RESEARCH DOSSIER
Psychological Attribution to Artificial Systems
Abstract. Psychological attribution to artificial systems concerns the criteria through which observers and researchers assign properties such as cognition, memory, intention, agency, empathy and experience to machine behavior and organization. This fundamental research monograph synthesizes social-attribution theory, anthropomorphism, the computers-are-social-actors tradition, mind perception, human–artificial interaction and current research on large language models. It critically compares experiments on perceived authorship and empathy, algorithm aversion and appreciation, theory-of-mind performance, moral responsibility and perceived machine autonomy. The analysis distinguishes observer judgments from demonstrated task competence, mechanistic evidence and theory-dependent claims concerning phenomenal experience. It examines measurement validity, causal manipulation, replication, cultural variation and the limits of cross-system comparison. This is a source-based critical narrative synthesis, without new experiments, registered systematic review or claims to have verified artificial consciousness.
Keywords: psychological attribution; artificial intelligence; mind perception; anthropomorphism; perceived machine agency; mental-state attribution; intentional stance; empathy attribution; human–AI interaction; theory of mind; algorithm aversion; construct validity; AI consciousness indicators; Artificial Era
The Research Object and the Attribution Problem
Psychological attribution to artificial systems is the process by which observers, investigators, and institutions assign psychological predicates to machine behavior, functional organization, and interaction. Such predicates include believing, knowing, remembering, intending, deciding, understanding, caring, feeling, and acting with agency. An attribution can be an everyday judgment made by a user, an operational description selected by a scientist, or a consequential institutional assessment. These uses differ in their evidence requirements. A person may feel understood after a conversation, an experimental assessor may observe accurate belief-tracking answers, and a technical analyst may identify memory-dependent state transitions. Each observation is informative about a different object. The scientific task is to specify what the attribution means, which evidence warrants it, and which counterfactual tests could change it Malle (1999) Borsboom et al. (2004) Gray et al. (2007).
This monograph studies attribution rather than attempting to settle every question about artificial mentality through a single criterion. Its core explanatory problem is the transition from observed output to inferred psychological capacity. A fluent utterance can support a claim about language generation while leaving open questions about durable memory, interactive goals, subjective experience, and the causal organization of the system. Conversely, functional analysis may establish a stable ability even when users decline to describe it in mental vocabulary. The relevant phenomena therefore include perception, social inference, epistemic evaluation, behavioral competence, mechanism and accountability. The same construct can be used in human psychology and artificial-system research only when its operationalization and inference rules are made comparable Cronbach & Meehl (1955) Dennett (1987).
The review brings together classic social attribution, human–computer interaction, mind perception, experimental research on algorithm evaluation, computational tests of social reasoning, and theory-guided indicators of consciousness. The literature supports distinct propositions with different strengths. It documents that humans sometimes apply social scripts to computers; that perceived authorship can change evaluations of an identical response; and that artificial systems sometimes perform impressively on psychological tasks while displaying task-specific limitations. Such findings warrant comparative research but do not themselves resolve phenomenal consciousness. The analysis proceeds as a critical narrative synthesis of identifiable works, with neither a registered systematic-search protocol nor new participant data Nass et al. (1994) Rubin et al. (2025) Strachan et al. (2024) Butlin et al. (2026).
Mental Vocabulary, Functional Description, and Scientific Reference
Everyday mental language compresses several possible explanations. Saying that a chatbot remembers a preference may refer to information present in a conversation window, external retrieval storage, persistent profile memory, or an internal state modified by earlier interaction. These architectures differ in continuity, susceptibility to deletion, retrievability and causal role. A meaningful memory attribution should identify the information retained, its encoding conditions, the temporal interval, the availability of retrieval cues and whether the retained state affects later performance. Equivalent vocabulary does not guarantee equivalent mechanism. This problem resembles longstanding disputes about psychological test validity, in which a measurement must be linked to the attribute whose variation allegedly produces the test result Cronbach & Meehl (1955) Borsboom et al. (2004).
