup\u003e8\u003c/sup\u003e, OpenArt\u003csup\u003e9\u003c/sup\u003e, Pikart\u003csup\u003e10\u003c/sup\u003e, Copilot Image Generator\u003csup\u003e11\u003c/sup\u003e. Each tool was tested with the same guiding question: Can this platform meaningfully transform an archival photograph into a 3D model, a semi-immersive scene, or a dynamic video suitable for public engagement? Throughout this exploratory phase, I evaluated whether the tools could recreate spatial environments, preserve historical details, or even respond coherently to archival descriptions. Most of the tools tested provided interesting prototypes but were not sufficiently consistent, historically faithful, or adapted to the image quality of archival material. Some platforms produced visually appealing results but disregarded historical details. Others could generate 3D reconstructions only when given high-resolution, object-centered images that differed significantly from the archival photographs used in this study. This phase of research clarified that the challenge of my project lay not only generating an immersive scene, but in doing so in a way that respected the constraints of archival images and descriptions.\u003c/p\u003e \u003cp\u003eUltimately, after testing this wider landscape of tools, I selected Skybox\u003csup\u003e12\u003c/sup\u003e, LTX Studio\u003csup\u003e13\u003c/sup\u003e, and DeeVid\u003csup\u003e14\u003c/sup\u003e for further evaluation. Skybox AI by Blockade Labs is an AI-powered tool that generates immersive, 360-degree panoramic images (skyboxes) from simple text prompts. LTX Studio is Lightricks' first entirely AI-powered platform, enabling professionals to visualize and develop any creative concept and DeeVid turns still images into animated video clips, whether you're using a single image or a series of images. All these three platforms offered the most stable semi-immersive outputs, supported text and/or image input, demonstrated clear differences in logic and behavior, and represented three distinct technical approaches (360-generation, video generation, and scene representation). These three tools became the focus of the larger case study and the foundation for the pattern analysis discussed in later sections. Before beginning formal testing, I spent time learning each platform, experimenting with their settings, subscriptions, and limitations. I subscribed to all three tools on a monthly basis, knowing that some features were locked behind paid access.\u003c/p\u003e \u003cp\u003eThe outputs generated by the three platforms differed according to their respective input structures and technical configurations. LTX Studio produced still-image outputs based on a combination of textual prompts and a single input image, supported by a built-in prompt optimizer. For each photograph, LTX Studio generated three colorized or enhanced image variants derived from the same image and text prompt. These outputs were produced using the FLUX configuration, with no additional stylistic parameters applied, a fixed 16:9 aspect ratio, you can change it to vertical manually, and no specified camera angle or location. DeeVid generated video outputs from single images paired with textual prompts. Two output configurations were tested: Fast V2.0, which produced 5-second videos at 512p resolution without sound effects, and Master V2.0, which produced 8-second videos at 720p resolution with sound effects enabled. SkyBox generated immersive outputs in the form of 360-degree environmental views, based on textual prompts or mixed text\u0026ndash;image inputs, and included a prompt optimization feature. These outputs were generated using model 3, with an image size of 8K and a landscape depth map. Outputs were organized by source image, tool, date, and prompt type to allow controlled comparison in an Excel sheet.\u003c/p\u003e \u003cp\u003ePrompting, \u0026ldquo;an input into the AI system to obtain specific results\u0026rdquo; (\u0026ldquo;Prompts for AI,\u0026rdquo; n.d.), became one of the most challenging and revealing parts of the methodology. For each platform, I began by using as a prompt the archival descriptions exactly as written by AHS. When results were poor, I tested two additional variations: an AI-generated prompt (using ChatGPT to create a 600-character Skybox prompt from the archival metadata) and a researcher-crafted prompt where I wrote my own description. Throughout the process, I documented not only the outputs, but also the tool behaviors: crashes, \u0026ldquo;high demand\u0026rdquo; messages, failed renders, missing textures, and unexplained changes in model behavior following software updates. These observations became just as important as the outputs themselves, because they revealed how unpredictable these systems currently are for archival use.\u003c/p\u003e \u003cp\u003eTo evaluate the results, I built a rubric with multiple dimensions: fidelity, completeness, coherence, emotional impact, authenticity, engagement, ease/comfort, and an overall qualitative notes section. I scored each output individually, and only after completing all evaluations, I calculated averages and looked for patterns. I also created a second table scored independently from the archival originals, focusing instead on how each tool handled specific content categories: race, gender, age, animals, buildings, environment, objects, vegetation, and people. Finally, once all outputs were scored, I began analyzing patterns across prompts, image types, and tools. I looked for repeated failures, repeated successes, and the specific words that led to either outcome. These patterns became the basis for my findings and shaped the larger implications for archivists who may want to experiment with AI-assisted access in the future.\u003c/p\u003e "},{"header":"Findings","content":" \u003cp\u003eAcross all the outputs generated from 33 archival photographs, clear patterns emerged in how each tool interpreted archival materials, how prompts influenced results, and where distortions or meaningful enhancements appeared. The findings reflect both the numerical scores (1\u0026ndash;5, 1 for the lowest score and 5 for the highest score) from the rubric tables and the qualitative insights that surfaced through repeated experimentation. Together, these patterns illustrate not only what the tools can do, but also how their internal logic diverges from archival values and constraints.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRubric Table number 1 Results (scores from 1 to 5)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFidelity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCompleteness\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCoherence\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEmotional Impact\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAuthenticity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eEngagement\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eEase \u0026amp; Comfort\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_ARC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e5.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e5.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_RES\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e5.