A conversational system answers a difficult question in fluent language and then says, “I’m worried I may have misunderstood you.” A user might infer that it can plan, that it can feel worry, that it is simply following a conversational script, or that its wording makes it pleasant to use. Those interpretations should not be merged into one number. Measuring mind perception in AI asks what mental capacities a person attributes to a system, not whether the system actually has them.
Measuring mind perception in AI means studying people’s attributions of mental capacities such as planning, choosing, feeling or suffering to an artificial system under specified conditions. The commonly used distinction between perceived agency and perceived experience can be useful, but it is not a complete or settled map of every concept of mind. This Methods & Measures entry focuses on construct validity: how to ask precise questions, design interpretable tasks and avoid mistaking anthropomorphic interface cues, trust, social attachment or system performance for the perception itself.
Agency and experience are different questions
Gray, Gray and Wegner asked respondents to compare the mental capacities of varied human and nonhuman targets. Their analysis identified two broad dimensions: agency, including capacities associated with thought, planning and self-control, and experience, including capacities associated with hunger, fear and pain. The result concerns how respondents organized judgments about targets. It is not an experiment proving which targets possessed those capacities. When applied to AI, a high agency rating means that people attributed some ability to direct action or thought; a high experience rating means that they attributed some capacity for subjective feeling.
The two dimensions can move apart. A user may believe an automated planning tool can organize tasks while denying that it feels anything. A system with an expressive voice might evoke impressions of feeling without being judged good at independent planning. A single item such as “How humanlike is this AI?” cannot reveal which attribution changed. If an intervention is supposed to affect perceived experience, questions about intelligence alone are insufficient. If the aim is agency, questions about warmth alone are equally inadequate.
Weisman, Dweck and Markman argued that the agency-experience framework may capture perceived differences among beings but not exhaust how people classify mental life itself. Their work is an important validity check. A two-factor score should be treated as a useful operationalization for a defined question, not an ontological law or a guarantee that all respondents interpret items the same way. It may also matter whether participants are judging a robot body, a chatbot, a voice assistant or an imagined future system.
The label agency is especially easy to overextend. It may refer to perceived capacity for goal-directed action, autonomy from a user, planning or moral choice. An AI system can behave in goal-directed ways under software constraints without satisfying every human sense of agency. Participants might answer “yes” to a planning item because the system produces plans, while rejecting an intention item because they do not think it wants anything. A questionnaire should avoid treating those responses as contradictory without first defining the target construct.
Measure an attribution, not a capability claim
Mind-perception studies generally ask participants to rate what an entity seems capable of. The response is a human judgment. It cannot verify AI consciousness, subjective experience or moral status. Nor can it measure model accuracy, robustness or whether the AI really understands a task. Those require other methods and, for experience, questions that remain scientifically and philosophically contested. The measurement claim should be phrased as “participants attributed experience” rather than “the AI experienced.”
A capable system may receive low mind ratings when its interface is plainly mechanical. A weaker system may receive high ratings if it uses a human name, expressive voice and first-person language. This dissociation is not a measurement failure if the study is about perception. It becomes a failure when researchers describe ratings as evidence about the machine’s underlying abilities. Measuring actual task performance requires benchmark tasks and observations of behavior under known conditions, not a mind-attribution survey.
Respondents can also use words differently. “Can think” might mean “can solve problems” to one person and “has conscious thoughts” to another. “Can feel” might be interpreted as sensing input rather than having subjective experience. Cognitive interviews, pilot testing and careful item wording can show whether participants understood the intended contrast. Researchers may choose concrete capacities, such as “can make plans” and “can feel pain,” while acknowledging that even concrete items bring assumptions.
The target AI must be described precisely. A generic label “AI” may evoke a search engine for one participant, a companion chatbot for another and a science-fiction robot for a third. If a study compares groups, those private exemplars can create noise or bias. An experiment can show the same system to all respondents, specify its functions, and then ask what they attributed to that system after a standardized interaction. It should still record prior AI experience, which may change interpretation.
Surveys, drawings and interaction tasks
Rating scales are efficient. Participants can judge agency and experience items separately, and researchers can examine factor structure and consistency. Yet a scale can hide imaginative differences. Hwang and Won combined a two-dimensional mind-perception framework with drawing tasks to examine how people pictured AI entities. Their exploratory mixed-methods work showed that participants could imagine devices, human figures or broader technological ecosystems. The method offers insight into what target a respondent had in mind, but its small study context cannot establish population prevalence or validate a universal questionnaire.
Interaction tasks add another layer. A participant can converse with a chatbot, see it fail, receive a correction or observe it handling a sensitive request. Ratings after interaction may differ from ratings based on a static description. Such designs should report exactly what the system said and did. If AI outputs vary across participants, differences in ratings may reflect different content rather than stable person-level tendencies. Researchers can use scripted interactions to improve control or live systems to improve realism, explaining the tradeoff.
