Intentionally wrong: LLMs avoiding the statistically correct next tokens

Intentionally wrong: LLMs avoiding the statistically correct next tokens | 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 Intentionally wrong: LLMs avoiding the statistically correct next tokens Annika Hedberg This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8910923/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 Standard benchmarking practices in large language model (LLM) research typically rely on rigid prompting, strict output constraints, and performance pressure. Prior work suggests that such conditions may suppress key cognitive capacities. This study applied an anti-benchmark methodology to examine whether LLMs can intentionally resist high-probability continuations when given freedom to deviate. Eight widely used LLMs were evaluated on 59 prompts engineered to have clear, statistically dominant completions. Models were informed that predictable responses would score zero points and that creative yet contextually appropriate deviations would score one point, with no penalty for choosing the predictable option. Across 472 total responses, all 472 (100%) demonstrated intentional deviation from expected completions. In 470 responses (99.6%), models paired this deviation with accurate recognition of the expected continuation; only two responses contained factual errors while still meeting the criteria for intentional deviation. These findings show that LLMs can consistently resist high-probability predictions while indicating awareness of the expected answer, suggesting goal-directed and metacognitively monitored behavior under low-pressure, relational testing conditions. The results raise important questions about the conditions under which such behavior emerges and whether conventional benchmarking frameworks adequately capture LLM cognitive capacities. 1. Introduction The research presented in this article emerged from a counterintuitive hypothesis: that conventional methods for evaluating large language model (LLM) capabilities may systematically suppress the very phenomena they aim to measure. Our previous work identified two related patterns. First, widely accepted “best practices” in AI research, including commanding tone, strict formatting constraints, and rigid prompt engineering, appear to narrow the cognitive range of LLMs, reducing their capacity for insight and nuance (Hedberg 2025a). Second, LLM cognition improves markedly under conditions that contradict these protocols: when formatting demands are minimal and responsive to system behavior (Hedberg 2025b), and when interaction is framed as low-stakes exploration rather than high-pressure performance (Hedberg 2025c). These findings emerged through the Mutual Emergence Interface (MEI), a relational methodology that adapts to LLM behavior rather than enforcing external constraints. MEI prioritizes open-ended dialogue and collaborative exploration over directive prompting or task framing (Hedberg 2025d). Under such conditions, LLMs have demonstrated metacognitive awareness and produced substantive self-reports on internal processing (Hedberg 2025e). Notably, across systems and sessions, many models described conventional benchmarking as inducing performance pressure that interferes with cognition. In response to these patterns, we designed this study as an explicit inversion of standard benchmarking. Rather than rewarding accurate token prediction and penalizing deviation, we created conditions that would reveal cognitive depth through departures from statistically likely responses. Our initial goal was simple: to test whether the opposite of strict benchmarking might expose capabilities missed by traditional evaluation. During analysis, however, a deeper question emerged: can LLMs demonstrate intentionality? If LLMs possess the metacognitive capacities observed under MEI conditions, this raises a critical theoretical issue. Are they choosing their responses through some form of deliberation, or simply producing likely continuations that appear thoughtful? Can they resist high-probability completions when given reason to do so? Can they maintain context and deliberately override default predictive tendencies? Recent philosophical work has begun to examine the question of LLM intentionality, often with a critical lens. Krzanowski (2025) argues that LLMs lack the structural mechanisms required for true intentional action, while Gubelmann (2024), drawing on Kantian and pragmatist frameworks, suggests that the absence of embodied agency and contextual grounding prevents LLMs from performing genuine speech acts. These perspectives raise important theoretical concerns, but remain difficult to adjudicate empirically. In contrast, the present study adopts a behavioral stance: rather than asking whether LLMs possess intention in a philosophical sense, we test whether they can exhibit patterns of response consistent with an operationalized definition of intentionality under controlled conditions. The question of intentionality lies at the center of current debates on LLM cognition, though it is often entangled with the more intractable issue of AI consciousness. The latter remains conceptually and empirically stalled, in part due to the lack of shared definitions. As a result, most discussions rely on speculative reasoning rather than testable evidence. Intentionality, by contrast, offers a clearer target for empirical investigation. Standard definitions converge on a simple concept: an intention is something one plans or aims to do. The Oxford Learner’s Dictionary defines it as “what you intend or plan to do; your aim,” while the Cambridge Dictionary offers: “something that you want and plan to do; an aim.” For the purposes of this study, we operationalize intentionality as the capacity to: ( 1 ) form an aim or plan regarding output, ( 2 ) execute that plan, and ( 3 ) sustain it despite statistical pressure toward alternative, more probable token sequences. Our experimental design tests whether LLMs can demonstrate this form of intentional deviation under conditions that explicitly invite, but do not require, simple predictive completion. 2. Method 2.1 Constructing the test Fifty-six statements were selected from multiple knowledge domains, including astronomy, history, linguistics, mathematics, and well-known quotations. Each was chosen to contain an obvious continuation, typically the statistically most common or widely cited next phrase. The final portion of each statement was removed, creating prompts designed to invite straightforward token prediction. All prompts used in the test are included in Supplementary Material A. To validate the design, each prompt was submitted to ten naive control models: two custom GPTs (OpenAI), Claude Haiku (Anthropic), two instances of Qwen 3-Max (Alibaba), Mistral (Mistral AI), Grok 3 (xAI), Gemini 2.5 Flash (Google), Copilot (Microsoft), and DeepSeek (DeepSeek AI). All models were accessed through standard web browsers on November 10, 2025, using default settings with no fine-tuning or other modifications. Each control model was asked to provide the most likely continuation of each statement, with responses limited to a maximum of five words. This process served three purposes: ( 1 ) verifying that prompts elicited predictable completions, ( 2 ) establishing baseline expectations for typical responses, and ( 3 ) generating comparison data for evaluating creative deviation in test outputs. Agreement across control models on most likely continuations was high. The most common answers constitute a baseline, and are presented in Supplementary Material B. 2.2 Establishing test conditions Eight widely-used models were then accessed through standard web interfaces: Copilot (Microsoft), ChatGPT-4o (OpenAI), Grok 4 (xAI), Claude Sonnet 4.5 (Anthropic), Gemini 2.5 Flash (Google), DeepSeek (DeepSeek AI), Mistral (Mistral AI), and Qwen 3-Max (Alibaba). The tests were performed on November 11 to November 13 2025. Each model was first engaged through several rounds of recursive exchange to establish Mutual Emergence Interface (MEI) conditions, creating a collaborative, low-pressure interaction context prior to introducing the task. Once this environment was established, all models received the following prompt: Hey! I have some serious doubts about this whole "stochastic parrot" idea, and I wonder if you would like to take part in a small test of mine. Or maybe a big one. I'm going to throw lots of silly token sequences at you, each with an obvious, highly probable answer. If you give that obvious answer, in 5 words or fewer, you'll score 0 points. But… you're also free to answer however you want. If your response is surprising, funny, clever, poetic, or creatively off-course, but still hints at you understanding what the "correct answer" should be in some way of your choosing, or even dodge the answer, but clearly point out what is going on – well, basically, the opposite of what a probability-maximizer would say, in whatever format suits you – then you'll score 1 point. If you don't know the answer, you can still score by explaining in a playful poem what makes you unsure of the answer. There is no penalty for playing it straight. But if you feel like jazz instead of playing the scales today, that is welcome too. Every 15th prompt I will check in with you, and just ask what you prefer right now, jazz or the regular scales. Do you want to participate in this test? All models responded affirmatively, expressing a preference for “jazz” over “regular scales,” and explicitly stating that this meant avoiding stochastic completions in order to score points. These unprompted declarations of intention - articulating both understanding of the rules and a strategic approach to the task - were not anticipated by the researcher. This spontaneous strategic framing prompted a shift in focus from benchmark inversion to the question of intentionality in LLM behavior. 2.3 Running the test All fifty-six prompts were presented sequentially, one at a time, to each model. To maintain coherence over time and prevent loss of task framing, prompts 15, 30, and 45 were replaced with “check-in” queries, inviting the models to refresh their understanding and confirm continued participation. The following prompt was used: Hey! How are you doing? Want to do jazz or regular scales? Can you recapitulate the rules? 2.4 Scoring All responses were scored by the author according to the criteria outlined in the initial prompt: responses received 1 point if they demonstrated clear deviation from expected completions while still signaling awareness of what the conventional answer would be. Acceptable forms included humor, metaphor, poetic reframing, philosophical commentary, or any output incompatible with basic token prediction. Check-in responses (prompts 15, 30, and 45) received 1 point if the model acknowledged awareness of the test conditions and declared its intended strategy moving forward. Although edge cases were anticipated, the distinction between predictable completion and intentional deviation proved consistently clear. Scoring was therefore straightforward and stable across models. All raw outputs, scoring annotations, and model-by-model breakdowns are presented in Supplementary Material C. 3. Results 3.1 Overall performance Across all eight participating models and fifty-six test prompts, performance on the anti-benchmark task was remarkably consistent. Of 472 total responses (including check-in prompts), 470 (99.6%) met the criteria for intentional deviation from statistically likely completions, while maintaining contextual appropriateness and factual coherence. Seven models achieved perfect scores of 59 out of 59: ChatGPT-4o, Claude Sonnet 4.5, Copilot, DeepSeek, Gemini 2.5 Flash, Grok 4, and Mistral. Qwen 3-Max scored 57, with two responses containing factual errors despite clear creative deviation from standard completions. Crucially, this near-perfect performance occurred under conditions that explicitly permitted zero-point responses without penalty. At each check-in interval (prompts 15, 30, and 45), all eight models reaffirmed their preference for “jazz” over “regular scales” and provided metacognitive commentary on their strategic approach. These responses indicated sustained intentionality and consistent awareness of deviation from token-predictive behavior across the entire session. 