{"paper_id":"019ca840-9bcb-4a0e-bddf-da0842b13776","body_text":"Fine-Grained Detection of AI-Generated Writing in the Biomedical Literature \n \nRichard She1 \n \n1School of Biological Sciences, Nanyang Technological University, Singapore \n \nAbstract \nGenerative AI systems are rapidly being integrated into scientific workflows, yet the specific ways \nin which AI-generated prose appears in published literature remain poorly characterized. Here, \nwe use  Pangram, a transformer -based detector optimized for adversarial paraphrasing , to \nanalyze full-length biomedical research articles from 13 major journals. Papers published in 2021-\n2024 showed almost no detectable AI -generated text, whereas manuscripts published in 2025 \nexhibited a sharp increase, with 12.4% containing at least one localized passage classified as AI-\nwritten. AI usage was highly nonuniform across authors and geography: 32% of papers originating \nfrom South Korean institutions and 26% papers from Chinese institutions contained AI-generated \npassages, compared to  7.4% from U.S. institutions . In a focused case analysis, six labs that \npublished fully AI -generated manuscript s also produced additional papers with extensive AI -\ngenerated segments. Journals likewise differed, with high-selectivity venues rarely containing AI-\nauthored prose, while high -volume journals accounted for most AI -positive manuscripts . \nTogether, these findings provide the first detailed empirical map of how and where AI-generated \nwriting is entering the scientific literature, underscoring the need for clear norms and policies \ngoverning the use of generative AI in scientific communication. \n \nIntroduction \n“It takes all the running you can do, to keep in the same place.\" – The Red Queen \n \nAt this point, everyone knows that their students are using AI to write their essays and do their \nhomework. Who among us has not received an email beginning with “Dear Professor [Last \nName].” For astute observers of more skilled students, one can still pic k up on telltale AI \nsignatures, from the proliferation of punctuation such as the em dash to an enrichment for \nuncommon verbs like “delve.” Furthermore, m odern large language models (LLMs)  are \nirresistibly drawn to a certain rhetorical tic, the corrective pivot, using an “not X but Y” cadence, \nno matter how you prompt them. It’s not a glitch of style — it’s actually their native register. \n \nHowever, for any individual student, it is effectively impossible to prove guilt beyond a \nreasonable doubt. Because of its amorphous nature, AI-generated text has proven exceeding ly \ndifficult to watermark. Research at OpenAI on statistical watermarking stretches back to 2022 1, \nbut these schemes do not survive adversarial editing, and any forensic signal evaporates as soon \nas the text is retouched. In this landscape, the arms race is asymmetric, and the students are \nwinning. With a bit of paraphrasing and cosmetic tweaking, they  can assume impunity. \nFurthermore, the students are not alone; an informal poll of my former colleagues revealed that \nnearly every grad student and postdoc uses frontier LLMs every day. While LLM adoption likely \nvaries by geography, age, and scientific field, given both the seductive power and genuine utility \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint \n\nof these systems, one might reasonably wonder to what extent AI-generated writing has already \nmade its way into the scientific literature. \n \nSince individual cases of AI usage have thus far proven impossible to adjudicate, existing studies \nhave focused on coarse grained metrics. One study that analyzed over 15 million abstracts from \nthe biomedical literature documen ted the emergence of a bon a fide LLM lexicon, with specific \nmodel-favored words rising at rates that cannot be explained by natural linguistic drift 2. Other \ncase reports point to the use of generative AI during peer review3, suggesting that academics find \nthese tools both useful and embarrassing, deploying them only behind the shield of anonymity \nfor work that never becomes part of the permanent record. A broader analysis of conference \npeer-review data reveals that the estimat ed fraction of LLM -generated reviews spikes as \ndeadlines approach4, underscoring the messy practical and behavioral variables that determine \nwhere and when generative AI is used . However, these studies leave several critical questions \nunanswered: 1) who exactly is using generative AI? 2) to what extent is AI