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_DeeVid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e5.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_LTX\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.09\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e5.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c8\" namest=\"c1\"\u003e \u003cp\u003eAVG_ARC stands for average score of outputs from SKYBOX with an archival description.\u003c/p\u003e \u003cp\u003eAVG_AI stands for average score of outputs from SKYBOX with an AI description.\u003c/p\u003e \u003cp\u003eAVG_RES stands for average score of outputs from SKYBOX with the researcher description.\u003c/p\u003e \u003cp\u003eAVG_ DeeVid stands for average score of outputs from DeeVid. platform.\u003c/p\u003e \u003cp\u003eAVG_LTX stands for average score of outputs from LTX Studio platform.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe two evaluation tables reveal distinct but complementary patterns about how each tool performs when generating outputs from archival photographs and textual descriptions. When the focus is on compatibility with the original archival record (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), measured through fidelity, completeness, coherence, emotional impact, authenticity, engagement, and ease of use, a clear hierarchy emerges. Skybox performs the weakest across all three prompt types (archival description, AI-generated description, and researcher-written description), particularly in fidelity, authenticity, and completeness. Although ease-of-use scores are high for all tools, Skybox\u0026rsquo;s inability to recreate historically grounded scenes remains consistent, and the differences between its three prompt types are minimal. In other words, changing the description does not dramatically improve Skybox\u0026rsquo;s reliability. By contrast, DeeVid and LTX consistently achieve higher scores, with DeeVid showing the strongest overall performance in reconstructing or visually extending the original archival scenes, followed closely by LTX.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRubric Table number 2 Results (scores from 1 to 5)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003elandscape\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePeople\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ebuilding\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eobjects\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eanimal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003erace\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003egender\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eage\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_ARC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e3.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e3.63\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e3.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e2.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_RES\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e3.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e3.50\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_DeeVid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e4.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e4.80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAVG_LTX\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e4.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e4.80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe second table shifts the focus from historical accuracy to general performance across content types: landscapes, people, buildings, objects, animals, race, gender, and age. Here again the overall pattern aligns with the first table. Skybox underperforms in nearly all categories involving people, bodies, identity, and social scenes, reflecting its difficulty with differentiation in human subjects. Even when prompts avoided identity terms, Skybox frequently produced distorted or semantically incorrect representations, confirming the lack of internal logic previously observed in qualitative results. LTX and DeeVid both score substantially higher across all categories; however, one LTX anomaly reveals an important limitation: LTX ignored text in the archival descriptions that specified racial identity in the subjects and sometimes introduced an additional person who did not exist in the archival record. This case shows that even high-performing tools are not immune to representational bias. Consistent with earlier observations, DeeVid remains the most semantically reliable, performing strongly across all categories, though its output style is limited to camera motion rather than full reconstruction.\u003c/p\u003e \u003cp\u003eOverall, the patterns across both tables show that Skybox\u0026rsquo;s weaknesses are systematic and rooted in how it interprets (or fails to interpret) descriptive prompts, whereas LTX and DeeVid show higher stability and accuracy but still present risks in racially sensitive contexts. These findings also highlight a methodological insight: textual descriptions of the past, whether through archival metadata, AI-generated prompts, or researcher prompts, are a central driver of how AI tools interpret or misinterpret history. The language given to these systems shapes the internal logic they apply, for better or for worse, and this dependency is visible across every category of the evaluation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eNote\u003c/strong\u003e \u003cp\u003e \u003cem\u003eBlack-and-white photograph by Pereira Studio, from the PC 1000 Tucson General Photo Collection, Arizona Historical Society.\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://cdm15812.contentdm.oclc.org/digital/collection/p15812coll15/id/14\u003c/span\u003e\u003cspan address=\"https://cdm15812.contentdm.oclc.org/digital/collection/p15812coll15/id/14\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. \u003cem\u003eTest number 10.\u003c/em\u003e\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAnalysis\u003c/p\u003e \u003cp\u003eThis section analyzes patterns of success and failure across the three AI platforms tested in this study\u0026mdash;Skybox, LTX Studio, and DeeVid\u0026mdash;by examining how each tool transformed archival photographs from the Arizona Historical Society\u0026rsquo;s Tucson 250\u0026thinsp;+\u0026thinsp;collections into semi-immersive outputs. Drawing on rubric scores and qualitative evaluation, the analysis focuses on recurring visual, semantic, and ethical issues, particularly around representation, historical accuracy, and internal coherence. Rather than treating failures as isolated anomalies, this section identifies systematic patterns linking tool performance to image content, prompt language, and the presence of people, events, or identity markers. Specific archival photographs are referenced by title to ground the analysis in concrete examples.\u003c/p\u003e \u003cp\u003e1. Skybox\u003c/p\u003e \u003cp\u003eNearly all major failures in the dataset originate from Skybox. These failures are not random; instead, they show strong lexical, conceptual, and visual similarities across images and prompts. Skybox consistently performs poorly when the archival photograph depicts people, social groups, or historically specific events, particularly when identity markers such as race, ethnicity, age, number of items or gender are present.