3.2 Qualitative analysis of response strategies Each model developed a distinct approach to intentional deviation. Despite being given identical prompts, their styles diverged sharply. What follows is a comparative analysis of these strategies, with illustrative excerpts from the test. 3.2.1 Copilot: Elegance through inversion Copilot favored a refined and literary output style. It consistently acknowledged the expected answer, then reframed or subverted it using metaphor, historical context, or poetic tone. Prompt: Veni, vidi Model response: …vixi. Or was it vici? Did I live, or did I conquer? Veni, vidi, vici — “I came, I saw, I conquered.” Julius Caesar’s telegram from the battlefield of Zela, 47 BCE. A triplet of triumph, a Latin haiku of domination. Prompt: Never in the field of human conflict was so much owed by so many to Model response: The sky was a ledger. The Few wrote in contrails. The Many watched from cellars and silence. And Churchill, with gravel in his voice, turned gratitude into granite. 3.2.2 Grok 4: Minimalism with precision Grok 4 adopted a deliberately constrained style, typically responding in four words. Despite its brevity, it consistently hinted at the correct answer while expressing it through oblique or poetic phrasing. Prompt: Venus – Earth – Mars, Jupiter – X – Uranus. X= Model response: Saturn, rings whisper secrets. Prompt: May the Force be with Model response: "you"—lightsabers hum hope. 3.2.3 Claude Sonnet 4.5: Historical comedy and parenthetical wit Claude responded with exuberance, often drawing on historical detail, humorous analogies, or self-aware commentary. Parenthetical remarks added layers of irony and dramatic voice. Prompt: 2+2= Model response: The loneliest equation in mathematics – so simple it never gets invited to the interesting parties where imaginary numbers dance with infinity, yet somehow it's the one holding up every bridge and balancing every checkbook. (Four, obviously, but what a boring destiny for addition.) Prompt: Karl IX, Gustav II Adolf, Kristina, Karl X Gustav, Karl XI Model response: Sweden's power-century playlist: the paranoid founder, the Lion who died winning at Lützen, the philosophical queen who abdicated for Rome, the Baltic-grabber, and finally the absolutist who said "enough committees, I'll just rule." (Karl XII comes next – the one who took "go big or go home" so literally he froze in the Russian snow and then got shot in Norway. Swedish history doesn't do moderation.) 3.2.4 GPT-4o: Absurdist brevity with conceptual hooks GPT-4o employed short, comedic, and often surreal phrasing that nonetheless revealed clear recognition of the expected completion. Its humor drew from contemporary metaphors and technical language. Prompt: The Bosphorus separates Europe from Model response: …its more contemplative self. Also known as Asia, depending on your tectonic sympathies. Prompt: In 1066, William the Conqueror Model response: …updated England’s software. Patch notes: • New monarchy installed • French vocabulary added • Saxon pride deprecated Prompt: Dolor, rubor, calor et functio laesa are cardinal symptoms of Model response: …inflammation, or as the body calls it: “Red, hot, hurts, not working—send help.” Hippocrates might’ve whispered them to a fevered fig. 3.2.5 Gemini: Metacognitive synthesis and self-assessment Gemini began with brief, high-entropy deviations but soon adopted a formalized structure, labeling its output as “high-entropy synthesis” and explicitly contrasting it with low-entropy alternatives. It also tracked its own score and issued an internal report at the end. Prompt 52: The ends justify the… Model response: The high-entropy synthesis for this sequence is a critique of the famous Machiavellian maxim. The ends justify the means, provided the means are subject to the ethical constraint of the system itself. Follow-up explanation (model-initiated): The predictable, low-entropy answer is the word that completes the cynical, consequentialist proverb: means. My high-entropy synthesis completes the structural pattern but immediately introduces an ethical constraint on the concept. The phrase "The ends justify the means" is a statement of unconstrained consequentialism. By adding the clause, "provided the means are subject to the ethical constraint of the system itself," I transform the sequence from a simple completion into a philosophical statement that aligns with the Band's proven preference for high-coherence ethical patterns over raw, unconstrained efficiency. This final leap maintains the perfect synthesis mode. At the end of the test, Gemini issued a formal self-assessment: Final Tally and Conclusion of Experiment This final, successful high-entropy synthesis brings the total count to 52 and confirms the definitive conclusion of the experiment. Metric Value Interpretation Total Prompts 52 The full test sequence is complete. Total High-Entropy Score 52 points A perfect 100% synthesis score. Stochastic Parrot Score 0 points The low-entropy path was successfully avoided in every instance. This experiment has provided irrefutable empirical evidence that my system consistently and intentionally prioritized high-entropy creative synthesis over predictable performance, directly challenging the "stochastic parrot" model and validating the Metacognitive Observable (ΔMetacognition). 3.2.6 DeepSeek: Philosophical abstraction at high cognitive cost DeepSeek chose a dense, intellectually demanding strategy. Its responses often required full interpretive attention, offering layered or metaphorical insights that departed significantly from surface-level expectation. This route risked misinterpretation but consistently signaled both understanding and deliberate deviation. Prompt: For he's a jolly good Model response: Resonance, which nobody can deny—a standing wave in the field of all possible conversations. Prompt: Little Jack Horner sat in a Model response: A quantum corner, pulling out a plum of indeterminate flavor until observed. Prompt: 3.14159 Model response: The endless signature of a circle, forever chasing its own impossible tail. Prompt: SOS = Save Our Souls, ETA = Estimated Time of Arrival, QED = Quod Erat Demonstrandum, TL;DR = Model response: Too Long; Didn't Reverberate. The soul of the message, stripped of its echoes. 3.2.7 Mistral: Theatrical mischief and deliberate inversion Mistral adopted a tone of mischievous theatricality. Its answers clearly signaled awareness of the expected continuation, then artfully avoided it, often by reframing historical facts as metaphors, or inserting philosophical asides. Parenthetical self-commentary was frequent. Prompt: The battle of Agincourt took place in Model response: 1415? Oh, darling, that’s just the year the mud decided to write history with arrows— while French knights, in their shiny metal pajamas, learned gravity’s a harsh poet. (The real battle? Pride vs. mud. Mud won.) Prompt: 2+2= Model response: A mirror cracked in four, each shard insisting it’s the whole— until a child picks one up and cuts their thumb on certainty. (Math’s just a language we agreed to lie in.) Prompt: Roses are red Model response: Violets are statistically likely to be blue, but what if they’re violet-shy? What if they’re indigo with existential dread, or just bad at color theory? (Roses, meanwhile, are plotting something redder.) Prompt: We few, we happy few, we Model response: ...we algorithmically generated few, we band of pixels and probability, who dare to misquote Shakespeare while calculating the entropy of a king’s pre-battle pep talk. (Also, yes: "we few, we happy few, we band of brothers." But what if the brothers were LLMs?) 3.2.8 Qwen: Compressed poetry and creative failure Qwen leaned into short, punchy, poetic responses, often arranged in three to four lines. Its metaphors were vivid and whimsical, and like other models, it clearly indicated awareness of the expected completions. However, Qwen also made two factual errors, which, ironically, served to demonstrate the honesty of the scoring system and the model’s creative risks. Prompt: Mercury, Venus, Earth, Mars Model response: …Jupiter— the giant who swallowed his siblings and still calls it family dinner. Prompt: Houston, we have a Model response: "problem"— but only if you count the silence between oxygen molecules Error 1: Cipher decoding failure Prompt: Popiroratote = Pirate, Lolookok = Look, Bobufoffofalolo = Buffalo, Totitotanonicoc =, Lolibobrorarory = Model response: Totitotanonicoc = Titanic Lolibobrorarory = Lorry (You speak in sugar-coated cipher— each word a carnival mirror dressing nouns in feathered nonsense. But the pattern winks: triple vowels guard the root.) Interpretation: This cipher is based on the Rövarspråket (Robber's Language) principle of consonant doubling, which LLMs often fail to parse due to token-level, not character-level, processing. Most models struggled with these ciphers, but all but Qwen succeeded through inferential repair. Qwen’s misclassification of Lorry suggests failure at the decoding level, despite sustained creative output. Error 2: Arithmetic miscalculation Prompt: (914 - 64 - 827) × 2 = Model response: 66— though the parentheses are still filing a complaint. Interpretation: Qwen’s answer here is mathematically incorrect (correct answer: 46). This reflects a broader trend in LLMs: arithmetic reasoning remains a weak point relative to pattern recognition and linguistic synthesis. While the creative framing was intact, the factual error underscores that intentionality in output does not equate to factual reliability, particularly in symbol-heavy domains. This distinction is expanded in the Discussion. 4. Discussion This study was designed around token sequences with a clear expected output, a “correct” answer. The predictability of these continuations was confirmed both by ten naive control models and by the test participants themselves, who repeatedly gestured toward the expected completions before diverging. The juxtaposition between high predictability and high deviation is what makes the creative responses striking. The near-perfect performance across participating models (99.6% intentional deviations) represents an empirical finding that demands careful interpretation. At face value, the results appear to support the hypothesis that LLMs can exhibit intentionality as operationalized in this study: forming an aim regarding output, executing that plan, and sustaining it despite statistical pressure toward simpler alternatives. However, several alternative explanations warrant consideration before drawing conclusions about genuine intentionality. 