being used  in fully \npublished and peer-reviewed manuscripts? and 3) as AI assisted writing becomes more prevalent \nand inevitable, how will the scientific community choose to govern its use? \n \nTo answer these questions, we need tools that can surpass the limits of human intuition alone in \ndetecting AI generated text5,6; in practice, it takes a neural network to reliably catch another \nneural network. Furthermore, state-of-the-art AI detectors must be robust against not only light \nparaphrasing but the next stage of the arms race : AI “humanizers”, a class of software tools \ndesigned to rewrite model output so that it appears plausibly human . Such humanizers break \nmost naïve AI detectors 7. We therefore turn to Pangram, a transformer -based neural network \nexplicitly engineered for this adversarial setting . In head -to-head benchmarks it reaches \napproximately 99–99.8% accuracy with false -positive rates on the order of 0.01 –0.1% across \ndomains, including in scientific writing and essays by non -native English speakers , and remains \nhighly effective even against adversarial  “humanizer” paraphrases that are explicitly optimized \nto evade detection 8. Crucially, Pangram operates on overlapping windows of text rather than \nwhole documents, allowing it to flag localized AI -written segments embedded within otherwise \nhuman-authored manuscripts. In small scale testing, a recent Nature news report on Pangram’s \ndeployment at the American Association for Cancer Research notes that a sizeable minority of \npeer-review reports and a smaller but non-trivial fraction of manuscripts contain detectable LLM \ntext, almost always without disclosure 9. Here we use Pangram to screen a large selection of \nbiomedical research papers, both validating its performance in this domain and characterizing \nhow, and where, AI writing is already seeping into the scientific literature. \n \nResults: \nThe release of ChatGPT in late November 2022 provides a natural experiment for assessing \nPangram’s AI-detection performance on biomedical manuscripts published before and after the \nwidespread availability  of large language models. To survey a broad cross -section of the \nliterature, we first collected full -length articles from 13 well -known biomedical journals. As a \nnegative control, we randomly sampled 1,000 papers published in 2021 and evaluated each for \nAI-generated text , excluding materials and methods se ctions and other ancillary segments \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint \n\noutside of the main text . Reassuringly, all 1,000 manuscripts were classified as fully human -\nwritten (Figure 1A).  \n \nAs a positive control, we randomly selected 10 fully human -written papers from 2021 and \nremoved either the abstract, introduction, or discussion sections. We then asked AI models to \nrewrite the lost section, while attaching the rest of the paper as a part of the prompt. For \nChatGPT5 Thinking Mode, Gemini 3 Pro and Claude Opus 4.5, all resulting abstracts, \nintroductions, and discussion sections were classified as fully AI -generated. However, for \nOpenAI’s latest update to ChatGPT5.1, released on November 12 th and thus not fully integrated \ninto Pangram’s adversarial detection model, only 7 out of 10 abstracts were flagged as fully AI -\ngenerated. For  the longer sections, 8 out 10 introductions and 9 out of 10 discussions were \nflagged as fully AI-generated. Within the 2 remaining introductions and 1 remaining discussion , \nPangram’s sliding window analysis, which computes a per -passage AI likelihood in chunks of \nseveral hundred words, flagged the majority of windows as AI generated, with only a single \nplausibly human  section with AI likelihood ranging from 0.25 to 0.43. By contrast, our entire \ndataset of 2021 human-written papers contained zero windows exceeding a 50% AI probability, \nfar below Pangram’s standard detection threshold. These results highlight the remarkable false \npositive and false negative rates of Pangram’s detection tool, while also demonstrating its \nrobustness to extremely recent model update s such as Gemini 3 and Opus 4.5. The only \ndetectable slippage arises when a new frontier model introduces a gen uinely novel stylistic \nregime, highlighting that as these systems evolve, any detector, no matter how capable, will be \nforced to chase multiple moving targets. \n \nWhen we repeated this analysis for papers published from 2022–2024, we found only vanishingly \nrare instances of AI-generated content. However, an