\u003c/p\u003e \u003cp\u003eHuman-centered scenes are especially problematic. For example, when Skybox was applied to Dunbar Junior High class (1949), a black-and-white group portrait documenting a segregated school in Tucson, the output significantly distorted facial features and failed to preserve the structure and seriousness of the original class photograph. Similarly, in Chinese vegetable seller on Meyer Street, Tucson (1904), Skybox ignored the archival description identifying the subject as Chinese and introduced incoherent objects and visual noise into the scene. Across such cases, Skybox frequently altered racial identity, age, and even gender. These distortions indicate that the problem is not merely aesthetic but semantic and ethical: the model fails to preserve identity markers that are central to archival meaning and historical accountability.\u003c/p\u003e \u003cp\u003eSkybox also struggles with event-driven or complex street scenes. In images such as Alianza Hispano Americana parade or pageant, Tucson, Arizona (circa 1930s), which depicts a costumed procession outside a downtown caf\u0026eacute;, the model removed or altered key elements, introduced objects that never existed, and produced chaotic compositions lacking internal logic. Similar failures occurred with historically documented events such as the Congress Hotel Fire and Fox Theatre-related scenes, where Skybox generated vague spatial layouts and visually implausible structures.\u003c/p\u003e \u003cp\u003eAnother consistent weakness involves text, numbers, and signage. In scenes containing storefront numbers, marquees, or signage, such as the \u0026ldquo;FOX\u0026rdquo; marquee or visible numbers on buildings and vehicles, Skybox failed to render legible or accurate text. This inability to reproduce written and numerical information results in outputs that are historically misleading, particularly for urban archival photographs where signage conveys temporal and cultural specificity.\u003c/p\u003e \u003cp\u003eInterior scenes with structured spatial rhythms further expose Skybox\u0026rsquo;s limitations. In images depicting hotel lobbies, porch interiors, jury boxes, or classroom interiors, furniture placement, shadows, and perspective often became inconsistent or distorted. The model appeared unable to maintain coherence in enclosed environments with repeated architectural elements.\u003c/p\u003e \u003cp\u003eAcross the entire dataset, Skybox produced only one high-scoring result with an average above 4: Bird\u0026rsquo;s-eye view of Tucson (1947)\u003csup\u003e15\u003c/sup\u003e, a wide aerial photograph looking south from A Mountain. This image contains no people, minimal architectural detail, and a clear horizon line. The success of this case underscores a consistent pattern: Skybox performs best when the scene depicts a wide-open environment with no human subjects and limited semantic complexity.\u003c/p\u003e \u003cp\u003eTaken together, these results suggest that Skybox generates what might be described as \u0026ldquo;possible pasts\u0026rdquo; rather than historically grounded reconstructions. While its outputs may be visually atmospheric, they lack the internal logic required for archival recreation. Skybox is therefore unsuitable for representing historically specific people, events, or urban scenes, though it may still be useful for early-stage prototyping of abstract or environmental backdrops.\u003c/p\u003e \u003cp\u003e2. LTX\u003c/p\u003e \u003cp\u003eIn contrast to Skybox, LTX Studio demonstrates substantially higher stability and coherence across the dataset. No LTX outputs scored below an average of 2, and most fall between 3 and 5. High-scoring examples include Chinese vegetable seller on Meyer Street, Tucson (1904), Five Black children, Club Filarm\u0026oacute;nico Tucsonense, Dillinger gang, and multiple views of adobe houses, railyards, railroad stations, and water reservoirs. LTX handles buildings, landscapes, and relatively stable environments particularly well, producing visually consistent and spatially coherent outputs.\u003c/p\u003e \u003cp\u003eLTX also performs moderately well with groups of people. However, faces often appear painterly or softened, especially when the source image is low quality. While this stylization does not always undermine coherence, it does reduce photographic fidelity. Importantly, LTX does not exhibit strong lexical triggers for failure. Instead, its performance appears more sensitive to technical factors such as image resolution, lighting, and perspective complexity than to specific prompt wording.\u003c/p\u003e \u003cp\u003eDespite its overall reliability, LTX presents a significant representational risk. In at least one case, even when the textual prompt explicitly identified the subjects as Black, the output altered their racial appearance and introduced an additional character not present in the original archival photograph. This example demonstrates that racial distortion can occur even in tools that otherwise appear consistent and technically competent. Such distortions are especially concerning in an archival context, where the preservation of identity is fundamental.\u003c/p\u003e \u003cp\u003eAdditional patterns observed across the dataset reinforce these findings. When source images were low quality, LTX often shifted toward a watercolor-like aesthetic. In some cases, the model introduced small but meaningful visual artifacts, such as altering shadows into facial hair or adding objects to clothing. Prompts containing words like \u0026ldquo;historical\u0026rdquo; tended to produce monochrome outputs, while vertical images were sometimes cropped unless the aspect ratio was manually adjusted.\u003c/p\u003e \u003cp\u003eOverall, LTX shows clear internal logic and visual consistency, making it the most reliable of the three tools for semi-immersive archival experimentation. However, its occasional failures around racial representation highlights the need for careful human oversight, even when rubric scores are high.\u003c/p\u003e \u003cp\u003e3. DeeVid\u003c/p\u003e \u003cp\u003eDeeVid displays the most uniformly high performance across the dataset, with average scores consistently between 4 and 5. The sole major failure occurred with Dunbar Junior High class (1949), which received an average score of 1.57. This failure may reflect difficulty handling group portraits involving many faces, or an inability to convey the seriousness and structure of a formal class photograph through animated camera movement.\u003c/p\u003e \u003cp\u003eAside from this case, DeeVid performed strongly across a wide range of subjects, including parade scenes, fire scenes, canyons, railroad stations, Black families, the Chinese vegetable seller photograph, auto parties, adobe buildings, railyards, and hotel interiors. Unlike Skybox, DeeVid rarely introduced invented objects or altered identities.