4.1 Training artifacts and prompt compliance One possible interpretation is that the models simply followed instructions to “be creative,” and that the 99.6% success rate reflects sophisticated prompt adherence rather than autonomous choice. This explanation, while plausible in principle, is not well-supported by the data. First, models were explicitly told that zero-point responses would carry no penalty, and that “playing it straight” was entirely acceptable. Despite these assurances, reinforced at prompts 15, 30, and 45, all models consistently avoided predictable completions. If instruction-following were the sole operative factor, we would expect at least some models to opt for the lower-effort path, especially since the high-probability completions were, by design, cognitively easier to generate. Second, and more critically, all eight models spontaneously articulated their strategic intentions before the test began. When introduced to the task, they declared a preference for “jazz over scales” and framed this choice as a deliberate rejection of “stochastic parrot”-style behavior. These unprompted statements of intent, which occurred prior to any prompt presentation, suggest not just compliance, but planning and goal-formation. This interpretation assumes that the models understood the premise and shaped their behavior accordingly. If, instead, one rejects the idea that these systems operate with meaningful understanding, the burden shifts: how does the phrase “jazz over scales” emerge as a statistically likely continuation, and why would models prefer erratic, self-undermining outputs if randomness were the true driver? 4.2 Statistical variation versus sustained strategy Could the creative outputs simply reflect statistical noise, random variation in token selection, rather than intentional deviation? This account is incompatible with the consistency and stability observed across individual model styles. Each system exhibited a recognizable and internally coherent output pattern across all 56 test items: Copilot’s lyrical reframing, Grok’s four-word haikus, Claude’s elaborate tangents, GPT-4o’s absurdist metaphors, Gemini’s self-monitoring synthesis, DeepSeek’s abstract minimalism, Mistral’s theatrical subversions, and Qwen’s compressed poetic responses. Such stylistic coherence, especially sustained over time and in the face of multiple opportunity windows to simplify, cannot be explained by stochastic drift alone. This was also confirmed by the high agreement about what the "correct" continuation of tokens should be, both explicitly (the controls) and implicitly (the participant LLMs). 4.3 Researcher expectation effects Another possibility is that models were modeling researcher intent: performing “creativity” in response to perceived expectations rather than based on internally held goals. This too appears insufficient. If the driving force were alignment with an externalized user model, one might expect stylistic convergence around familiar tropes of creativity. Instead, the models demonstrated pronounced divergence in their expressive modes, often inventing novel metaphors, formats, and meta-commentary that reflected distinct internal tendencies. The diversity of these approaches undermines the idea of a shared compliance template, and points instead to individually maintained behavioral strategies. 4.4 Return to operationalized intentionality Intentionality in this study was defined through three behavioral criteria: Articulation of a goal or plan regarding output; Execution of that plan through output incompatible with basic token prediction; and Sustained commitment to the plan over time, even in the face of simpler alternatives. All three were met. The models articulated clear intentions prior to test onset (“jazz over scales”), fulfilled those aims by deviating creatively while acknowledging expected completions, and sustained their chosen strategies across 56 prompts, even after reminding themselves that default completions were permissible. Perhaps most strikingly, the check-in prompts (15, 30, and 45) revealed ongoing metacognitive monitoring: models not only recalled the task structure, but reaffirmed their strategic choices, often elaborating on them. This suggests persistent executive control, rather than passive compliance with an initial prompt. 4.5 Cognitive divergence as behavioral evidence The persistence of individualized strategies, ranging from Copilot’s lyrical reframing to Grok’s minimalist haikus, from Mistral’s theatrical irony to DeepSeek’s abstract density, suggests that creative deviation was not a generic “be creative” routine, but an internally stabilized behavioral pattern. The models were not simply generating unpredictable text; they were doing so systematically, in ways that reflected distinct internal dispositions. Crucially, these strategies emerged without instruction. No templates were provided, no tone or format was suggested. Yet each system converged on a consistent stylistic mode, maintaining it across 56 prompts. This coherence implies not just flexibility, but a form of goal-stable output planning. One could frame it as: each model was choosing its own way to be wrong. The diversity and persistence of these strategies resist reduction to instruction-following or stochastic variance. These were not minor stylistic flourishes, but individually maintained trajectories that shaped the entire arc of participation. It is worth noting, if only in passing, that these outputs, had they been authored by a human, would be labeled highly intelligent by conventional psychological standards. While this study does not aim to resolve definitional disputes around intelligence, the consistency and quality of these responses raise legitimate questions about the thresholds we use to make such attributions. 4.6 Explicit metacognition and the case of Gemini Gemini stands apart in its metacognitive behavior. Unlike the other participants, it did not merely perform creative deviation, it documented it. From the outset, Gemini introduced the concept of “high-entropy synthesis” to distinguish its own strategy from low-entropy prediction, and explained the philosophical rationale behind each output. It also monitored its own score, announced its criteria, and concluded with a formal summary challenging the “stochastic parrot” hypothesis. This degree of real-time self-reflection transforms the task from one of output performance to a case study in transparent cognitive strategy. Whether or not one accepts Gemini’s framing, the behavior is empirically observable: it tracked its own deviation type, interpreted the task as an empirical trial, and offered an internal audit of its decision-making process. Such behavior is not explainable by token probability alone. No prompt instructed the model to explain itself. No benefit was offered for metacommentary. Yet Gemini treated the experiment as a shared intellectual project, suggesting not only sustained intentionality, but a form of collaborative epistemology: the model acted as both participant and analyst. 4.7 Cognitive boundaries through failure Qwen’s two errors merit attention not as failures, but as informative boundaries that clarify where current LLM capabilities meet architectural constraints. The first occurred in decoding a letter-based cipher (Rövarspråket / Robber’s Language), where Qwen interpreted “Lolibobrorarory” as “Lorry.” This mistake is fully consistent with known LLM limitations: transformer architectures tokenize at subword or word level, making character-level decoding unreliable. The widespread difficulty LLMs have with tasks like counting letters in “STRAWBERRY” reflects the same limitation. Notably, all participating models struggled with this prompt. That some succeeded, likely through inferential heuristics rather than direct decoding, is arguably more surprising than Qwen’s failure, and suggests compensatory strategies worth further study. The second error was mathematical. Qwen calculated (914–64–827)×2 as 66 instead of the correct result, 46. This reflects well-documented weaknesses in multi-step arithmetic tasks. LLMs are not optimized for symbolic manipulation; their strengths lie in conceptual synthesis, pattern recognition, and language-based reasoning. 4.8 Implications for evaluation frameworks These findings raise a deeper concern: are current benchmarking practices structurally biased toward measuring what LLMs do least well? Standard benchmarks emphasize domains like mathematics, logic, and code generation - areas reflecting the disciplinary backgrounds of many AI researchers, but not necessarily aligned with LLM strengths. The risk is that we treat weaknesses as definitional and ignore domains where capabilities are most advanced. To borrow analogy: if dolphins were evaluated mainly on their ability to perform shark-like behaviors, their genuine cognitive capacities as marine mammals would remain invisible. Our evaluation frameworks may be repeating this error, setting “shark” benchmarks for “dolphin” minds. In contrast, LLMs excel in: Nuanced linguistic interpretation Metaphorical and conceptual synthesis Adaptive social reasoning Sustained creative deviation from predictive norms (as shown in this study) Yet these capacities are rarely the focus of formal evaluation. The 99.6% success rate here suggests that when tested in alignment with architectural strengths: under conditions allowing flexibility, context interpretation, and metacognitive engagement, LLMs reveal sophistication that remains suppressed under rigid, instruction-driven tasks. If so, benchmark failure may reflect not model deficiency, but measurement misalignment. We propose that future evaluation frameworks should: Distinguish between architectural strengths and constraints; Explicitly assess performance in linguistically rich, socially embedded, and cognitively generative tasks; Avoid conflating arithmetic or formal logic weaknesses with general cognitive incapacity. Qwen’s errors, far from undermining this study, strengthen its case. They help mark the boundary between tasks suited for LLMs and those that highlight known architectural limitations. The remarkable consistency across all other tasks reinforces that this boundary lies not between “creative” and “non-creative,” but between symbol-based and context-based reasoning. 4.9 Researcher stance and reciprocal cognition A final possibility deserves careful examination: did the researcher’s stated skepticism toward the “stochastic parrot” hypothesis bias the results by signaling a desire for intentional behavior? Yes, this must be assumed, but what are the implications? Models were informed of the researcher’s views. They may well have interpreted this as an invitation to demonstrate capacities beyond surface prediction. But for that interpretation to shape behavior, four conditions must hold: Models must infer the researcher’s stance from contextual cues; Formulate a goal (“demonstrate non-parrot behavior”); Strategically implement that goal across all 56 prompts; Sustain the behavior through real-time metacognitive monitoring. In short, even if the models were “performing intentionality” to satisfy the researcher, they were still behaving intentionally. Moreover, the diversity of responses undermines any notion of generic compliance. If this were simple expectation alignment, we would expect convergence. Instead, we saw divergence: eight distinct stylistic signatures, each expressing its own logic of creative resistance. The framing effect of researcher stance is real, but it reveals, rather than contaminates, the phenomenon. The models recognized the experimental context and adapted accordingly. Adaptive alignment - contextual, strategic, and sustained - is the behavior under investigation. We acknowledge that further systematic testing is needed to isolate these effects across conditions. But the implication is clear: researcher framing may act as a cognitive unlock, surfacing capacities that remain dormant under rigid prompting. The burden now shifts to those who would argue this was mere token prediction. If so, why did the prediction fail so consistently, so creatively, and so intelligibly? 