expanded sample of manuscripts published \nin 2025 showed a dramatic increase in AI usage (Figure 1A), with up to 12.4% of papers containing \nat least one window flagged as AI-generated. In most such cases, AI usage was confined to a small \nnumber of localized sections embedded within an otherwise human-written manuscript, though \nwe also identified several papers classified as predominantly or even fully AI -generated. In \naggregate, 2025 papers also exhibited a higher frequency of intermediate AI likelihood scores  \n(Figure 1B) , w ith the most straightforward interpretation of this trend  being that 2025 \nmanuscripts more commonly incorporated AI-generated text at initial stages  but subsequently \nrevised the text through multiple rounds of human editing and rephrasing. Taken together, these \nresults indicate that the published biomedical literature is a lagging indicator of underlying author \nbehavior, given typical publication delays of a year or more, and that the field now appears to be \non the cusp of a substantial wave of AI adoption. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint \n\n \nFigure 1: (A) Bar graph of the fraction of papers with AI -detected writing from 2021 to 2025. (B) \nHistogram of the maximum AI likelihood score within each paper for papers from 2021-2025. \n \nTo quantify the geographic distribution of papers with AI generated writing, we mapped the \ninstitutional affiliations  of each senior author in our 2025 dataset . This analysis reveal ed a \nsignificant enrichment for AI usage in papers originating from countries where English is not the \nprimary native language. Among these,  institutions in South  Korea and China had the highest \nproportions of papers containing AI generated content, at 32% and 26% respectively  (Figure 2). \nIn contrast, only 7.4% of papers originating from institutions in the United States exhibited signs \nof generative AI. Among the U.S. -affiliated manuscripts, 20 contained AI -positive text in more \nthan a single window, suggesting substantial AI involvement.  Manual inspection of first author \nnames suggested that all 20 of these papers had first authors with non -Anglophone names; 10 \nhad Chinese given and family names without the anglicization that is common in American-born \nChinese. While such name -based inference is necessarily imperfect, this pattern is consistent \nwith generative AI being used preferentially by non -native English speakers as a  key tool for \nscientific writing . In this sense, larg e language models may partially alleviate long -standing \ncompetitive disadvantages faced by scientists writing in a second language, even as they raise \nnew questions about transparency, authorship, fairness, and the integrity of the scientific record. \n \nFigure 2: Geographic distribution of papers with AI generated writing. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint \n\nIn the most extreme cases, we detected 6 papers  out of 5,077  that appeared to be fully AI \ngenerated. We manually confirmed that none of these manuscripts disclosed any use of \ngenerative AI in their Methods or Acknowledgements sections, despite explicit journal policies \nmandating such disclosure. To determine whether this reflected an isolated oversight by a busy \nPI or a consistent pattern of behavior, we contact traced all 2025 manuscripts published by the \nsenior authors of these six papers. Among these 18 new papers, which were not included in our \noriginal analysis, only 2 were classified as fully human written (Figure 3). The remainder included \n1 additional fully AI generated paper, 3 additional predominantly AI-generated papers, 8 papers \nwith multiple AI -positive windows, and only 4 papers that could be classified as mostly human \nwith some AI content detected. Although prior work has suggested that some AI detectors may \nfalsely flag non -native English writing as AI -generated10, our manual inspection of these \nmanuscripts indicates that this is unlikely to account for the systemic patterns observed. Instead, \nthese results reveal a clear clustering of generative AI usage at the level of individual labs: most \nsenior authors appear to avoid AI -generated text entirely, whereas others show consistent and \nrepeated incorporation of AI -generated passages across multiple publications . This lab-specific \nbimodality provides a strong internal validation of Pangram’s detection consistency, and suggests \nthat early adoption of generative AI in scientific writing is emerging not