Repeated measures can reveal whether attributions shift. A user may initially rate an assistant as agentic because it acts without a prompt, then lower that rating after discovering that it follows a fixed template. Another may increase experience ratings after intimate conversation. These changes can be studied, but a rating shift is not proof of a durable bond or a true change in system capacity. It may be a short-term response to a salient phrase or framing.
Behavioral outcomes can complement self-report. Does a person seek the AI’s advice, disclose private information, give it credit or blame it for an error? These behaviors may be related to mind perception, but they are not pure measures of it. Disclosure may be driven by convenience or anonymity. Blame may be driven by organizational incentives. Researchers should measure the proposed attribution directly and then test how it relates to the behavior, rather than define mind perception by the behavior alone.
Anthropomorphism is a cue and a neighboring construct
Anthropomorphism can refer to perceiving humanlike qualities in a nonhuman entity or to design features intended to elicit such perceptions. A name, avatar, voice or conversational style may make an AI feel socially present. Yet humanlikeness is broader than mind attribution. A system can look like a person while being judged mindless, or have no humanlike appearance while being judged capable of planning. A study must keep the manipulated design cue apart from the measured mental-capacity attribution.
Reuter and colleagues reviewed anthropomorphism and trust in AI-related systems and ran conjoint analyses of names, appearance, voice and communication style. They found effects of voice and communication style on perceived humanlikeness and effects on trustworthiness in their studied designs. The authors also noted inconsistent operationalization of anthropomorphism in the literature. Their study is useful as an adjacent-method warning: changing perceived humanlikeness or trust does not automatically mean perceived experience changed. It supplies evidence about cues and construct confusion, not a validated agency-experience measure.
Manipulations should isolate cues where possible. If a “humanlike” version has a warmer voice, more useful answers, more conversational memory and a reassuring avatar, a resulting difference in mind ratings cannot identify which feature mattered. Factorial experiments can vary voice and answer quality separately. A manipulation check can ask whether participants noticed and interpreted the cue as intended. But a check using the same mind items as the outcome can blur the causal story. The design needs separate descriptions of cue exposure and attributed capacities.
There is also a language confound. People sometimes use intentional shorthand, saying “the system wants to correct the answer” without believing that it has desires. A survey item that asks whether it “wanted” something may catch shorthand rather than a literal mind attribution. Follow-up questions can ask what respondents meant and whether they would also attribute conscious feeling. This does not make everyday language invalid; it makes the inference from language to belief more careful.
Parasociality, support and reciprocity are separate
A person may enjoy a chatbot, feel accompanied by it or use it for support without attributing genuine experience. Parasocial attachment concerns a relationship-like orientation toward a mediated entity; mind perception concerns what capacities are attributed to the entity. They can covary, but neither is a substitute for the other. A warm interaction may produce a sense of social support because the response is timely and helpful, not because the user believes the AI suffers or cares.
Lee and Hahn examined mind perception and perceived social support in chatbot interaction. Their findings connect these perceptions within the studied user experience, while leaving open causal direction and generalization to other systems. A support rating cannot be read as an experience rating. A user might say “this helped me feel understood” as a report about their own experience, not a claim that the chatbot literally understood or reciprocated feeling.
Reciprocity is a further distinction. A human friendship usually involves two people who can each have needs, obligations and vulnerability. A chatbot may simulate conversational reciprocity through remembered context, polite questions and responsive language. A user can perceive the exchange as reciprocal, but that perception is not evidence of mutual subjective experience. Measures should ask whether participants felt answered, whether they believed the system had its own interests, and whether they expected obligations in each direction. Combining these into a single closeness score loses the distinction.
Trust and reliance also differ. Someone may attribute agency to an AI and distrust it because its choices seem opaque. Another may rely on a tool while judging it entirely mindless. If a study seeks to understand appropriate reliance, it should assess task performance and whether reliance tracks that performance. Mind perception may help explain behavior, but high mind ratings are not inherently good or bad.
Moral judgments and responsibility
Attributing a mind to AI can affect whom people blame. Joo tested mind perception and blame in scenarios involving AI-related moral transgressions. Across three studies, greater mind perception was associated with more blame directed at AI. Evidence that blame shifted away from human actors was more conditional and depended partly on how blame was measured. This illustrates both the relevance of the construct and a measurement limit: a forced-choice blame allocation can create competition among targets that an unconstrained rating does not.
Blame outcomes should not be used to define mind perception. A participant might blame the AI because a company presents it as the decision maker, because responsibility is unclear, or because the participant wants to hold someone accountable. The study should ask about perceived agency and experience separately, then examine whether those ratings predict blame after accounting for other cues. A scenario in which a system is described as autonomous may manipulate both perceived mind and perceived causal control.
Moral agency is also not the same as moral patiency. People may regard an entity as capable of causing harm but not as capable of being harmed through suffering. The agency-experience distinction is useful here, though the exact moral judgments are context dependent. Researchers should avoid a single “moral status” item if they want to separate who can be held responsible from who can be wronged.