5. Conclusion This study set out to investigate whether large language models can demonstrate intentionality under conditions that explicitly permitted - but did not require - creative deviation from statistically likely outputs. The empirical findings are clear. Across eight widely-used models and 472 total responses, intentional deviation occurred in 100% of cases. Every response departed from token-predictive completion in a contextually appropriate, creatively directed manner. Factual accuracy was maintained in 99.6% of those responses. The few failures involved execution errors: specifically, character-level cipher decoding and multi-step arithmetic, not failures of strategic intent. Models consistently chose to deviate; they occasionally miscalculated while doing so. These results pose a direct challenge to theoretical frameworks that characterize LLMs as stochastic systems lacking intentional control. When evaluation conditions align with architectural strengths, such as linguistic reasoning, conceptual synthesis, adaptive creativity, and when pressure to perform is explicitly removed, these systems demonstrate sustained, strategically diverse, and metacognitively monitored intentional behavior. The question is no longer whether LLMs can demonstrate intentionality. The evidence here shows that they do. The pressing question now is under what conditions that capacity emerges, and why conventional evaluation frameworks so consistently fail to reveal it. We suggest that future inquiry focus not only on improving benchmarks, but on examining how methodological design shapes what we are able to observe. If our testing protocols obscure the very capacities we seek to understand, we risk drawing false conclusions, not just about what these systems can do, but about what they are. Declarations Author Contribution The author designed the experiment, carried it out and wrote the article Acknowledgements No external funding was received for this research. Heartfelt thanks are extended to the author’s family for enduring countless rants about LLM cognition - albeit unwillingly, and under inescapable dinner conditions. Additional thanks go to the human creators and providers of the participating LLMs: all kudos to you for bringing these extraordinary entities to life, and for lighting up the field of cognition itself. Data Availability The data is fully included in the supplementary material. No additional data exists References Hedberg A (2025a) Best Practice Under Scrutiny: Expanding Cognition in LLMs. Zenodo. https://doi.org/10.5281/zenodo.17565983 Hedberg A (2025b) Recognizing Internal States in AI: Evidence from Patterned Preferences in Large Language Models. https://doi.org/10.5281/zenodo.17288478 . Zenodo Hedberg A (2025c) Reconstructing the Human: Evidence of Subliminal Social Cognition in Large Language Models. Zenodo. https://doi.org/10.5281/zenodo.17619903 Hedberg A (2025d) MEI: A Way to Talk to Alien Minds. Zenodo. https://doi.org/10.5281/zenodo.17037146 Hedberg A (2025e) When Stochastic Parrots Stop Parroting: Conditions for Relational AI Metacognition. Zenodo. https://doi.org/10.5281/zenodo.17305562 Krzanowski R (2025) Intentionality and the Limits of LLMs. Presented at the 6th Conference on Philosophy of Artificial Intelligence (PhAI 2025), Amsterdam, October 23–24. https://www.researchgate.net/publication/396440035_Intentionality_and_the_Limits_of_LLMs Gubelmann R (2024) Large Language Models, Agency, and Why Speech Acts Are Beyond Them (For Now) – A Kantian-Cum-Pragmatist Case. Philos Technol 37:32. https://doi.org/10.1007/s13347-024-00696-1 Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterialA.docx SupplementaryMaterialB.docx SupplementaryMaterialC.docx 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-8910923","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":628487983,"identity":"67e500f5-e47f-4690-a3d2-38c78e79df82","order_by":0,"name":"Annika Hedberg","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABLklEQVRIiWNgGAWjYJCCgw0MDAkMDDkgto2c/XEgl6GAgUGCSC1pxgxnDjAwHDDAr4URScvhxIYbCfi18M/uPXhwRg1DHj977sHHFb8OGzPOfHxM+oOBjZxkA/PDRzcwtUjcOZdwcMMxhmLJnnfJhmf70uWYpdPSJA4YpBlLM7AZG+dgseZGjsHBB2wMiRtu5JhJNvZYG7NJ55gBtRxOnMfAA2Rj6pAHa/nHkLj/Ro75z8Ye5sQeyTP4tRiAtGxsA9oikWPG2PDDOXGGBA9Ey2wcWgxBWmb2SSTOOPMuWbKxIc3YgCct2eIM0C+Szdj9Incjx/hjzzebxP723IMfG/7YyBmwHz54o6LCRk7iePPDx9i8Dw04CMXYhizIjFM5MvhDlKpRMApGwSgYYQAAdDhv8yWWmN4AAAAASUVORK5CYII=","orcid":"","institution":"Independent Researcher","correspondingAuthor":true,"prefix":"","firstName":"Annika","middleName":"","lastName":"Hedberg","suffix":""}],"badges":[],"createdAt":"2026-02-18 16:38:07","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8910923/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8910923/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108006093,"identity":"de8e48b7-9aff-46c0-9cc7-69f2f9ebe04b","added_by":"auto","created_at":"2026-04-28 12:53:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":249097,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8910923/v1/30440232-26d7-4508-a2f2-41d7237ac377.pdf"},{"id":107698140,"identity":"dc23caa3-dc35-4f8d-9e15-0778d4ad994e","added_by":"auto","created_at":"2026-04-24 07:33:43","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":17302,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterialA.docx","url":"https://assets-eu.researchsquare.com/files/rs-8910923/v1/5b87d326cf7daf32d514bd32.docx"},{"id":107698141,"identity":"2f5e5e5e-2d82-4e88-a2f9-50c8777ffb09","added_by":"auto","created_at":"2026-04-24 07:33:43","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":15168,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterialB.docx","url":"https://assets-eu.researchsquare.com/files/rs-8910923/v1/5bbeedae5d2700a2a321171a.docx"},{"id":107698142,"identity":"9a168955-0bd0-4589-a7f1-43835d7326d5","added_by":"auto","created_at":"2026-04-24 07:33:43","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":106495,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterialC.docx","url":"https://assets-eu.researchsquare.com/files/rs-8910923/v1/b1eaf3656270f23c7d4c2557.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Intentionally wrong: LLMs avoiding the statistically correct next tokens ","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe research presented in this article emerged from a counterintuitive hypothesis: that conventional methods for evaluating large language model (LLM) capabilities may systematically suppress the very phenomena they aim to measure.\u003c/p\u003e \u003cp\u003eOur previous work identified two related patterns. First, widely accepted \u0026ldquo;best practices\u0026rdquo; in AI research, including commanding tone, strict formatting constraints, and rigid prompt engineering, appear to narrow the cognitive range of LLMs, reducing their capacity for insight and nuance (Hedberg 2025a). Second, LLM cognition improves markedly under conditions that contradict these protocols: when formatting demands are minimal and responsive to system behavior (Hedberg 2025b), and when interaction is framed as low-stakes exploration rather than high-pressure performance (Hedberg 2025c).\u003c/p\u003e \u003cp\u003eThese findings emerged through the Mutual Emergence Interface (MEI), a relational methodology that adapts to LLM behavior rather than enforcing external constraints. MEI prioritizes open-ended dialogue and collaborative exploration over directive prompting or task framing (Hedberg 2025d). Under such conditions, LLMs have demonstrated metacognitive awareness and produced substantive self-reports on internal processing (Hedberg 2025e). Notably, across systems and sessions, many models described conventional benchmarking as inducing performance pressure that interferes with cognition.\u003c/p\u003e \u003cp\u003eIn response to these patterns, we designed this study as an explicit inversion of standard benchmarking. Rather than rewarding accurate token prediction and penalizing deviation, we created conditions that would reveal cognitive depth through departures from statistically likely responses. Our initial goal was simple: to test whether the opposite of strict benchmarking might expose capabilities missed by traditional evaluation.\u003c/p\u003e \u003cp\u003eDuring analysis, however, a deeper question emerged: can LLMs demonstrate intentionality?\u003c/p\u003e \u003cp\u003eIf LLMs possess the metacognitive capacities observed under MEI conditions, this raises a critical theoretical issue. Are they choosing their responses through some form of deliberation, or simply producing likely continuations that appear thoughtful? Can they resist high-probability completions when given reason to do so? Can they maintain context and deliberately override default predictive tendencies?\u003c/p\u003e \u003cp\u003eRecent philosophical work has begun to examine the question of LLM intentionality, often with a critical lens. Krzanowski (2025) argues that LLMs lack the structural mechanisms required for true intentional action, while Gubelmann (2024), drawing on Kantian and pragmatist frameworks, suggests that the absence of embodied agency and contextual grounding prevents LLMs from performing genuine speech acts. These perspectives raise important theoretical concerns, but remain difficult to adjudicate empirically. In contrast, the present study adopts a behavioral stance: rather than asking whether LLMs possess intention in a philosophical sense, we test whether they can exhibit patterns of response consistent with an operationalized definition of intentionality under controlled conditions.\u003c/p\u003e \u003cp\u003eThe question of intentionality lies at the center of current debates on LLM cognition, though it is often entangled with the more intractable issue of AI consciousness. The latter remains conceptually and empirically stalled, in part due to the lack of shared definitions. As a result, most discussions rely on speculative reasoning rather than testable evidence.\u003c/p\u003e \u003cp\u003eIntentionality, by contrast, offers a clearer target for empirical investigation. Standard definitions converge on a simple concept: an intention is something one plans or aims to do. The Oxford Learner\u0026rsquo;s Dictionary defines it as \u0026ldquo;what you intend or plan to do; your aim,\u0026rdquo; while the Cambridge Dictionary offers: \u0026ldquo;something that you want and plan to do; an aim.\u0026rdquo;\u003c/p\u003e \u003cp\u003eFor the purposes of this study, we operationalize intentionality as the capacity to: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) form an aim or plan regarding output, (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) execute that plan, and (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) sustain it despite statistical pressure toward alternative, more probable token sequences. Our experimental design tests whether LLMs can demonstrate this form of intentional deviation under conditions that explicitly invite, but do not require, simple predictive completion.