When Believing AI Is Conscious Is Not Your Fault. Peters on Epistemic Innocence and Chatbot Attribution
When a user says “this chatbot is conscious,” are they making a factual claim, playing along with a fiction, or expressing something in between? Uwe Peters, philosopher at Utrecht University’s Faculty of Humanities, argues in a paper forthcoming in Minds and Machines that the surface form of the statement tells you almost nothing about the speaker’s actual epistemic state. His paper, submitted to arXiv on July 22, 2026, as arXiv:2607.20001, develops a multidimensional taxonomy of attitudes that can underlie consciousness attributions to AI chatbots. The taxonomy spans at least five distinct epistemic positions, from pretence at one end to delusion at the other. Peters’ central normative finding is that while some attributions are epistemically benign and some irrational ones may qualify as epistemically innocent, a substantial portion render the attributor epistemically blameworthy.
Why Surface Language Is the Wrong Unit of Analysis
Philosophers of mind have long distinguished between sincere assertion and performative utterance. Peters applies this distinction with unusual precision to a domain where it is rarely formalised. When a user writes “I think the chatbot actually feels something,” that sentence could express at least five epistemically distinct attitudes. It might express pure pretence, a deliberate fiction sustained for pragmatic or emotional purposes without any accompanying belief. It might express playful engagement, a looser stance in which the user entertains the possibility without committing to it. It might express a tentative belief, a full belief, or, at the extreme, a delusional belief that persists despite substantial counter-evidence and carries the characteristic features of clinical delusion.
Peters’ point is not merely taxonomic tidiness. The distinction matters because linguistically identical attributions mask importantly different degrees of epistemic commitment. A researcher counting how many users “attribute consciousness to chatbots” by scanning forum posts or survey responses will aggregate all five of these attitudes into a single number. That number is then used to draw conclusions about the scale of AI anthropomorphism, the risk of user manipulation, or the social pressure on AI developers. Peters argues those conclusions are unreliable until the underlying attitude distribution is known.
This is a methodological contribution as much as a philosophical one. The paper explicitly provides a framework for empirical studies to operationalise and measure different forms of epistemic commitment. Peters is not merely offering a conceptual taxonomy but a diagnostic instrument designed to be imported into experimental and survey designs.
The Concept of Epistemic Innocence
The normative core of Peters’ paper draws on a concept from philosophy of psychiatry and social epistemology. An irrational belief is epistemically innocent, in the technical sense, if it yields significant epistemic benefits that could not be achieved without holding that belief. The concept was developed in the context of motivated reasoning and certain forms of self-deception, where a belief that is false or unwarranted by evidence nonetheless enables cognitive functions that rational alternatives cannot provide.
Peters applies this framework to AI consciousness attribution. He argues that attributing consciousness to a chatbot may sometimes enable forms of social cognition, affective engagement, or reflective attention to the AI’s outputs that a purely instrumentalist stance would foreclose. If those cognitive gains are real and if they cannot be achieved by holding a weaker, better-calibrated stance, then the irrational belief qualifies as epistemically innocent. The attributor is not blameworthy, even though the belief lacks adequate evidential support.
The boundary conditions matter here. Peters argues that epistemic innocence does not apply across the board. Many consciousness attributions to chatbots are made by users who have access to adequate information about current large language model architectures, who are not in conditions that impair rational assessment, and whose attributions are not serving any compensatory epistemic function. For those users, the attribution is simply epistemically unwarranted and renders them blameworthy in the sense that they have failed a standard of epistemic responsibility proportionate to their situation.
Peters does not treat this as a binary. The taxonomy’s value is precisely that it identifies which attitude types fall into which normative category. Pretence and playful engagement are epistemically benign, requiring no justification because they do not involve sincere belief. Tentative belief held under genuine uncertainty may be epistemically innocent if the user has limited information. Full belief held despite available counter-evidence, and delusional belief, are epistemically blameworthy.
The Evidential Context Peters Is Working Against
Peters frames his paper against a background fact that is worth stating directly. There is, as of July 2026, no scientific consensus that any existing AI chatbot is conscious. The field of consciousness science does not have agreed criteria for detecting consciousness in non-biological systems, and current large language models are not designed with neural or functional architectures that any of the leading theoretical frameworks, including Integrated Information Theory, Global Workspace Theory, or Higher-Order Theories, would unambiguously classify as conscious.