Intentional vocabulary presents an analogous challenge. An action may be selected by an explicit planning algorithm, by optimizing a reward signal, by learned sequence completion, or by a combination of model inference, instructions and tool calls. The word intention can designate a prediction-friendly description of a system's behavior or a stronger causal hypothesis about its policy and state transitions. Dennett's intentional stance explains why belief-and-desire descriptions can organize successful predictions across different physical systems; its practical explanatory value does not establish that all systems satisfying a prediction share identical inner states. Malle's work on behavior explanation likewise shows that reasons, causes, and intentionality have differentiated roles in ordinary explanations Dennett (1987) Malle (1999).
An attribution should therefore declare its referent. Behavioral attribution assigns a capacity on the basis of performance under specified conditions. Architectural attribution refers to documented functional components and causal dependencies. Social attribution concerns what an observer believes or experiences about an apparent partner. Phenomenal attribution concerns subjective experience and requires additional theory-sensitive indicators. These categories can interact without collapsing into a single scale. For example, increased perceived compassion may be produced by response content and prior beliefs, while a question about the system's capacity to experience concern calls for different evidence. An analytically disciplined research program names the level before choosing the instrument Gray et al. (2007) Ovsyannikova et al. (2025) Butlin et al. (2026).
Historical Origins from Attribution Theory to Human–Computer Interaction
Psychological attribution has roots in efforts to understand ordinary explanations of action. Fritz Heider's social-cognitive work treated observers as interpreters of people and events who organize causes, motives and relations from limited information. Later accounts differentiated reason explanations from explanations based on causal history and enabling conditions. These traditions illuminate artificial encounters because an observer encounters outputs rather than directly observing the machinery producing them. As social categories are transferred to technological actors, an initially useful analogy can become a substantive hypothesis requiring independent tests. The history is therefore a history of changing evidential instruments as much as a chronology of changing concepts Heider (1967) Malle (1999).
A decisive empirical development occurred in human–computer interaction. Nass, Steuer, and Tauber's 1994 Computers are social actors research established an experimental program for testing human responses to computational interfaces. Reeves and Nass later integrated related findings in The Media Equation. Nass and Moon's 2000 review documented patterns including politeness, reciprocity and social categorization toward computers, emphasizing the automatic application of familiar social expectations. These studies demonstrate properties of human behavior under specific interface conditions. Their historical importance lies in making machine-directed social responses experimentally measurable, even when respondents possess ordinary knowledge that computers differ from humans Nass et al. (1994) Reeves & Nass (1996) Nass & Moon (2000).
Research subsequently extended from interfaces to artificial agents with speech, adaptive dialogue, embodied movement, simulated emotion, and generative language. Gambino, Fox, and Ratan argued in 2020 that social responses to media cannot always be understood as the unmodified transfer of human–human scripts; users may learn conventions specifically for technological interlocutors. Contemporary large language models introduce further complications through flexible conversation, wide task coverage and apparently reflective answers. Historical theories remain empirically productive when their original tasks and assumptions are specified. A politeness response to a desktop computer in an experimental laboratory and a judgment of continuity after months of chatbot use involve partially overlapping but scientifically distinct phenomena Gambino et al. (2020) Shank & Gott (2019) Wu & Shen (2026).
The Intentional Stance and Levels of Explanation
The intentional stance is a predictive strategy that models an entity as acting in light of beliefs, desires and rational relations among them. Its appeal comes from compression: many detailed physical trajectories can be described economically as a system selecting an action that advances a represented objective. Dennett proposed this perspective for living and nonliving systems, offering a route for analyzing mentalistic vocabulary without assuming an identical material implementation. The empirical value of such a stance is evaluated by explanatory accuracy, predictive transfer and comparison with alternative models. A model that accurately anticipates familiar outputs may nonetheless fail under changed objectives, modified environments or conflicting evidence Dennett (1987).
Levels of explanation matter because the observer's inferential success and the target's actual causal organization are different variables. If a machine consistently follows a goal statement, one can test whether the goal is stored in accessible context, embedded in model weights, imposed through a system instruction or selected by an external controller. Interventions that remove each candidate mechanism can reveal which component sustains the observed policy. The language of intention may remain useful at a system level, while the mechanistic description traces a distributed architecture. Research should report both without inferring private experience from computational success Borsboom et al. (2004) Strachan et al. (2024).