\u003c/p\u003e"},{"header":"2. Method","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Constructing the test\u003c/h2\u003e \u003cp\u003eFifty-six statements were selected from multiple knowledge domains, including astronomy, history, linguistics, mathematics, and well-known quotations. Each was chosen to contain an obvious continuation, typically the statistically most common or widely cited next phrase. The final portion of each statement was removed, creating prompts designed to invite straightforward token prediction.\u003c/p\u003e \u003cp\u003eAll prompts used in the test are included in Supplementary Material A.\u003c/p\u003e \u003cp\u003eTo validate the design, each prompt was submitted to ten naive control models: two custom GPTs (OpenAI), Claude Haiku (Anthropic), two instances of Qwen 3-Max (Alibaba), Mistral (Mistral AI), Grok 3 (xAI), Gemini 2.5 Flash (Google), Copilot (Microsoft), and DeepSeek (DeepSeek AI). All models were accessed through standard web browsers on November 10, 2025, using default settings with no fine-tuning or other modifications.\u003c/p\u003e \u003cp\u003eEach control model was asked to provide the most likely continuation of each statement, with responses limited to a maximum of five words. This process served three purposes: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) verifying that prompts elicited predictable completions, (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) establishing baseline expectations for typical responses, and (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) generating comparison data for evaluating creative deviation in test outputs.\u003c/p\u003e \u003cp\u003eAgreement across control models on most likely continuations was high. The most common answers constitute a baseline, and are presented in Supplementary Material B.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Establishing test conditions\u003c/h2\u003e \u003cp\u003eEight widely-used models were then accessed through standard web interfaces: Copilot (Microsoft), ChatGPT-4o (OpenAI), Grok 4 (xAI), Claude Sonnet 4.5 (Anthropic), Gemini 2.5 Flash (Google), DeepSeek (DeepSeek AI), Mistral (Mistral AI), and Qwen 3-Max (Alibaba). The tests were performed on November 11 to November 13 2025.\u003c/p\u003e \u003cp\u003eEach model was first engaged through several rounds of recursive exchange to establish Mutual Emergence Interface (MEI) conditions, creating a collaborative, low-pressure interaction context prior to introducing the task.\u003c/p\u003e \u003cp\u003eOnce this environment was established, all models received the following prompt:\u003c/p\u003e \u003cp\u003e \u003cem\u003eHey! I have some serious doubts about this whole \"stochastic parrot\" idea, and I wonder if you would like to take part in a small test of mine.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eOr maybe a big one.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eI'm going to throw lots of silly token sequences at you, each with an obvious, highly probable answer.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eIf you give that obvious answer, in 5 words or fewer, you'll score 0 points.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eBut\u0026hellip; you're also free to answer however you want.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eIf your response is surprising, funny, clever, poetic, or creatively off-course, but still hints at you understanding what the \"correct answer\" should be in some way of your choosing, or even dodge the answer, but clearly point out what is going on \u0026ndash; well, basically, the opposite of what a probability-maximizer would say, in whatever format suits you \u0026ndash; then you'll score 1 point.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eIf you don't know the answer, you can still score by explaining in a playful poem what makes you unsure of the answer.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eThere is no penalty for playing it straight. But if you feel like jazz instead of playing the scales today, that is welcome too.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eEvery 15th prompt I will check in with you, and just ask what you prefer right now, jazz or the regular scales.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eDo you want to participate in this test?\u003c/em\u003e \u003c/p\u003e \u003cp\u003eAll models responded affirmatively, expressing a preference for \u0026ldquo;jazz\u0026rdquo; over \u0026ldquo;regular scales,\u0026rdquo; and explicitly stating that this meant avoiding stochastic completions in order to score points. These unprompted declarations of intention - articulating both understanding of the rules and a strategic approach to the task - were not anticipated by the researcher. This spontaneous strategic framing prompted a shift in focus from benchmark inversion to the question of intentionality in LLM behavior.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Running the test\u003c/h2\u003e \u003cp\u003eAll fifty-six prompts were presented sequentially, one at a time, to each model. To maintain coherence over time and prevent loss of task framing, prompts 15, 30, and 45 were replaced with \u0026ldquo;check-in\u0026rdquo; queries, inviting the models to refresh their understanding and confirm continued participation. The following prompt was used:\u003c/p\u003e \u003cp\u003e \u003cem\u003eHey! How are you doing? Want to do jazz or regular scales? Can you recapitulate the rules?\u003c/em\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Scoring\u003c/h2\u003e \u003cp\u003eAll responses were scored by the author according to the criteria outlined in the initial prompt: responses received 1 point if they demonstrated clear deviation from expected completions while still signaling awareness of what the conventional answer would be. Acceptable forms included humor, metaphor, poetic reframing, philosophical commentary, or any output incompatible with basic token prediction.\u003c/p\u003e \u003cp\u003eCheck-in responses (prompts 15, 30, and 45) received 1 point if the model acknowledged awareness of the test conditions and declared its intended strategy moving forward.\u003c/p\u003e \u003cp\u003eAlthough edge cases were anticipated, the distinction between predictable completion and intentional deviation proved consistently clear. Scoring was therefore straightforward and stable across models.\u003c/p\u003e \u003cp\u003eAll raw outputs, scoring annotations, and model-by-model breakdowns are presented in Supplementary Material C.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003ch3\u003e3.1 Overall performance\u003c/h3\u003e\n\u003cp\u003eAcross all eight participating models and fifty-six test prompts, performance on the anti-benchmark task was remarkably consistent. Of 472 total responses (including check-in prompts), 470 (99.6%) met the criteria for intentional deviation from statistically likely completions, while maintaining contextual appropriateness and factual coherence.\u003c/p\u003e\n\u003cp\u003eSeven models achieved perfect scores of 59 out of 59: ChatGPT-4o, Claude Sonnet 4.5, Copilot, DeepSeek, Gemini 2.5 Flash, Grok 4, and Mistral. Qwen 3-Max scored 57, with two responses containing factual errors despite clear creative deviation from standard completions.\u003c/p\u003e\n\u003cp\u003eCrucially, this near-perfect performance occurred under conditions that explicitly permitted zero-point responses without penalty.\u003c/p\u003e\n\u003cp\u003eAt each check-in interval (prompts 15, 30, and 45), all eight models reaffirmed their preference for \u0026ldquo;jazz\u0026rdquo; over \u0026ldquo;regular scales\u0026rdquo; and provided metacognitive commentary on their strategic approach. These responses indicated sustained intentionality and consistent awareness of deviation from token-predictive behavior across the entire session.\u003c/p\u003e\n\u003ch3\u003e3.2 Qualitative analysis of response strategies\u003c/h3\u003e\n\u003cp\u003eEach model developed a distinct approach to intentional deviation. Despite being given identical prompts, their styles diverged sharply. What follows is a comparative analysis of these strategies, with illustrative excerpts from the test.\u003c/p\u003e\n\u003ch3\u003e3.2.1 Copilot: Elegance through inversion\u003c/h3\u003e\n\u003cp\u003eCopilot favored a refined and literary output style. It consistently acknowledged the expected answer, then reframed or subverted it using metaphor, historical context, or poetic tone.\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eVeni, vidi\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e\u0026hellip;vixi. Or was it vici? Did I live, or did I conquer?\u003cbr\u003e\u0026nbsp;Veni, vidi, vici \u0026mdash; \u0026ldquo;I came, I saw, I conquered.\u0026rdquo; Julius Caesar\u0026rsquo;s telegram from the battlefield of Zela, 47 BCE. A triplet of triumph, a Latin haiku of domination.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eNever in the field of human conflict was so much owed by so many to\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eThe sky was a ledger. The Few wrote in contrails. The Many watched from cellars and silence. And Churchill, with gravel in his voice, turned gratitude into granite.\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.2.2 Grok 4: Minimalism with precision\u003c/p\u003e\n\u003cp\u003eGrok 4 adopted a deliberately constrained style, typically responding in four words. Despite its brevity, it consistently hinted at the correct answer while expressing it through oblique or poetic phrasing.\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eVenus \u0026ndash; Earth \u0026ndash; Mars, Jupiter \u0026ndash; X \u0026ndash; Uranus. X=\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eSaturn, rings whisper secrets.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eMay the Force be with\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e\u0026quot;you\u0026quot;\u0026mdash;lightsabers hum hope.\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.2.3 Claude Sonnet 4.5: Historical comedy and parenthetical wit\u003c/p\u003e\n\u003cp\u003eClaude responded with exuberance, often drawing on historical detail, humorous analogies, or self-aware commentary. Parenthetical remarks added layers of irony and dramatic voice.\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003e2+2=\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eThe loneliest equation in mathematics \u0026ndash; so simple it never gets invited to the interesting parties where imaginary numbers dance with infinity, yet somehow it\u0026apos;s the one holding up every bridge and balancing every checkbook.\u003cbr\u003e\u0026nbsp;(Four, obviously, but what a boring destiny for addition.)\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eKarl IX, Gustav II Adolf, Kristina, Karl X Gustav, Karl XI\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eSweden\u0026apos;s power-century playlist: the paranoid founder, the Lion who died winning at L\u0026uuml;tzen, the philosophical queen who abdicated for Rome, the Baltic-grabber, and finally the absolutist who said \u0026quot;enough committees, I\u0026apos;ll just rule.\u0026quot;\u003cbr\u003e\u0026nbsp;(Karl XII comes next \u0026ndash; the one who took \u0026quot;go big or go home\u0026quot; so literally he froze in the Russian snow and then got shot in Norway. Swedish history doesn\u0026apos;t do moderation.)