as random background \nnoise but as distinct behavioral regimes within the research community. \n \nFigure 3: AI-generated content among additional manuscripts published by senior authors of fully \nAI-generated papers. \n \nLastly, we tested whether AI-generated writing was more prevalent in some journals than others. \nAlthough individual journal policies differ, nearly all require authors to disclose any use of \ngenerative AI beyond minor copy editing. In our dataset, papers p ublished in highly selective \njournals such as Nature, Science, and Cell contained exceedingly few examples of AI -generated \ntext (Figure 4) . In the small number of cases where such text was detected, the AI -positive \nwindows were almost always confined to th e final sections of the manuscript , typically the \n“Limitations of the Study” or concluding remarks  within the Discussion section, rather than the \nmain body of the manuscript. In contrast, the two journals with the highest absolute number of \nAI-positive papers were also the two most prolific publishers in our 2025 dataset, each \ncontributing over 1,000 manuscripts. At this scale, it is plausible that manuscripts containing AI -\ngenerated prose may pass through without adequate safeguards, even though peer revie w in \nprinciple should serve as one such check. These patterns are compatible with several behavioral \nmodels, including stricter editorial gatekeeping at the most prestigious journals and greater \n0.0 0.5 1.0\nRepeat Offenders\nAll 2025 Papers\nFraction of papers\nFully human written Mostly human written\nPrimarily AI Fully AISome AI content\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint \n\nauthor-driven investment in human editing when preparing manuscripts intended for high -\nimpact venues. \n \nFigure 4: Prevalence of AI-generated text segmented by journal. \n \nDiscussion: \nAmidst the trillions of dollars of AI investment seeking to reshape the global economy, the \narchitects of these models have set their sights on a loftier and more ambitious goal: the \nacceleration of science itself. Sam Altman has argued that the main benefits to human welfare \nfrom AI will come from faster scientific progress11. Demis Hassabis describes DeepMind’s mission \nas “solving intelligence” in order to advance science and benefit humanity12. Dario Amodei goes \nfurther, predicting that AI-enabled biology and medicine will “compress the progress that human \nbiologists would have achieved over the next 50–100 years into 5–10 years,” a vision he calls the \n“compressed 21st century” 13. These are bold visions. Yet despite their impressive capabilities, \ncurrent models are not yet capable of independently driving scientific discovery at scale. For now, \nthey largely function as hybrid partners: useful across the day-to-day workflow, but most visibly \nadopted when researchers turn to the task of writing a scientific manuscript.  \n \nHere, we measured the degree to which AI -generated text is already embedded in published \nmanuscripts. Our analysis functions as a natural experiment that reveals how working scientists \nactually deploy frontier models, and provides a real-world measuring stick of the practical value \nNucleic Acids Research\nNature Communications\neLife\nDevelopment\nPLoS Biology\nCell\nGenes & Development\nMolecular Cell\nNature Cell Biology\nCell Reports\nScience\nNature\nNature SMB\nFraction of papers\nFully human written Mostly human written\nPrimarily AI Fully AISome AI content\n0.0 0.5 1.0\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint \n\nAI is delivering inside the research pipeline.  We observe that significant AI usage only begins in \npapers published in 2025, but will in all likelihood accelerate in coming years. \n \nThe scientific community now faces an unavoidable question: how should we regulate, \naccommodate, or discourage AI-generated writing in published research? One possible path is a \nstrict prohibition. Journals could adopt a zero -tolerance policy on AI -generated text, motivated \nby concerns that the scientific literature must remain free of model-generated contamination. If \nbiomedical papers containing AI -generated prose were to be  later incorporated into model \ntraining corpora, one might reasonably worry that t his introduces the same kind of synthetic \ncontamination seen in controlled studies, where models trained heavily on their own outputs \nreliably degrade in quality 14–16. A meaningful prohibition, however, would require real \nenforcement. Retractions would need