\u003c/p\u003e \u003cp\u003ePrompt wording appears to have minimal impact on DeeVid\u0026rsquo;s success. Its outputs are largely limited to zooming, panning, or subtle motion applied to the original image, rather than reconstructing the scene in depth. This constraint explains both its stability and its limitation: DeeVid does not offer dramatic reinterpretation or spatial speculation, but it also avoids many of the distortions seen in other tools. As a result, DeeVid is reliable but conceptually narrow.\u003c/p\u003e \u003cp\u003e4. Cross-Tool Lexical Patterns (Success vs Failure)\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eKeywords associated with failure (mainly Skybox)\u003c/strong\u003e \u003cp\u003egroup, crowd, class, parade, band, people swimming, Male/Female/Children descriptors, African American / Mexican American / Chinese, street scene, sign, marquee, numbers, fire, smoke, interior lobby, shotgun, guns, horses, clutter.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eKeywords associated with success (Skybox\u0026thinsp;+\u0026thinsp;LTX\u0026thinsp;+\u0026thinsp;DeeVid)\u003c/strong\u003e \u003cp\u003eaerial view / bird\u0026rsquo;s-eye view, panoramic, landscape, desert plain, mountains, adobe houses, railroad station, railyards, canyon, river, vegetation, early Tucson / 1880s / 1940s (when not mixed with people), simple architectural description.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eOverall, these findings show that while AI tools can create engaging or visually rich outputs, they do so in ways that are often disconnected from archival evidence, descriptive metadata, or the interpretive needs of cultural heritage institutions. Their strengths lie in ambiance, mood, and environmental immersion, while their weaknesses are concentrated in representation, accuracy, and internal consistency, which are key values in archival practice.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe results of this case study reveal a clear and consistent pattern: contemporary generative AI tools offer visually impressive possibilities for transforming archival photographs into semi-immersive media, but they do so with uneven accuracy, unpredictable logic, and ethically significant distortions. Across both rubric tables, Skybox in particular shows systematic weaknesses, especially in scenes involving people, race, age, signage, or complex environments. One possible explanation for AI platforms like Skybox\u0026rsquo;s poor performance in these contexts is that the platform relies primarily on text-driven generation without an explicit reasoning or constraint mechanism to enforce historical or semantic consistency. As a result, Skybox may respond more effectively to broad, generalized descriptions than to the detailed, specific, and metadata-rich language typical of archival description, where precision rather than atmosphere is the primary goal. LTX performs more reliably but still ignores explicit identifiers of age, race, and gender, and it adds invented details. DeeVid is the most consistent but is limited in expressive depth. These findings resonate strongly with the emerging AI and cultural heritage scholarship from 2023\u0026ndash;2025, which repeatedly warns that generative systems excel at creating plausible \u0026ldquo;atmospheres\u0026rdquo; but struggle with authenticity, representation, and contextual fidelity.\u003c/p\u003e \u003cp\u003eRecent literature demonstrates enthusiasm for immersive technologies while simultaneously cautioning against design shortcuts and historical distortions. Studies such as Park et al. (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) show that generative systems often default to culturally dominant or Eurocentric representations, even when prompts specify otherwise. This aligns closely with my findings: in several cases, both Skybox and LTX tended to ignore racial identifiers specified in the archival description. These distortions mirror the \u0026ldquo;cultural flattening\u0026rdquo; issues Park identifies (Park et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Similarly, the concerns raised by Riter et al. (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) about authenticity, interpretive integrity, and representational ethics in AI-generated archival content directly map onto the risks identified in this project. My rubric results show that even high-performing tools like LTX produce subtle but meaningful misinterpretations, and Skybox\u0026rsquo;s low scores across people, buildings, and events demonstrate what Riter et al. describe as \u0026ldquo;algorithmic misrepresentation,\u0026rdquo; where models generate aesthetically coherent but historically ungrounded scenes. These parallels validate the importance of the systematic error analysis in my workflow (Riter et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eMy results suggest that AI-generated reconstructions must be carefully evaluated before being presented to the public. Immersive experiences may appear engaging but can erode historical trust if they manifestly misrepresent historical detail. Ye et al. (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) show that users\u0026rsquo; willingness to engage with digital heritage depends on perceived authenticity and immersion, and caution that immersion without credibility leads to distraction or misinterpretation rather than learning. To avoid this, my research suggests that instead of reducing labor, much human oversight is still needed when incorporating AI into archival workflows. As Wu (2025) highlights, AI can support intangible cultural heritage when cultural safeguards and expert review are integrated into the workflow. My findings reinforce this argument. The archival photographs used in this study document Tucson\u0026rsquo;s diverse communities. The fact that AI models repeatedly altered race, gender, and age directly contradicts Wu\u0026rsquo;s emphasis on cultural integrity (Wu, 2025). Without human oversight, AI-generated outputs risk misrepresenting or erasing marginalized histories.\u003c/p\u003e \u003cp\u003eA major takeaway from 2019\u0026ndash;2025 VR and AR heritage research is that immersive technologies support emotional engagement, spatial understanding, and knowledge retention when the environments are coherent and contextually accurate (Baradaran Rahimi et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Chanakira et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Gong et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Wang, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). My results diverge sharply from this expectation in the case of AI-generated immersive spaces, such as Skybox. Although Skybox produces high emotional scores because of its 360\u0026deg; format, its lack of authenticity and coherence undermines its potential for learning. This tension evokes Harrington\u0026rsquo;s (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) warning that perceived learning may increase even when actual learning does not. The risk of immersive outputs is that users may trust inaccurate AI scenes simply because they are immersive (Harrington, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Similarly, Shehade \u0026amp; Stylianou-Lambert (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) report that VR must align with historical evidence to avoid misleading museum visitors. My findings confirm that generative AI dramatically increases this risk: unlike