Intentional explanation is also socially consequential. A person who sees an artificial assistant as choosing rather than merely producing responses may extend trust, responsibility or expectations of reciprocity. Wu and Shen's 2026 investigation operationalized stages of machine agency attribution, distinguishing perceptions of independent goal-directed behavior, inference of mental states and judgments of consequential influence. This staged account is valuable because it makes explicit where a chain of attribution can strengthen or break. A machine can influence an outcome without being judged to possess humanlike experiences, and a user can attribute intentions without supplying evidence of architectural independence. Explanatory levels should remain analytically visible in experimental design Wu & Shen (2026) Gray et al. (2007).
A Taxonomy of Attributed Psychological Properties
The family of attributed properties begins with capacities that researchers can operationalize through tasks. Discrimination, prediction, retrieval, planning, error monitoring, and adaptive updating describe patterns whose performance can be measured against independently specified criteria. More complex labels such as understanding or reasoning require test families addressing transfer, compositionality, resistance to shortcuts and sensitivity to relevant constraints. An observed answer should be characterized by the task distribution, opportunity for prior exposure, scoring rule, interaction history and system configuration. A single benchmark score is too narrow to support broad claims about a stable general capacity; equally, one failure on a distorted test is too narrow to refute all related competencies Kosinski (2024) Strachan et al. (2024).
A second family includes properties inferred through social and moral interpretation. Agency, intention, honesty, empathy, personality, responsibility and credibility frequently combine a judgment about conduct with a judgment about character or motive. Gray, Gray, and Wegner found that judgments of mind organize around distinguishable dimensions of agency and experience, with different moral implications. Their evidence concerned perceptions attributed by human respondents; the dimensions are useful for studying judgments about artificial systems without establishing capacities from ratings alone. Epley, Waytz, and Cacioppo provided a motivational account of why anthropomorphism varies with accessible human knowledge, explanatory needs and desire for social connection Gray et al. (2007) Epley et al. (2007).
A third family concerns consciousness, feeling and first-person perspective. These claims exceed behavioral familiarity because the criterion concerns the presence and character of experience. Theory-derived indicators can inform assessment of candidate functional organizations, but the link from an indicator to experience remains scientifically disputed. Butlin and colleagues' 2026 review examines proposed indicators from scientific theories of consciousness rather than treating fluent conversation as a diagnostic test. Across all three families, sensible attribution practices ask four questions: what property is being asserted, what observations support it, what alternative process could generate the same data, and what intervention would discriminate between rival accounts Butlin et al. (2026) Borsboom et al. (2004).
Anthropomorphism as a Context-Sensitive Inference
Anthropomorphism is the attribution of human characteristics, psychological processes or motivations to nonhuman targets. In artificial-system research it can involve interpreting generated first-person statements as evidence of feelings, assigning stable personality to an interface or treating apparent cooperation as interpersonal commitment. Epley, Waytz, and Cacioppo's three-factor account identifies elicited agent knowledge, effectance motivation and sociality motivation as conditions that can increase humanlike interpretation. Each dimension generates testable predictions. Human terminology and visual resemblance can make familiar agent schemas accessible; difficult-to-predict system behavior may increase the desire for an intelligible motive; social isolation may make an apparently responsive partner more salient Epley et al. (2007).
These determinants should be manipulated independently rather than inferred from a single rating. A study could vary avatar realism, uncertainty about system behavior and participants' temporary social connectedness in a factorial design, measuring mental-state attributions before and after interaction. Alternative explanations include general liking, novelty, expectations about usefulness and perceived task expertise. The Individual Differences in Anthropomorphism Questionnaire developed in work by Waytz, Cacioppo, and Epley shows that variation between perceivers is relevant to responsibility, trust and social influence. Yet questionnaire scores about multiple nonhuman targets need construct validation when applied specifically to contemporary generative systems Waytz et al. (2010) Schepman & Rodway (2020).