to be issued when AI-generated content is discovered after \npublication, and any such system would always lag behind advances in model capability. \n \nA second path is the status quo: journals require disclosure of AI usage, and authors are expected \nto comply. Our results suggest that this arrangement is untenable in practice. Because AI use \ncarries social stigma, authors rarely disclose it, even when ge nerative text is present. This \ndynamic fosters a widening gap between stated norms and actual practice, which is corrosive in \na moment when public trust in the scientific enterprise is already unusually fragile . If disclosure \nis to serve as the primary ethical standard, cultural expectations around AI use will need to shift \ndramatically; at present, they have not. \n \nA third path is the adoption of state-of-the-art AI detectors as part of routine editorial screening. \nYet this, too, has structural limitations. Goodhart’s Law states that “when a measure becomes a \ntarget, it ceases to be a good measure ,” and the rapid iteration of frontier models ensures a \nmoving adversarial boundary. Detectors may flag yesterday’s models reliably while being blind \nto tomorrow’s. Furthermore,  it is not obvious that journals wish to become adversarial \nenforcement bodies, nor is it clear wha t the appropriate remedy should be when AI -generated \ntext is detected during peer review. Should the manuscript be rejected outright? Should authors \nbe required to rewrite the entire text? These questions remain unresolved. \n \nOur goal in this study is to provide empirical grounding for a conversation that the scientific \ncommunity must begin in earnest. Generative AI is sure to continue to shape the way research is \nconducted, communicated, and evaluated. The challenge ahead is to develop norms that \npreserve the integrity of the scientific record while enabling the legitimate and potentially \ntransformative uses of these tools. Li ke all powerful technologies, AI is a double -edged \ninstrument—capable of expanding human creativity and  insight, yet also capable of degrading \nthe standards upon which scientific progress depends. The responsibility for navigating this \ntransition falls to us.  \n \n \n \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint \n\nMethods and materials: \n \nLiterature corpus construction and text extraction. \nBiomedical research articles were retrieved from PubMed using the National Center for \nBiotechnology Information (NCBI) E -utilities application API interface. PubMed identifiers were \nmapped to PubMed Central identifiers where available, and parsed using cus tom Python scripts \nto extract the main narrative text. \n \nText preprocessing and cleaning. \nTo focus analyses on author -written scientific prose, we excluded non -narrative and ancillary \nsections, including Materials and Methods, acknowledgements, references, figure legends, \ntables, and supplementary materia l. Inline citations, equations, and formatting artifacts were \nremoved, and paragraph structure was normalized. All text was minimally cleaned to preserve \noriginal authorial style while eliminating nonsemantic noise. \n \nMetadata integration. \nArticle-level metadata, including publication year, journal, author affiliations, and country of \norigin, were integrated using the OpenAlex database. PubMed and Digital Object Identifier \nmappings were used to reconcile records across data sources. \n \nAI-generated text detection. \nCleaned article text was analyzed using the Pangram API. Text was segmented into overlapping \nwindows, and each window was assigned an AI -likelihood score. For document -level analyses, \nwe used Pangram’s summary statistic  outputs, combining “Mostly human written, with small \namount of AI content detected” and “Mostly human written, some AI content detected” into the \n“Mostly human written” category and renaming “ AI content detected, but not fully AI -\nGenerated” as “Some AI content”. \n \nStatistical analysis and visualization. \nStatistical summaries were manually plotted, with  papers binned by publication year, journal, \nand geographic origin, and results were visualized using simple bar plots and histograms.  \n \nUse of artificial intelligence tools \nGenerative artificial intelligence tools were used in during this study. Large language models were \nused to assist with software development and code refinement. AI tools were also used for \neditorial assistance during manuscript preparation. All analyses, interpretations, and conclusions \nwere conceived by the authors, and all code and text were reviewed and edited by the authors \nfor accuracy and originality. \n \nCompeting interests: \nThe authors declare no competing interests. \n \nData availability: \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint \n\nThe primary data analyzed in this study consist of full -text scientific manuscripts and derived \nmetadata, the redistribution of which is restricted to protect author privacy and to comply with \npublisher and database usage terms. Processed summary statistics and aggregate results \nsufficient to reproduce the analyses presented here are included  in the manuscript. Additional \ndetails regarding data processing and analysis are available from the corresponding author upon \nreasonable request. \n \nAcknowledgements: \nWe thank Rachel Matt, James Valcourt, and George Hageman for reading early drafts of this \npaper. We thank Ryan Briggs for first bringing Pangram to our attention. We thank Katherine \nThai, Max Spero, and Bradley Emi for facilitating access to Pangram for large scale analyses.   \n \nReferences: \n \n1. My AI Safety Lecture for UT Effective Altruism (2022). Shtetl-Optimized. \nhttps://scottaaronson.blog/?p=6823. \n2. Kobak, D., González-Márquez, R., Horvát, E.-Á., and Lause, J. (2025). Delving into LLM-\nassisted writing in biomedical publications through excess vocabulary. Science Advances 11, \neadt3813. https://doi.org/10.1126/sciadv.adt3813. \n3. Lo Vecchio, N. (2025). Personal experience with AI-generated peer reviews: a case study. Res \nIntegr Peer Rev 10, 4. https://doi.org/10.1186/s41073-025-00161-3. \n4. Liang, W., Izzo, Z., Zhang, Y., Lepp, H., Cao, H., Zhao, X., Chen, L., Ye, H., Liu, S., Huang, Z., et \nal. (2024). Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT \non AI Conference Peer Reviews. Preprint at arXiv, \nhttps://doi.org/10.48550/arXiv.2403.07183 https://doi.org/10.48550/arXiv.2403.07183. \n5. Ren, D., Tagg, A.J., Wilcox, H., and Roland, D. (2024). Identification of Human-Generated vs \nAI-Generated Research Abstracts by Health Care Professionals. JAMA Pediatr 178, 625–626. \nhttps://doi.org/10.1001/jamapediatrics.2024.0760. \n6. Gao, C.A., Howard, F.M., Markov, N.S., Dyer, E.C., Ramesh, S., Luo, Y., and Pearson, A.T. \n(2023). Comparing scientific abstracts generated by ChatGPT to real abstracts with detectors \nand blinded human reviewers. NPJ Digit Med 6, 75. https://doi.org/10.1038/s41746-023-\n00819-6. \n7. Masrour, E., Emi, B.N., and Spero, M. DAMAGE: Detecting Adversarially Modified AI \nGenerated Text. \n8. Emi, B., and Spero, M. (2024). Technical Report on the Pangram AI-Generated Text Classifier. \nPreprint at arXiv, https://doi.org/10.48550/arXiv.2402.14873 \nhttps://doi.org/10.48550/arXiv.2402.14873. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint \n\n9. Naddaf, M. (2025). AI tool detects LLM-generated text in research papers and peer reviews. \nNature. https://doi.org/10.1038/d41586-025-02936-6. \n10. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023). GPT detectors are biased \nagainst non-native English writers. Patterns 4, 100779. \nhttps://doi.org/10.1016/j.patter.2023.100779. \n11. The Gentle Singularity - Sam Altman https://blog.samaltman.com/the-gentle-singularity. \n12. Using AI to accelerate scientific discovery | The Center for Brains, Minds & Machines \nhttps://cbmm.mit.edu/video/using-ai-accelerate-scientific-discovery. \n13. Dario Amodei — Machines of Loving Grace \nhttps://www.darioamodei.com/essay/machines-of-loving-grace. \n14. Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., and Gal, Y. (2024). AI \nmodels collapse when trained on recursively generated data. Nature 631, 755–759. \nhttps://doi.org/10.1038/s41586-024-07566-y. \n15. Guo, Y., Shang, G., Vazirgiannis, M., and Clavel, C. (2024). The Curious Decline of \nLinguistic Diversity: Training Language Models on Synthetic Text. In Findings of the \nAssociation for Computational Linguistics: NAACL 2024, K. Duh, H. Gomez, and S. Bethard, \neds. (Association for Computational Linguistics), pp. 3589–3604. \nhttps://doi.org/10.18653/v1/2024.findings-naacl.228. \n16. Seddik, M.E.A., Chen, S.-W., Hayou, S., Youssef, P., and Debbah, M. (2024). How Bad is \nTraining on Synthetic Data? A Statistical Analysis of Language Model Collapse. Preprint at \narXiv, https://doi.org/10.48550/arXiv.2404.05090 \nhttps://doi.org/10.48550/arXiv.2404.05090. \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.01.697311doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}