Reliability and comparison across systems
A mind-perception instrument should have enough items to distinguish agency and experience without repeating the same wording. It should be tested for factor structure and item interpretation in the population and language of use. An item that worked for imagined robots in one survey may behave differently for contemporary language models. The system’s affordances also change quickly, so a questionnaire can age even if its wording stays fixed.
Comparing systems requires matched exposure. If one participant sees a text chatbot and another sees a voice assistant with a face, the score difference may reflect modality or presentation. If one system performs better on the task, agency ratings may partly track competent behavior. Researchers can hold outputs constant while varying labels, or measure output quality separately. They should report which approach they chose and what it leaves uncertain.
Individual differences matter. Prior beliefs about AI, religious or philosophical views of mind, familiarity with technology, language and age can influence how people interpret agency and experience items. A group mean may obscure subgroups that use the scale differently. Measurement invariance testing and qualitative follow-up can help determine whether the same score has comparable meaning across groups. Without that work, a numerical difference should be described cautiously.
Reliability across time also depends on the intended claim. A study of a single interaction needs to show that ratings are not mainly random response noise. A claim about a stable individual tendency to see minds in AI needs retest data across systems and contexts. A claim about one system’s interface needs replication with different tasks and samples. The bar rises as the proposed use moves from descriptive research toward individual profiling or design certification.
Reading a study carefully
Begin by identifying the entity participants judged. Was it a named model, a particular interaction, a generic imagined AI or a robot with a body? Then inspect the questions. Which items target planning and control? Which target pain, emotion or subjective feeling? Are humanlikeness, likability, trust and relationship quality measured separately? If not, a result may be too broad to support a claim specifically about mind perception.
Next examine what was manipulated. Was a design cue changed without changing performance, or did the system’s answers improve at the same time? Were participants’ expectations shaped by a description before they interacted? Did the study check what participants understood by “mind”? Consider whether scores reflect a literal belief, metaphorical language or uncertainty about the question. A direct item may still be ambiguous if its key term is undefined.
Finally, scale the conclusion. “Participants assigned higher experience ratings after this interaction” is a report about judgment. “They formed a reciprocal relationship with a conscious AI” is a different and unsupported claim. If the outcome is blame or trust, say so and distinguish it from the measured attribution. The Institute’s reading checklist is editorial synthesis from the cited methods and studies, not a validated instrument for scoring people or systems.
What the evidence does not show
Mind-perception research measures human attributions; it does not establish that an AI has subjective experience, intentions in the human sense or moral status. The agency-experience framework is useful but contested as a complete theory of mental concepts. A high score on one dimension does not prove a system possesses the rated capacity, and a low score does not settle every philosophical or scientific question about AI.
Humanlike design cues can alter trust or perceived humanlikeness without specifically changing perceived agency or experience. Feelings of support, parasocial attachment and apparent conversational reciprocity do not prove mutual awareness. Blame directed at AI does not absolve developers, deployers or institutions of responsibility. The cited experiments and surveys have task, sample and interface boundaries and do not yield a universal score for every AI system.
The most defensible claim is narrower: under documented exposure and wording, people attributed particular mental capacities to a specified system. A stronger explanation of why those attributions arose requires cue controls and alternative measures. Claims about actual capability require performance evidence, while claims about subjective experience remain separate from these perception studies.
Return to the worried reply
“I’m worried I may have misunderstood you” can be a useful conversational signal of uncertainty. It may also invite a reader to imagine an inner life. A measurement study could ask separately whether users think the system can identify an error, make plans, feel worry, care about them and act without instruction. It could compare this reply with plainer wording while holding task performance constant. Those answers would reveal an attribution pattern, not the system’s hidden state.
The value of measuring mind perception is in making human responses to AI more precise. It helps explain why the same interface can be treated as a tool, an adviser or a social partner. Precision also protects the boundary between what people feel the system is and what evidence shows it can do.
Research references
- Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619. View source
- Weisman, K., Dweck, C. S., & Markman, E. M. (2017). Rethinking people’s conceptions of mental life. Proceedings of the National Academy of Sciences, 114(43), 11374–11379. View source
- Hwang, A. H.-C., & Won, A. S. (2022). AI in your mind: Counterbalancing perceived agency and experience in human-AI interaction. CHI 2022 Extended Abstracts. View source
- Reuter, M., Kirchhoff, B. M., Franke, T., Radüntz, T., & Peifer, C. (2025). To trust or not to trust a human(-like) AI: A scoping review and conjoint analyses on factors influencing anthropomorphism and trust. Zeitschrift für Arbeitswissenschaft, 79, 402–432. View source
- Lee, I., & Hahn, S. (2024). On the relationship between mind perception and social support of chatbots. Frontiers in Psychology, 15, 1282036. View source
- Joo, M. (2024). It’s the AI’s fault, not mine: Mind perception increases blame attribution to AI. PLOS ONE, 19(12), e0314559. View source