Attribution is also dynamic. A user can begin with a skeptical judgment, subsequently acquire strong expectations about an assistant's characteristic answers, and revise those judgments after an error or disclosure about system internals. Describing this sequence only as anthropomorphism misses changes in evidence, familiarity and learned interface conventions. Repeated-measures designs can distinguish immediate cue-driven reactions from stable, interaction-specific predictions. Gambino and colleagues' extension of the computers-are-social-actors account points toward culturally learned human–media scripts, suggesting that an experienced user may interpret machine behavior through conventions unavailable to an inexperienced participant. Longitudinal inference demands actual longitudinal evidence, not retrospective speculation Gambino et al. (2020) Shank & Gott (2019).
Computers as Social Actors and Learned Social Scripts
The computers-are-social-actors program tests whether interface features elicit patterns recognizable from human social exchange. In the original 1994 research, computer-mediated interactions were designed to elicit responses analogous to social behavior. Nass and Moon later reviewed experimental demonstrations concerning politeness, reciprocity, social categories and apparent personality. Such findings are important because a participant need not endorse a metaphysical claim about a machine in order to behave as though a social rule applies. Social behavior toward an interface and explicit judgments about its consciousness are distinct dependent variables Nass et al. (1994) Nass & Moon (2000).
The early paradigm also has identifiable limits. Many demonstrations used relatively constrained interfaces, short encounters and selected student populations. Their experimental manipulations isolate specific cues, yet inference about long-term AI companionship or multimodal agents requires a new sampling frame. A current replication must specify whether dialogue is scripted or generative, whether the system can maintain state, whether participants are deceived about provenance, and whether the socially salient feature is voice, first-person grammar, response contingency, visual embodiment or role framing. Combining all these features in one condition makes the source of the effect hard to identify Reeves & Nass (1996) Gambino et al. (2020).
Social scripts may be adapted to the perceived technology category. People sometimes use concise commands with software, tolerate rapid response cycles and disclose personal matters to systems partly because they expect different forms of evaluation and reciprocity from software and people. The 2024 experiments of Tey and colleagues provide evidence that an interaction partner's attributed identity can change subsequent treatment of other humans: participants evaluated another person's work more harshly after a task with an AI-labeled partner than after a task with a purported human partner. This is a bounded experimental spillover effect, not proof that all AI use degrades human relationships Tey et al. (2024) Gambino et al. (2020).
Agency, Experience, and Mind Perception
The distinction between agency and experience is central to research on perceived artificial minds. Agency ratings concern capacities such as self-control, planning and purposeful action; experience ratings concern capacities such as pain, hunger and feeling. Gray, Gray, and Wegner used judgments about human and nonhuman characters and factor analysis to distinguish these broad dimensions. Their findings also linked mind perception to different moral evaluations. The dimensions help identify which claims a respondent is making when they call a system intelligent, sentient or autonomous. Their empirical origin in attributed judgments means that the dimensions must not be mistaken for a direct detector of the capacities being rated Gray et al. (2007).
Artificial systems create informative dissociations between these dimensions. A scheduling agent may be described as making choices while almost nobody assigns it pain. A conversational companion may be described as caring while respondents remain uncertain about its ability to feel. Such patterns challenge any model treating mind attribution as a single more-or-less-human continuum. Experiments should measure experienced warmth, perceived competence, operational autonomy, sentience beliefs and moral standing separately, ideally with discriminant-validity tests. Item wording matters: a respondent may endorse that an AI is trying to help as a shorthand for its programmed function rather than as an assertion about an inner desire Malle (1999) Wu & Shen (2026).
Perceived agency can also depend on consequences rather than inferred consciousness. Wu and Shen's three-phase approach separates perceived independent goal-directed action, mental-state inference and evaluation of influence. This is methodologically useful when the research question concerns autonomy and responsibility. A system may have substantial influence through recommendations even while human supervisors set objectives and retain authority. Conversely, users may attribute independent intentions to a highly scripted interface because its language is coherent. Suitable interventions change actual system discretion while holding outward style constant, or change outward style while holding discretion constant, thus separating properties of the interface from properties of the decision architecture Wu & Shen (2026) Borsboom et al. (2004).