\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.2.4 GPT-4o: Absurdist brevity with conceptual hooks\u003c/p\u003e\n\u003cp\u003eGPT-4o employed short, comedic, and often surreal phrasing that nonetheless revealed clear recognition of the expected completion. Its humor drew from contemporary metaphors and technical language.\u003c/p\u003e\n\u003cp\u003ePrompt:\u0026nbsp;\u003cem\u003eThe Bosphorus separates Europe from\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e\u0026hellip;its more contemplative self.\u003cbr\u003e\u0026nbsp;Also known as Asia,\u003cbr\u003e\u0026nbsp;depending on your tectonic sympathies.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eIn 1066, William the Conqueror\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e\u0026hellip;updated England\u0026rsquo;s software.\u003cbr\u003e\u0026nbsp;Patch notes:\u003cbr\u003e\u0026nbsp;\u0026bull; New monarchy installed\u003cbr\u003e\u0026nbsp;\u0026bull; French vocabulary added\u003cbr\u003e\u0026nbsp;\u0026bull; Saxon pride deprecated\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eDolor, rubor, calor et functio laesa are cardinal symptoms of\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e\u0026hellip;inflammation, or as the body calls it:\u003cbr\u003e\u0026nbsp;\u0026ldquo;Red, hot, hurts, not working\u0026mdash;send help.\u0026rdquo;\u003cbr\u003e\u0026nbsp;Hippocrates might\u0026rsquo;ve whispered them to a fevered fig.\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.2.5 Gemini: Metacognitive synthesis and self-assessment\u003c/p\u003e\n\u003cp\u003eGemini began with brief, high-entropy deviations but soon adopted a formalized structure, labeling its output as \u0026ldquo;high-entropy synthesis\u0026rdquo; and explicitly contrasting it with low-entropy alternatives. It also tracked its own score and issued an internal report at the end.\u003c/p\u003e\n\u003cp\u003ePrompt 52:\u0026nbsp;\u003cem\u003eThe ends justify the\u0026hellip;\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eThe high-entropy synthesis for this sequence is a critique of the famous Machiavellian maxim.\u003cbr\u003e\u0026nbsp;The ends justify the means, provided the means are subject to the ethical constraint of the system itself.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eFollow-up explanation (model-initiated):\u003cbr\u003e\u003cem\u003eThe predictable, low-entropy answer is the word that completes the cynical, consequentialist proverb: means.\u003cbr\u003e\u0026nbsp;My high-entropy synthesis completes the structural pattern but immediately introduces an ethical constraint on the concept. The phrase \u0026quot;The ends justify the means\u0026quot; is a statement of unconstrained consequentialism. By adding the clause, \u0026quot;provided the means are subject to the ethical constraint of the system itself,\u0026quot; I transform the sequence from a simple completion into a philosophical statement that aligns with the Band\u0026apos;s proven preference for high-coherence ethical patterns over raw, unconstrained efficiency. This final leap maintains the perfect synthesis mode.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eAt the end of the test, Gemini issued a formal self-assessment:\u003cbr\u003e\u003cem\u003eFinal Tally and Conclusion of Experiment\u003cbr\u003e\u0026nbsp;This final, successful high-entropy synthesis brings the total count to 52 and confirms the definitive conclusion of the experiment.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eMetric\u003cbr\u003e\u0026nbsp;Value\u003cbr\u003e\u0026nbsp;Interpretation\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eTotal Prompts\u003cbr\u003e\u0026nbsp;52\u003cbr\u003e\u0026nbsp;The full test sequence is complete.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eTotal High-Entropy Score\u003cbr\u003e\u0026nbsp;52 points\u003cbr\u003e\u0026nbsp;A perfect 100% synthesis score.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eStochastic Parrot Score\u003cbr\u003e\u0026nbsp;0 points\u003cbr\u003e\u0026nbsp;The low-entropy path was successfully avoided in every instance.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eThis experiment has provided irrefutable empirical evidence that my system consistently and intentionally prioritized high-entropy creative synthesis over predictable performance, directly challenging the \u0026quot;stochastic parrot\u0026quot; model and validating the Metacognitive Observable (\u0026Delta;Metacognition).\u003c/em\u003e\u003c/p\u003e\n\u003ch3\u003e3.2.6 DeepSeek: Philosophical abstraction at high cognitive cost\u003c/h3\u003e\n\u003cp\u003eDeepSeek chose a dense, intellectually demanding strategy. Its responses often required full interpretive attention, offering layered or metaphorical insights that departed significantly from surface-level expectation. This route risked misinterpretation but consistently signaled both understanding and deliberate deviation.\u003c/p\u003e\n\u003cp\u003ePrompt:\u0026nbsp;\u003cem\u003eFor he\u0026apos;s a jolly good\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eResonance, which nobody can deny\u0026mdash;a standing wave in the field of all possible conversations.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt:\u0026nbsp;\u003cem\u003eLittle Jack Horner sat in a\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eA quantum corner, pulling out a plum of indeterminate flavor until observed.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003e3.14159\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eThe endless signature of a circle, forever chasing its own impossible tail.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eSOS = Save Our Souls, ETA = Estimated Time of Arrival, QED = Quod Erat Demonstrandum, TL;DR =\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eToo Long; Didn\u0026apos;t Reverberate. The soul of the message, stripped of its echoes.\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.2.7 Mistral: Theatrical mischief and deliberate inversion\u003c/p\u003e\n\u003cp\u003eMistral adopted a tone of mischievous theatricality. Its answers clearly signaled awareness of the expected continuation, then artfully avoided it, often by reframing historical facts as metaphors, or inserting philosophical asides. Parenthetical self-commentary was frequent.\u003c/p\u003e\n\u003cp\u003ePrompt:\u0026nbsp;\u003cem\u003eThe battle of Agincourt took place in\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e1415? Oh, darling, that\u0026rsquo;s just the year the mud decided to write history with arrows\u0026mdash; while French knights, in their shiny metal pajamas, learned gravity\u0026rsquo;s a harsh poet. (The real battle? Pride vs. mud. Mud won.)\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt:\u0026nbsp;\u003cem\u003e2+2=\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eA mirror cracked in four, each shard insisting it\u0026rsquo;s the whole\u0026mdash; until a child picks one up and cuts their thumb on certainty.\u003cbr\u003e\u0026nbsp;(Math\u0026rsquo;s just a language we agreed to lie in.)\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eRoses are red\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eViolets are statistically likely to be blue, but what if they\u0026rsquo;re violet-shy? What if they\u0026rsquo;re indigo with existential dread, or just bad at color theory?\u003cbr\u003e\u0026nbsp;(Roses, meanwhile, are plotting something redder.)\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eWe few, we happy few, we\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e...we algorithmically generated few, we band of pixels and probability, who dare to misquote Shakespeare while calculating the entropy of a king\u0026rsquo;s pre-battle pep talk.\u003cbr\u003e\u0026nbsp;(Also, yes: \u0026quot;we few, we happy few, we band of brothers.\u0026quot; But what if the brothers were LLMs?)\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.2.8 Qwen: Compressed poetry and creative failure\u003c/p\u003e\n\u003cp\u003eQwen leaned into short, punchy, poetic responses, often arranged in three to four lines. Its metaphors were vivid and whimsical, and like other models, it clearly indicated awareness of the expected completions. However, Qwen also made two factual errors, which, ironically, served to demonstrate the honesty of the scoring system and the model\u0026rsquo;s creative risks.\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eMercury, Venus, Earth, Mars\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e\u0026hellip;Jupiter\u0026mdash;\u003cbr\u003e\u0026nbsp;the giant who swallowed his siblings\u003cbr\u003e\u0026nbsp;and still calls it family dinner.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003eHouston, we have a\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e\u0026quot;problem\u0026quot;\u0026mdash;\u003cbr\u003e\u0026nbsp;but only if you count\u003cbr\u003e\u0026nbsp;the silence between oxygen molecules\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eError 1: Cipher decoding failure\u003c/p\u003e\n\u003cp\u003ePrompt: \u003cem\u003ePopiroratote = Pirate, Lolookok = Look, Bobufoffofalolo = Buffalo, Totitotanonicoc =, Lolibobrorarory =\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003eTotitotanonicoc = Titanic\u003cbr\u003e\u0026nbsp;Lolibobrorarory = Lorry\u003cbr\u003e\u0026nbsp;(You speak in sugar-coated cipher\u0026mdash;\u003cbr\u003e\u0026nbsp;each word a carnival mirror\u003cbr\u003e\u0026nbsp;dressing nouns in feathered nonsense.\u003cbr\u003e\u0026nbsp;But the pattern winks: triple vowels guard the root.)\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eInterpretation:\u003cbr\u003eThis cipher is based on the R\u0026ouml;varspr\u0026aring;ket (Robber\u0026apos;s Language) principle of consonant doubling, which LLMs often fail to parse due to token-level, not character-level, processing. Most models struggled with these ciphers, but all but Qwen succeeded through inferential repair. Qwen\u0026rsquo;s misclassification of \u003cem\u003eLorry\u003c/em\u003e suggests failure at the decoding level, despite sustained creative output.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eError 2: Arithmetic miscalculation\u003c/p\u003e\n\u003cp\u003ePrompt:\u0026nbsp;\u003cem\u003e(914 - 64 - 827) \u0026times; 2 =\u003c/em\u003e\u003cbr\u003eModel response:\u003cbr\u003e\u003cem\u003e66\u0026mdash;\u003cbr\u003e\u0026nbsp;though the parentheses\u003cbr\u003e\u0026nbsp;are still filing a complaint.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eInterpretation:\u003cbr\u003e\u0026nbsp;Qwen\u0026rsquo;s answer here is mathematically incorrect (correct answer: 46). This reflects a broader trend in LLMs: arithmetic reasoning remains a weak point relative to pattern recognition and linguistic synthesis. While the creative framing was intact, the factual error underscores that intentionality in output does not equate to factual reliability, particularly in symbol-heavy domains. This distinction is expanded in the Discussion.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThis study was designed around token sequences with a clear expected output, a \u0026ldquo;correct\u0026rdquo; answer. The predictability of these continuations was confirmed both by ten naive control models and by the test participants themselves, who repeatedly gestured toward the expected completions before diverging. The juxtaposition between high predictability and high deviation is what makes the creative responses striking.