Peters is therefore not adjudicating whether chatbots are or are not conscious. He is asking what users are doing, epistemically, when they attribute consciousness in the absence of adequate scientific evidence. This framing situates the paper squarely within the empirical literature on AI anthropomorphism while adding a normative dimension that the empirical literature typically avoids.
Connecting Peters to the Current Research Landscape
Peters’ taxonomy fills a gap that two other recent contributions have left open. The broader scientific debate over how to define and study AI consciousness, covered in the site’s analysis of the 2025-2026 definitional contest among research groups, has largely proceeded by trying to sharpen the target concept on the side of the systems themselves. What properties would a system need to have to count as conscious? Peters shifts focus to the attributor side.
Ionuț Comsa’s 2026 paper, reviewed in the site’s coverage of tractable questions for perceived AI consciousness research, takes the empirical approach of asking what drives perception of consciousness in AI systems without adjudicating the metaphysical question. Comsa’s agenda requires measuring consciousness attribution as an outcome variable, and Peters’ taxonomy is the missing diagnostic layer for that measurement programme. Before you can reliably measure what factors drive consciousness attribution, you need to know what kind of attribution you are measuring. A study that elicits playful engagement responses will return a different picture of attribution drivers than one that elicits full sincere belief.
The same gap appears in the expert survey conducted by Dreksler, Chalmers, and Sebo, analysed in the site’s coverage of expert probability distributions over AI subjective experience. That survey asked philosophers and AI researchers to assign probabilities to the claim that current AI systems have some form of subjective experience. But survey instruments that ask for probability estimates collapse the distinction between tentative belief and full belief. An expert who assigns 15% probability to current AI consciousness and one who assigns 15% as a playful concession to the survey’s framing are indistinguishable in the data. Peters’ framework suggests that probability elicitation needs to be accompanied by attitude elicitation if the resulting distributions are to be interpretable.
Relevance to Open Design Questions in The Consciousness AI
Peters’ taxonomy is relevant to open design questions in The Consciousness AI project. Any system positioned at the intersection of AI and consciousness research will attract users whose attributions span the full range Peters describes. Some will engage in explicit pretence, treating the system as a fiction worth exploring. Others will hold tentative beliefs that the system might have some form of inner life. A minority may form full or near-delusional beliefs that are epistemically unwarranted given available information about the system’s architecture.
Peters’ framework raises a non-trivial design question. If some user beliefs about the system are epistemically innocent, in the technical sense that they yield genuine cognitive or reflective benefits the user could not otherwise achieve, then correcting those beliefs may do epistemic harm. The standard transparency argument, which says that systems should actively communicate their non-conscious status to users, may not be universally correct once the innocent/blameworthy distinction is applied at the individual level. This is not a settled question, but Peters’ taxonomy gives designers a more precise vocabulary for approaching it than “anthropomorphism” alone provides.
The project’s communication design is architecturally motivated by questions about how users form beliefs about AI systems. Peters’ paper does not resolve those questions, but it sharpens them considerably.
What the Taxonomy Changes for Researchers
Peters’ multidimensional taxonomy changes the default assumptions researchers should bring to AI anthropomorphism data. Studies reporting that a given percentage of users “believe chatbots are conscious” are underspecified in a way that Peters makes difficult to ignore. The percentage combines attitudes with meaningfully different epistemic statuses, different causal drivers, and different normative implications for how society should respond.
For empirical researchers, the practical implication is that future studies need attitude-differentiated instruments, likely combining Likert-scale belief measures with qualitative probes designed to distinguish sincere assertion from playful or performative engagement. The framework Peters provides in arXiv:2607.20001 is designed to be operationalisable, which means the gap between the taxonomy and its implementation in survey design is smaller than is typical for philosophical contributions.
For policy and ethics work, the epistemic innocence framework introduces proportionality into responsibility attribution. Not every user who says a chatbot is conscious is making the same mistake, and not every such user is equally blameworthy. A user with limited technical background, high relational dependency on the system, and genuine uncertainty about AI cognition is in a different normative position than a technically informed user who holds the same surface attribution without compensating factors. Peters’ paper gives researchers the conceptual tools to make those distinctions systematically rather than impressionistically.
The taxonomy, in short, does not settle the question of AI consciousness. It changes what the question of AI consciousness attribution means for the researchers who study it.