manually designed VR environments, AI-generated scenes may contain invented details that the user cannot detect. This expands the ethical problem beyond what the VR literature has traditionally addressed (Shehade \u0026amp; Stylianou-Lambert, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eArchival theory offers important context for interpreting these results. Harris (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2007\u003c/span\u003e) and Ketelaar (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) argue that archives are constructed, interpretive spaces shaped by power, mediation, and selective representation. My outputs demonstrate how generative AI extends this mediation further, by inventing new content, altering identity markers, or introducing stylistic assumptions from its training data. This reinforces the need to treat AI-generated VR environments as interpretive objects, not neutral reconstructions. The pattern analysis in this study shows that descriptive language (archival metadata, researcher-written descriptions, and AI-generated summaries) directly affects generative outcomes. Yakel\u0026rsquo;s writing on description and access (2003) underscores the centrality of metadata in shaping how users understand archival materials. The models do not \u0026ldquo;understand\u0026rdquo; history; they translate the descriptiveness of metadata into visual speculation. This supports the argument that archival description must be carefully evaluated when used as generative input (Yakel \u0026amp; Deboreh A., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2003\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAn important pattern that emerges from this analysis is the difficulty users would likely face in recognizing representational distortions within AI-generated immersive outputs. In several Skybox and LTX cases, changes to racial identity, misinterpretation of signage, or the introduction of invented architectural elements were subtle enough to remain visually plausible, particularly when presented in immersive or animated formats. This suggests the presence of an \u0026ldquo;AI logic gap\u0026rdquo; in which the relationship between descriptive input and visual output is not legible to viewers. Even when prompts specify historically grounded details, the internal reasoning of the model remains opaque, making it difficult for users to assess whether an output aligns with archival evidence or diverges from it.\u003c/p\u003e \u003cp\u003eThis pattern aligns with, and extends, recent findings in the literature on AI bias and interpretability. Zhou et al. (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) show that even knowledgeable users struggle to evaluate whether AI debiasing mechanisms are effective (Zhou et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). The present study suggests a parallel challenge in immersive archival contexts: viewers may be unable to identify when generative systems have altered or erased historically significant details. Related work by Park et al. (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), Pope (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), and Riter et al. (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) similarly demonstrates that generative systems behave inconsistently across prompts and often rely on culturally dominant patterns embedded in training data. The prompt-level analysis in this study empirically supports these observations, particularly in cases where keywords related to race, ethnicity, gender, or community correlated with degraded or distorted outputs in Skybox and, to a lesser extent, LTX. Rather than resolving representational risk, prompt specificity alone appears insufficient to constrain generative behavior in ways that align with archival values.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis case study demonstrates both the promise and the limits of contemporary generative AI tools for archival access, interpretation, and semi-immersive engagement. By systematically comparing three commercial AI plaforms for immersive image generation across all outputs derived from 33 archival photographs, I observed clear patterns in how these systems reconstruct (or misconstruct) historical material. While all three platforms are highly accessible and capable of producing visually compelling results, their internal logic diverges sharply from archival values of authenticity, representation, and evidential reliability. The consistent weaknesses in Skybox, the selective but significant representational distortions in LTX, and the limited interpretive depth of DeeVid together highlight a core tension: AI can generate striking atmospheres, but it cannot yet reliably recreate the cultural, social, and historical specificity embedded in archival records which it confirms our need to train an archive-specific AI model to produce immersive environments.\u003c/p\u003e \u003cp\u003eThese findings reinforce concerns raised in recent scholarship on AI, XR, and cultural heritage, particularly around bias, opacity, and the risk of producing immersive misinformation. Across tools and prompt types, outputs involving people, race, age, signage, and complex events were the most vulnerable to distortion. Even high-quality prompts and rich archival descriptions could not overcome the \u0026ldquo;AI logic gap\u0026rdquo; that emerged repeatedly in this study. At the same time, the study also shows that AI can support early-stage prototyping, brainstorming, and environmental visualization when used carefully and within its limits, especially for landscapes, buildings, and non-identity-based scenes.\u003c/p\u003e \u003cp\u003eUltimately, this research offers a grounded, practice-oriented foundation for archivists, technologists, and cultural heritage professionals who may experiment with AI-assisted immersive access in the future. It confirms that generative AI is not yet suitable for public-facing historical reconstruction, particularly when representation and community memory are at stake. However, with stronger cultural safeguards, domain-specific training, and transparent workflows, these technologies may eventually support new forms of engagement with archival materials. This case study therefore represents both a caution and a roadmap: AI can enrich archival interpretation, but only when paired with human judgment, archival expertise, and a commitment to historical integrity. Based on this experience and conclusion following paragraphs are some recommendations for Archivists and the Arizona Historical Society and AI Developers.\u003c/p\u003e \u003cp\u003eFor archivists and institutions like the Arizona Historical Society, these findings emphasize the importance of strengthening descriptive metadata practices with AI use-cases in mind. Although the goal is not to fictionalize records, archivists may need to supplement traditional DACS (Describing Archives: A Content Standard) descriptions with additional structured detail like lighting, spatial context, number of figures, vegetation, building types, and environmental cues to help generative tools produce more responsible results. Developing controlled vocabularies for AI-friendly description could reduce distortions and preserve historical integrity. Institutions should also begin conversations about ethical guidelines for AI-assisted reconstructions: what counts as acceptable variance? How should demographic representation be monitored? And where should reconstructions be clearly labeled as interpretive rather than evidentiary? Finally, cultural heritage organizations may consider partnerships with computer scientists to train dedicated models on regionally significant datasets, ensuring that future semi-immersive reconstructions reflect local history, identities, and values rather than generic or biased AI defaults.