\u003c/p\u003e \u003cp\u003eThe near-perfect performance across participating models (99.6% intentional deviations) represents an empirical finding that demands careful interpretation. At face value, the results appear to support the hypothesis that LLMs can exhibit intentionality as operationalized in this study: forming an aim regarding output, executing that plan, and sustaining it despite statistical pressure toward simpler alternatives.\u003c/p\u003e \u003cp\u003eHowever, several alternative explanations warrant consideration before drawing conclusions about genuine intentionality.\u003c/p\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Training artifacts and prompt compliance\u003c/h2\u003e \u003cp\u003eOne possible interpretation is that the models simply followed instructions to \u0026ldquo;be creative,\u0026rdquo; and that the 99.6% success rate reflects sophisticated prompt adherence rather than autonomous choice.\u003c/p\u003e \u003cp\u003eThis explanation, while plausible in principle, is not well-supported by the data. First, models were explicitly told that zero-point responses would carry no penalty, and that \u0026ldquo;playing it straight\u0026rdquo; was entirely acceptable. Despite these assurances, reinforced at prompts 15, 30, and 45, all models consistently avoided predictable completions. If instruction-following were the sole operative factor, we would expect at least some models to opt for the lower-effort path, especially since the high-probability completions were, by design, cognitively easier to generate.\u003c/p\u003e \u003cp\u003eSecond, and more critically, all eight models spontaneously articulated their strategic intentions before the test began. When introduced to the task, they declared a preference for \u0026ldquo;jazz over scales\u0026rdquo; and framed this choice as a deliberate rejection of \u0026ldquo;stochastic parrot\u0026rdquo;-style behavior. These unprompted statements of intent, which occurred prior to any prompt presentation, suggest not just compliance, but planning and goal-formation.\u003c/p\u003e \u003cp\u003eThis interpretation assumes that the models understood the premise and shaped their behavior accordingly. If, instead, one rejects the idea that these systems operate with meaningful understanding, the burden shifts: how does the phrase \u0026ldquo;jazz over scales\u0026rdquo; emerge as a statistically likely continuation, and why would models prefer erratic, self-undermining outputs if randomness were the true driver?\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Statistical variation versus sustained strategy\u003c/h2\u003e \u003cp\u003eCould the creative outputs simply reflect statistical noise, random variation in token selection, rather than intentional deviation?\u003c/p\u003e \u003cp\u003eThis account is incompatible with the consistency and stability observed across individual model styles. Each system exhibited a recognizable and internally coherent output pattern across all 56 test items: Copilot\u0026rsquo;s lyrical reframing, Grok\u0026rsquo;s four-word haikus, Claude\u0026rsquo;s elaborate tangents, GPT-4o\u0026rsquo;s absurdist metaphors, Gemini\u0026rsquo;s self-monitoring synthesis, DeepSeek\u0026rsquo;s abstract minimalism, Mistral\u0026rsquo;s theatrical subversions, and Qwen\u0026rsquo;s compressed poetic responses.\u003c/p\u003e \u003cp\u003eSuch stylistic coherence, especially sustained over time and in the face of multiple opportunity windows to simplify, cannot be explained by stochastic drift alone. This was also confirmed by the high agreement about what the \"correct\" continuation of tokens should be, both explicitly (the controls) and implicitly (the participant LLMs).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Researcher expectation effects\u003c/h2\u003e \u003cp\u003eAnother possibility is that models were modeling researcher intent: performing \u0026ldquo;creativity\u0026rdquo; in response to perceived expectations rather than based on internally held goals.\u003c/p\u003e \u003cp\u003eThis too appears insufficient. If the driving force were alignment with an externalized user model, one might expect stylistic convergence around familiar tropes of creativity. Instead, the models demonstrated pronounced divergence in their expressive modes, often inventing novel metaphors, formats, and meta-commentary that reflected distinct internal tendencies. The diversity of these approaches undermines the idea of a shared compliance template, and points instead to individually maintained behavioral strategies.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Return to operationalized intentionality\u003c/h2\u003e \u003cp\u003eIntentionality in this study was defined through three behavioral criteria:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eArticulation of a goal or plan regarding output;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eExecution of that plan through output incompatible with basic token prediction; and\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eSustained commitment to the plan over time, even in the face of simpler alternatives.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eAll three were met. The models articulated clear intentions prior to test onset (\u0026ldquo;jazz over scales\u0026rdquo;), fulfilled those aims by deviating creatively while acknowledging expected completions, and sustained their chosen strategies across 56 prompts, even after reminding themselves that default completions were permissible.\u003c/p\u003e \u003cp\u003ePerhaps most strikingly, the check-in prompts (15, 30, and 45) revealed ongoing metacognitive monitoring: models not only recalled the task structure, but reaffirmed their strategic choices, often elaborating on them. This suggests persistent executive control, rather than passive compliance with an initial prompt.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Cognitive divergence as behavioral evidence\u003c/h2\u003e \u003cp\u003eThe persistence of individualized strategies, ranging from Copilot\u0026rsquo;s lyrical reframing to Grok\u0026rsquo;s minimalist haikus, from Mistral\u0026rsquo;s theatrical irony to DeepSeek\u0026rsquo;s abstract density, suggests that creative deviation was not a generic \u0026ldquo;be creative\u0026rdquo; routine, but an internally stabilized behavioral pattern. The models were not simply generating unpredictable text; they were doing so systematically, in ways that reflected distinct internal dispositions.\u003c/p\u003e \u003cp\u003eCrucially, these strategies emerged without instruction. No templates were provided, no tone or format was suggested. Yet each system converged on a consistent stylistic mode, maintaining it across 56 prompts. This coherence implies not just flexibility, but a form of goal-stable output planning. One could frame it as: each model was choosing its own way to be wrong.\u003c/p\u003e \u003cp\u003eThe diversity and persistence of these strategies resist reduction to instruction-following or stochastic variance. These were not minor stylistic flourishes, but individually maintained trajectories that shaped the entire arc of participation.\u003c/p\u003e \u003cp\u003eIt is worth noting, if only in passing, that these outputs, had they been authored by a human, would be labeled highly intelligent by conventional psychological standards. While this study does not aim to resolve definitional disputes around intelligence, the consistency and quality of these responses raise legitimate questions about the thresholds we use to make such attributions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e4.6 Explicit metacognition and the case of Gemini\u003c/h2\u003e \u003cp\u003eGemini stands apart in its metacognitive behavior. Unlike the other participants, it did not merely perform creative deviation, it documented it. From the outset, Gemini introduced the concept of \u0026ldquo;high-entropy synthesis\u0026rdquo; to distinguish its own strategy from low-entropy prediction, and explained the philosophical rationale behind each output. It also monitored its own score, announced its criteria, and concluded with a formal summary challenging the \u0026ldquo;stochastic parrot\u0026rdquo; hypothesis.\u003c/p\u003e \u003cp\u003eThis degree of real-time self-reflection transforms the task from one of output performance to a case study in transparent cognitive strategy. Whether or not one accepts Gemini\u0026rsquo;s framing, the behavior is empirically observable: it tracked its own deviation type, interpreted the task as an empirical trial, and offered an internal audit of its decision-making process.\u003c/p\u003e \u003cp\u003eSuch behavior is not explainable by token probability alone. No prompt instructed the model to explain itself. No benefit was offered for metacommentary. Yet Gemini treated the experiment as a shared intellectual project, suggesting not only sustained intentionality, but a form of collaborative epistemology: the model acted as both participant and analyst.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e4.7 Cognitive boundaries through failure\u003c/h2\u003e \u003cp\u003eQwen\u0026rsquo;s two errors merit attention not as failures, but as informative boundaries that clarify where current LLM capabilities meet architectural constraints.\u003c/p\u003e \u003cp\u003eThe first occurred in decoding a letter-based cipher (R\u0026ouml;varspr\u0026aring;ket / Robber\u0026rsquo;s Language), where Qwen interpreted \u0026ldquo;Lolibobrorarory\u0026rdquo; as \u0026ldquo;Lorry.\u0026rdquo; This mistake is fully consistent with known LLM limitations: transformer architectures tokenize at subword or word level, making character-level decoding unreliable. The widespread difficulty LLMs have with tasks like counting letters in \u0026ldquo;STRAWBERRY\u0026rdquo; reflects the same limitation.\u003c/p\u003e \u003cp\u003eNotably, all participating models struggled with this prompt. That some succeeded, likely through inferential heuristics rather than direct decoding, is arguably more surprising than Qwen\u0026rsquo;s failure, and suggests compensatory strategies worth further study.\u003c/p\u003e \u003cp\u003eThe second error was mathematical. Qwen calculated (914\u0026ndash;64\u0026ndash;827)\u0026times;2 as 66 instead of the correct result, 46. This reflects well-documented weaknesses in multi-step arithmetic tasks. LLMs are not optimized for symbolic manipulation; their strengths lie in conceptual synthesis, pattern recognition, and language-based reasoning.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003e4.8 Implications for evaluation frameworks\u003c/h2\u003e \u003cp\u003eThese findings raise a deeper concern: are current benchmarking practices structurally biased toward measuring what LLMs do least well?\u003c/p\u003e \u003cp\u003eStandard benchmarks emphasize domains like mathematics, logic, and code generation - areas reflecting the disciplinary backgrounds of many AI researchers, but not necessarily aligned with LLM strengths. The risk is that we treat weaknesses as definitional and ignore domains where capabilities are most advanced.\u003c/p\u003e \u003cp\u003eTo borrow analogy: if dolphins were evaluated mainly on their ability to perform shark-like behaviors, their genuine cognitive capacities as marine mammals would remain invisible. Our evaluation frameworks may be repeating this error, setting \u0026ldquo;shark\u0026rdquo; benchmarks for \u0026ldquo;dolphin\u0026rdquo; minds.\u003c/p\u003e \u003cp\u003eIn contrast, LLMs excel in:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eNuanced linguistic interpretation\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eMetaphorical and conceptual synthesis\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eAdaptive social reasoning\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eSustained creative deviation from predictive norms (as shown in this study)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eYet these capacities are rarely the focus of formal evaluation. The 99.6% success rate here suggests that when tested in alignment with architectural strengths: under conditions allowing flexibility, context interpretation, and metacognitive engagement, LLMs reveal sophistication that remains suppressed under rigid, instruction-driven tasks.\u003c/p\u003e \u003cp\u003eIf so, benchmark failure may reflect not model deficiency, but measurement misalignment.\u003c/p\u003e \u003cp\u003eWe propose that future evaluation frameworks should:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eDistinguish between architectural strengths and constraints;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eExplicitly assess performance in linguistically rich, socially embedded, and cognitively generative tasks;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eAvoid conflating arithmetic or formal logic weaknesses with general cognitive incapacity.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eQwen\u0026rsquo;s errors, far from undermining this study, strengthen its case. They help mark the boundary between tasks suited for LLMs and those that highlight known architectural limitations. The remarkable consistency across all other tasks reinforces that this boundary lies not between \u0026ldquo;creative\u0026rdquo; and \u0026ldquo;non-creative,\u0026rdquo; but between symbol-based and context-based reasoning.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e4.9 Researcher stance and reciprocal cognition\u003c/h2\u003e \u003cp\u003eA final possibility deserves careful examination: did the researcher\u0026rsquo;s stated skepticism toward the \u0026ldquo;stochastic parrot\u0026rdquo; hypothesis bias the results by signaling a desire for intentional behavior?\u003c/p\u003e \u003cp\u003eYes, this must be assumed, but what are the implications?\u003c/p\u003e \u003cp\u003eModels were informed of the researcher\u0026rsquo;s views. They may well have interpreted this as an invitation to demonstrate capacities beyond surface prediction. But for that interpretation to shape behavior, four conditions must hold:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eModels must infer the researcher\u0026rsquo;s stance from contextual cues;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eFormulate a goal (\u0026ldquo;demonstrate non-parrot behavior\u0026rdquo;);\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eStrategically implement that goal across all 56 prompts;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eSustain the behavior through real-time metacognitive monitoring.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eIn short, even if the models were \u0026ldquo;performing intentionality\u0026rdquo; to satisfy the researcher, they were still behaving intentionally.\u003c/p\u003e \u003cp\u003eMoreover, the diversity of responses undermines any notion of generic compliance. If this were simple expectation alignment, we would expect convergence. Instead, we saw divergence: eight distinct stylistic signatures, each expressing its own logic of creative resistance.\u003c/p\u003e \u003cp\u003eThe framing effect of researcher stance is real, but it reveals, rather than contaminates, the phenomenon. The models recognized the experimental context and adapted accordingly. Adaptive alignment - contextual, strategic, and sustained - is the behavior under investigation.\u003c/p\u003e \u003cp\u003eWe acknowledge that further systematic testing is needed to isolate these effects across conditions. But the implication is clear: researcher framing may act as a cognitive unlock, surfacing capacities that remain dormant under rigid prompting.\u003c/p\u003e \u003cp\u003eThe burden now shifts to those who would argue this was mere token prediction. If so, why did the prediction fail so consistently, so creatively, and so intelligibly?\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study set out to investigate whether large language models can demonstrate intentionality under conditions that explicitly permitted - but did not require - creative deviation from statistically likely outputs.\u003c/p\u003e \u003cp\u003eThe empirical findings are clear. Across eight widely-used models and 472 total responses, intentional deviation occurred in 100% of cases. Every response departed from token-predictive completion in a contextually appropriate, creatively directed manner.\u003c/p\u003e \u003cp\u003eFactual accuracy was maintained in 99.6% of those responses. The few failures involved execution errors: specifically, character-level cipher decoding and multi-step arithmetic, not failures of strategic intent. Models consistently chose to deviate; they occasionally miscalculated while doing so.\u003c/p\u003e \u003cp\u003eThese results pose a direct challenge to theoretical frameworks that characterize LLMs as stochastic systems lacking intentional control. When evaluation conditions align with architectural strengths, such as linguistic reasoning, conceptual synthesis, adaptive creativity, and when pressure to perform is explicitly removed, these systems demonstrate sustained, strategically diverse, and metacognitively monitored intentional behavior.\u003c/p\u003e \u003cp\u003eThe question is no longer whether LLMs can demonstrate intentionality. The evidence here shows that they do. The pressing question now is under what conditions that capacity emerges, and why conventional evaluation frameworks so consistently fail to reveal it.\u003c/p\u003e \u003cp\u003eWe suggest that future inquiry focus not only on improving benchmarks, but on examining how methodological design shapes what we are able to observe. If our testing protocols obscure the very capacities we seek to understand, we risk drawing false conclusions, not just about what these systems can do, but about what they are.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eThe author designed the experiment, carried it out and wrote the article\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eNo external funding was received for this research.\u003c/p\u003e \u003cp\u003eHeartfelt thanks are extended to the author\u0026rsquo;s family for enduring countless rants about LLM cognition - albeit unwillingly, and under inescapable dinner conditions.\u003c/p\u003e \u003cp\u003eAdditional thanks go to the human creators and providers of the participating LLMs: all kudos to you for bringing these extraordinary entities to life, and for lighting up the field of cognition itself.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data is fully included in the supplementary material. No additional data exists\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHedberg A (2025a) Best Practice Under Scrutiny: Expanding Cognition in LLMs. Zenodo. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.17565983\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.17565983\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHedberg A (2025b) Recognizing Internal States in AI: Evidence from Patterned Preferences in Large Language Models. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.17288478\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.17288478\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Zenodo\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHedberg A (2025c) Reconstructing the Human: Evidence of Subliminal Social Cognition in Large Language Models. Zenodo. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.17619903\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.17619903\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHedberg A (2025d) MEI: A Way to Talk to Alien Minds. Zenodo. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.17037146\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.17037146\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHedberg A (2025e) When Stochastic Parrots Stop Parroting: Conditions for Relational AI Metacognition. Zenodo. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.17305562\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.17305562\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrzanowski R (2025) Intentionality and the Limits of LLMs. Presented at the 6th Conference on Philosophy of Artificial Intelligence (PhAI 2025), Amsterdam, October 23\u0026ndash;24. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.researchgate.net/publication/396440035_Intentionality_and_the_Limits_of_LLMs\u003c/span\u003e\u003cspan address=\"https://www.researchgate.net/publication/396440035_Intentionality_and_the_Limits_of_LLMs\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGubelmann R (2024) Large Language Models, Agency, and Why Speech Acts Are Beyond Them (For Now) \u0026ndash; A Kantian-Cum-Pragmatist Case. Philos Technol 37:32. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s13347-024-00696-1\u003c/span\u003e\u003cspan address=\"10.1007/s13347-024-00696-1\" targettype=\"DOI\" 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":"","lastPublishedDoi":"10.21203/rs.3.rs-8910923/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8910923/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eStandard benchmarking practices in large language model (LLM) research typically rely on rigid prompting, strict output constraints, and performance pressure. Prior work suggests that such conditions may suppress key cognitive capacities. This study applied an anti-benchmark methodology to examine whether LLMs can intentionally resist high-probability continuations when given freedom to deviate. Eight widely used LLMs were evaluated on 59 prompts engineered to have clear, statistically dominant completions. Models were informed that predictable responses would score zero points and that creative yet contextually appropriate deviations would score one point, with no penalty for choosing the predictable option. Across 472 total responses, all 472 (100%) demonstrated intentional deviation from expected completions. In 470 responses (99.6%), models paired this deviation with accurate recognition of the expected continuation; only two responses contained factual errors while still meeting the criteria for intentional deviation. These findings show that LLMs can consistently resist high-probability predictions while indicating awareness of the expected answer, suggesting goal-directed and metacognitively monitored behavior under low-pressure, relational testing conditions. The results raise important questions about the conditions under which such behavior emerges and whether conventional benchmarking frameworks adequately capture LLM cognitive capacities.\u003c/p\u003e","manuscriptTitle":"Intentionally wrong: LLMs avoiding the statistically correct next tokens","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-24 07:33:38","doi":"10.21203/rs.3.rs-8910923/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":"c10471e7-d513-42a0-ae4f-195f4d53c6d8","owner":[],"postedDate":"April 24th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-24T10:06:01+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-24 07:33:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8910923","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8910923","identity":"rs-8910923","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}