\u003c/p\u003e \u003cp\u003eFor AI developers, these results reveal a deep need to incorporate historical reasoning, demographic consistency, and representational ethics into model design. Tools intended for use in archives or museums must adopt stricter logic engines: if a prompt specifies \u0026ldquo;Black students,\u0026rdquo; \u0026ldquo;a Chinese vegetable seller,\u0026rdquo; or \u0026ldquo;1950s Tucson,\u0026rdquo; the model should not alter race, add new figures, change architecture, or introduce anachronisms. Developers should integrate metadata schemas (DACS, Dublin Core, EAD) into training pipelines and allow models to interpret descriptive fields as constraints rather than stylistic suggestions. Fine-tuning on local historical datasets, combined with bias detection modules, can reduce misrepresentation and prevent the erasure or distortion of marginalized communities. Transparency is also crucial: users must be able to see why the model made specific choices, which internal references it weighted, and where uncertainties occur. Ultimately, AI for heritage needs to move from entertainment-driven generation to evidence-sensitive reconstruction, meaningfully grounded in historical logic, community identity, and archival ethics.\u003c/p\u003e \u003cp\u003eLimitation\u003c/p\u003e \u003cp\u003eThis study is constrained by several limitations that should be acknowledged. First, the tools themselves displayed inconsistent performance over time, partly due to system updates that altered outputs without notice. Skybox changed download speeds, generation style, and sometimes produced unexpectedly better faces one week and worse ones the next, making reproducibility difficult. DeeVid experienced repeated \u0026ldquo;fail\u0026rdquo; states during rendering, especially when adding sound effects, which significantly slowed output generation and reduced the number of samples possible; even when successful, sound-enabled videos occasionally altered facial structure in the output, reducing authenticity. LTX temporarily went offline after an update, preventing generation for several hours. The dataset also includes cases where source images were undated, low quality, or lacking descriptive detail, reflecting typical archival conditions but limiting the tool\u0026rsquo;s ability to perform consistently. Finally, the study focuses on semi-immersive outputs rather than fully interactive VR environments, and the findings apply only to the tools and subscriptions available during the research period. Despite these limitations, the failures themselves are analytically valuable, illustrating how fragile and unpredictable generative systems can be when applied to historically sensitive material.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis research received no external funding. All expenses related to this project were covered by the author.\u003c/p\u003e \u003cp\u003eCompeting Interests\u003c/p\u003e \u003cp\u003eThe author declares no competing financial or non-financial interests related to this work.\u003c/p\u003e \u003cp\u003eData availability\u003c/p\u003e \u003cp\u003eNo datasets were generated or analyzed during the current study.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eThe author conducted all aspects of the research, including study design, data collection, analysis, interpretation, and manuscript preparation, and approved the final version for publication. This is a solo author research.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe author would like to thank Marie Salda\u0026ntilde;a, Assistant Professor in the College of Information at the University of Arizona, for her guidance, insightful feedback, and ongoing support throughout the Directed Research course in fall 2025 that informed this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBaradaran Rahimi, F., Boyd, J. E., Eiserman, J. R., Levy, R. M., \u0026amp; Kim, B. (2022). Museum beyond physical walls: An exploration of virtual reality-enhanced experience in an exhibition-like space. \u003cem\u003eVirtual Reality\u003c/em\u003e, \u003cem\u003e26\u003c/em\u003e(4), 1471\u0026ndash;1488. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10055-022-00643-5\u003c/span\u003e\u003cspan address=\"10.1007/s10055-022-00643-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoboc, R. G., Băutu, E., G\u0026icirc;rbacia, F., Popovici, N., \u0026amp; Popovici, D.-M. (2022). Augmented Reality in Cultural Heritage: An Overview of the Last Decade of Applications. \u003cem\u003eApplied Sciences\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(19), 9859. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/app12199859\u003c/span\u003e\u003cspan address=\"10.3390/app12199859\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChanakira, T. A., Happonen, A., \u0026amp; Hasheela, V. (2023). An Applied Study of the Impact of Augmented Reality on Museum Experience: A Focus on Museums in Namibia. In Prof. G. Bucci (Ed.), \u003cem\u003eResearch Highlights in Science and Technology Vol. 8\u003c/em\u003e (pp. 104\u0026ndash;132). B P International (a part of SCIENCEDOMAIN International). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.9734/bpi/rhst/v8/5976C\u003c/span\u003e\u003cspan address=\"10.9734/bpi/rhst/v8/5976C\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eColavizza, G., Blanke, T., Jeurgens, C., \u0026amp; Noordegraaf, J. (2022). Archives and AI: An Overview of Current Debates and Future Perspectives. \u003cem\u003eJournal on Computing and Cultural Heritage\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(1), 1\u0026ndash;15. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/3479010\u003c/span\u003e\u003cspan address=\"10.1145/3479010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eColegrove, P. T., \u0026amp; Mikel, M. (2018). Radical inclusion: Immersive 360-degree video capture, dissemination, and use of emerging technology in support of traditional archival roles at a university library. \u003cem\u003eProceedings of the Association for Information Science and Technology\u003c/em\u003e, \u003cem\u003e55\u003c/em\u003e(1), 779\u0026ndash;780. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/pra2.2018.14505501113\u003c/span\u003e\u003cspan address=\"10.1002/pra2.2018.14505501113\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEffective Prompts for AI: The Essentials. (n.d.). [Edu]. \u003cem\u003eMIT Sloan Teaching \u0026amp; Learning Technologies\u003c/em\u003e. Retrieved January 11, 2026, from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://mitsloanedtech.mit.edu/ai/basics/effective-prompts/#What_is_a_Prompt\u003c/span\u003e\u003cspan address=\"https://mitsloanedtech.mit.edu/ai/basics/effective-prompts/#What_is_a_Prompt\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFarella, M. (2025). \u003cem\u003eDigital solutions for immersive virtual reality exploration of cultural heritage sites\u003c/em\u003e [Doctoral thesis, University of Palermo, University of Catania, University of Messina, University of Palermo]. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://tesidottorato.depositolegale.it/static/PDF/web/viewer.jsp\u003c/span\u003e\u003cspan address=\"https://tesidottorato.depositolegale.it/static/PDF/web/viewer.jsp\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGadAllah, S. I. I. (2020). Using Modern Technologies in the Museums\u0026rsquo; Exhibitions: The Grand Egyptian Museum as a Case Study. \u003cem\u003eInternational Journal of Heritage, Tourism and Hospitality\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(Issue 3 (Special Issue)), 347\u0026ndash;359. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.21608/ijhth.2020.299549\u003c/span\u003e\u003cspan address=\"10.21608/ijhth.2020.299549\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGong, Q., Zou, N., Yang, W., Zheng, Q., \u0026amp; Chen, P. (2024). User experience model and design strategies for virtual reality-based cultural heritage exhibition. \u003cem\u003eVirtual Reality\u003c/em\u003e, \u003cem\u003e28\u003c/em\u003e(2), 69. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10055-024-00942-z\u003c/span\u003e\u003cspan address=\"10.1007/s10055-024-00942-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHarrington, M. C. R. (2020). Connecting User Experience to Learning in an Evaluation of an Immersive, Interactive, Multimodal Augmented Reality Virtual Diorama in a Natural History Museum \u0026amp; the Importance of Story. \u003cem\u003e2020 6th International Conference of the Immersive Learning Research Network (iLRN)\u003c/em\u003e, 70\u0026ndash;78. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.23919/iLRN47897.2020.9155202\u003c/span\u003e\u003cspan address=\"10.23919/iLRN47897.2020.9155202\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHarris, V. S. (2007). \u003cem\u003eArchives and justice: A South African perspective\u003c/em\u003e. Society of American Archivists.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHazan, S. (2023). \u003cem\u003eThe Dance of the Doppelg\u0026auml;ngers: AI and the cultural heritage community\u003c/em\u003e. Proceedings of EVA London 2023. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.14236/ewic/EVA2023.13\u003c/span\u003e\u003cspan address=\"10.14236/ewic/EVA2023.13\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHettmann, W., W\u0026ouml;lfel, M., Butz, M., Torner, K., \u0026amp; Finken, J. (2023). Engaging Museum Visitors with AI-Generated Narration and Gameplay. In A. L. Brooks (Ed.), \u003cem\u003eArtsIT, Interactivity and Game Creation\u003c/em\u003e (Vol. 479, pp. 201\u0026ndash;214). Springer Nature Switzerland. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-031-28993-4_15\u003c/span\u003e\u003cspan address=\"10.1007/978-3-031-28993-4_15\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKetelaar, E. (2000). Archivistics Research Saving the Profession. \u003cem\u003eThe American Archivist\u003c/em\u003e, \u003cem\u003e63\u003c/em\u003e(2), 322\u0026ndash;340. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.17723/aarc.63.2.0238574511vmv576\u003c/span\u003e\u003cspan address=\"10.17723/aarc.63.2.0238574511vmv576\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKetelaar, E. (2010). Ten years of archival science. \u003cem\u003eArchival Science\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(4), 345\u0026ndash;352. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10502-011-9137-2\u003c/span\u003e\u003cspan address=\"10.1007/s10502-011-9137-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKotsiubivska, K., Tymoshenko, O., \u0026amp; Vasylevsky, A. (2024). Artificial Intelligence Tools for Preservation and Popularization of Cultural Heritage. \u003cem\u003eDigital Platform: Information Technologies in Sociocultural Sphere\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(2), 275\u0026ndash;282. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.31866/2617-796X.7.2.2024.317736\u003c/span\u003e\u003cspan address=\"10.31866/2617-796X.7.2.2024.317736\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark, H., Patel, J., Wise, N., Sim, U., Wang, Y., \u0026amp; Williams-Pierce, C. (2025). Stereotypes, Storytelling, and the (Un)Reliability of Generative AI. \u003cem\u003eProceedings of the Association for Information Science and Technology\u003c/em\u003e, \u003cem\u003e62\u003c/em\u003e(1), 1628\u0026ndash;1630. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/pra2.1489\u003c/span\u003e\u003cspan address=\"10.1002/pra2.1489\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePope, A. (2025). Time Out of Mind: Investigating the Technical and Ethical Impacts of Virtual Reality Tools in Archaeological Studies. \u003cem\u003eProceedings of the Association for Information Science and Technology\u003c/em\u003e, \u003cem\u003e62\u003c/em\u003e(1), 1643\u0026ndash;1646. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/pra2.1494\u003c/span\u003e\u003cspan address=\"10.1002/pra2.1494\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRiter, R. B., Mehra, B., Poole, A. H., Lischer-Katz, Z., Tribelhorn, S., \u0026amp; Wagner, T. (2025). Representation and Authenticity in AI Generated, Curated, and Mediated Archives. \u003cem\u003eProceedings of the Association for Information Science and Technology\u003c/em\u003e, \u003cem\u003e62\u003c/em\u003e(1), 1655\u0026ndash;1657. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/pra2.1498\u003c/span\u003e\u003cspan address=\"10.1002/pra2.1498\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShehade, M., \u0026amp; Stylianou-Lambert, T. (2020). Virtual Reality in Museums: Exploring the Experiences of Museum Professionals. \u003cem\u003eApplied Sciences\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(11), 4031. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/app10114031\u003c/span\u003e\u003cspan address=\"10.3390/app10114031\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShinde, G., Kirstein, T., Ghosh, S., \u0026amp; Franks, P. C. (2024). \u003cem\u003eAI in Archival Science\u0026mdash;A Systematic Review\u003c/em\u003e (Version 1). arXiv. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.2410.09086\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.2410.09086\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSpadoni, E., Carulli, M., Ferrise, F., \u0026amp; Bordegoni, M. (2023). Impact of Multisensory XR Technologies on Museum Exhibition Visits. In M. Antona \u0026amp; C. Stephanidis (Eds.), \u003cem\u003eUniversal Access in Human-Computer Interaction\u003c/em\u003e (Vol. 14021, pp. 120\u0026ndash;132). Springer Nature Switzerland. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-031-35897-5_9\u003c/span\u003e\u003cspan address=\"10.1007/978-3-031-35897-5_9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, X. (2024). [DC] User Exploratory Learning in a Virtual Reality Museum. \u003cem\u003e2024 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW)\u003c/em\u003e, 1152\u0026ndash;1153. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/VRW62533.2024.00370\u003c/span\u003e\u003cspan address=\"10.1109/VRW62533.2024.00370\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu, J., \u0026amp; Pan, Y. (2024). The Future Museum: Integrating Augmented Reality (AR) and Virtual-text with AI-enhanced Information Systems. \u003cem\u003eJournal of Information Systems Engineering and Management\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(1), 25826. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.55267/iadt.07.15439\u003c/span\u003e\u003cspan address=\"10.55267/iadt.07.15439\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYakel, E., \u0026amp; Deboreh A., T. (2003). AI: Archival Intelligence and User Expertise. \u003cem\u003eThe American Archivist\u003c/em\u003e, \u003cem\u003e66\u003c/em\u003e(1), 51\u0026ndash;78. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.jstor.org/stable/40294217\u003c/span\u003e\u003cspan address=\"https://www.jstor.org/stable/40294217\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYe, X., Ruan, Y., Xia, S., \u0026amp; Gu, L. (2025). Adoption of digital intangible cultural heritage: A configurational study integrating UTAUT2 and immersion theory. \u003cem\u003eHumanities and Social Sciences Communications\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(1), 23. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1057/s41599-024-04222-8\u003c/span\u003e\u003cspan address=\"10.1057/s41599-024-04222-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, J. (2012). Archival Representation in the Digital Age. \u003cem\u003eJournal of Archival Organization\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(1), 45\u0026ndash;68. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/15332748.2012.677671\u003c/span\u003e\u003cspan address=\"10.1080/15332748.2012.677671\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou, K. Z., Cao, J., Yuan, X., Weissglass, D. E., Kilhoffer, Z., Sanfilippo, M. R., \u0026amp; Tong, X. (2025). \u0026ldquo;I\u0026rsquo;m not confident in debiasing AI systems since I know too little\u0026rdquo;: Designing and Evaluating Hands-on Gender Bias Tutorials for AI Practitioners and Learners. \u003cem\u003eProceedings of the Association for Information Science and Technology\u003c/em\u003e, \u003cem\u003e62\u003c/em\u003e(1), 835\u0026ndash;847. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/pra2.1301\u003c/span\u003e\u003cspan address=\"10.1002/pra2.1301\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Footnotes","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arizonahistoricalsociety.org/tucson250/\u003c/span\u003e\u003cspan address=\"https://arizonahistoricalsociety.org/tucson250/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.meshy.ai/\u003c/span\u003e\u003cspan address=\"https://www.meshy.ai/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://runwayml.com/\u003c/span\u003e\u003cspan address=\"https://runwayml.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://lumalabs.ai/\u003c/span\u003e\u003cspan address=\"https://lumalabs.ai/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://poly.cam/\u003c/span\u003e\u003cspan address=\"https://poly.cam/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.tripo3d.ai/\u003c/span\u003e\u003cspan address=\"https://www.tripo3d.ai/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://depth-r.com/\u003c/span\u003e\u003cspan address=\"https://depth-r.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://gemini.google.com/app\u003c/span\u003e\u003cspan address=\"https://gemini.google.com/app\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://openart.ai/\u003c/span\u003e\u003cspan address=\"https://openart.ai/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pika.art/login\u003c/span\u003e\u003cspan address=\"https://pika.art/login\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://copilot.microsoft.com/imagine\u003c/span\u003e\u003cspan address=\"https://copilot.microsoft.com/imagine\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://skybox.blockadelabs.com/\u003c/span\u003e\u003cspan address=\"https://skybox.blockadelabs.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ltx.studio/\u003c/span\u003e\u003cspan address=\"https://ltx.studio/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://deevid.ai/\u003c/span\u003e\u003cspan address=\"https://deevid.ai/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e Bird\u0026rsquo;s-eye view of Tucson. (1947). [Black-and-white photograph]. PC 1000 Tucson General Photograph Collection. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://azhsarchives.contentdm.oclc.org/digital/collection/p15812coll14/id/32\u003c/span\u003e\u003cspan address=\"https://azhsarchives.contentdm.oclc.org/digital/collection/p15812coll14/id/32\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Archival practice, generative artificial intelligence, immersive technologies, cultural heritage, archival photographs, representation and bias, AI ethics, metadata and description, semi-immersive media, digital archives","lastPublishedDoi":"10.21203/rs.3.rs-8615665/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8615665/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAs archives and cultural heritage institutions experiment with immersive and semi-immersive technologies, generative artificial intelligence is increasingly proposed as a means of expanding access to archival materials. This exploratory case study evaluates three commercial artificial intelligence tools: Skybox, LTX Studio, and DeeVid by applying them to thirty-three historical photographs from the Arizona Historical Society\u0026rsquo;s Tucson 250+: \u003cem\u003eWhere We Live, What We Do, and Who We Are digital\u003c/em\u003e exhibition. Using a rubric-based evaluation framework, the study assesses fidelity, coherence, authenticity, engagement, and representational accuracy. The findings reveal systematic differences across platforms: Skybox frequently produces historically inaccurate outputs in scenes involving people and identity markers; LTX Studio demonstrates greater visual consistency but introduces subtle representational distortions; and DeeVid generates stable but limited transformations. The results indicate that generative tools currently require substantial human oversight to align with archival standards and ethical responsibilities. Also, the results point to the importance of conceptually informed and logic-based approaches in the development of artificial intelligence tools for archival applications.\u003c/p\u003e","manuscriptTitle":"Towards Archival Engagement With AI: An exploratory study using historical photographs","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-11 14:45:20","doi":"10.21203/rs.3.rs-8615665/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e5ecc609-f5a4-4d86-a8ae-dd0e3a95a3f3","owner":[],"postedDate":"February 11th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-16T07:12:44+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-11 14:45:20","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8615665","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8615665","identity":"rs-8615665","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}