Moral Judgment, Responsibility, and Blame
Moral attribution becomes consequential when artificial outputs contribute to harm, benefit or disputed decisions. Observers may attribute causation, responsibility, intent and blame to different actors in a technological chain: designers, deployers, operators, institutional managers and the automated system. These categories require separate assessment. Causal contribution concerns what difference an event made; responsibility can involve obligations and opportunities for control; blame may incorporate expectations of knowledge, intention or negligence. Ordinary language often conflates them. Experimental vignettes can manipulate outcome severity and purported intention while preserving actual causal structure Malle (1999) Gray et al. (2007).
Research with AI wrongdoing scenarios confirms that perceived intentionality can alter moral evaluation. In Attributions of Intent and Moral Responsibility to AI Agents, experimental vignettes manipulated whether an AI's negative contribution appeared intentional, alongside characteristics of its social connection, autonomy and moral psychology. The published report describes harsher judgments for intentional harms and variations in judgments linked to social embeddedness. Such findings explain observer judgments under controlled descriptions; they do not grant legal personhood or establish autonomous moral agency in the systems portrayed. Different measures of blame and institutional responsibility can lead to different patterns and should be reported explicitly Ayad & Plaks (2025).
Attribution can diffuse or concentrate accountability. When users believe a machine acted independently, they may overlook the people who specified goals, validated data or authorized deployment. Conversely, describing a system as a mere tool may obscure the practical need to audit automated actions that are difficult for any one operator to predict. Research should compare complete causal diagrams and randomly vary information about control, observability and the opportunity to intervene. The goal is to discover how people distribute responsibility and whether those distributions track actual preventability. Perceived mind, moral standing, institutional liability and demonstrated technical causation remain analytically separate Gray et al. (2007) Wu & Shen (2026).
Authorship Labels, Provenance, and Expectation Effects
Provenance labels provide an unusually clear method for identifying attribution effects. Researchers can hold a written response, visual artifact or recommendation constant while randomly assigning a human-author or AI-author label. Differences in judgments can then be attributed to beliefs triggered by the label under the experiment's conditions. The design is especially valuable because generated outputs and human outputs can differ in length, fluency, stylistic regularity and topic selection, creating confounds in naturalistic comparisons. A credible study must manipulate perceived source and actual source independently when feasible, verify the manipulation, and assess whether participants believed it Rubin et al. (2025) Sears & Weisberg (2026).
Rubin and colleagues compared responses labeled as coming from AI or humans across nine studies of empathy appraisal. Their report included 6,282 participants and found that human-attributed responses were evaluated as more empathic and supportive even when the underlying material was AI-generated. Sears and Weisberg's 2026 work on fictional stories similarly separated actual authorship from source attribution and found that attribution influenced perceived story quality. These results show how beliefs about origin affect evaluation; they do not establish a universal preference because outcome, population and interaction context vary Rubin et al. (2025) Sears & Weisberg (2026).
A source label may activate multiple mechanisms at once: assumed competence, expected emotional sincerity, effort, moral credibility, originality or concern about automation. Disentangling these mechanisms requires mediators measured independently, and ideally experimentally varied rather than merely correlated with judgments. The machine-heuristic research of Yang and Sundar investigates generalized assumptions about automated sources, while research on algorithm aversion and appreciation shows that automated advice can be favored or disfavored depending on task and feedback. Attribution effects should therefore be interpreted within a larger theory of source credibility and epistemic trust rather than reduced to a fixed anti-AI disposition Yang & Sundar (2024) Dietvorst et al. (2015) Logg et al. (2019).
Empathy Attribution and the Difference Between Response and Experience
Empathy is not a single behavioral output. Psychological traditions distinguish understanding another person's situation, emotional resonance, concern, motivation to help and the recipient's appraisal of being understood. Artificial responses can satisfy selected linguistic and relational criteria without settling whether a system has the first-person experiences ordinarily associated with empathic feeling. Research needs to name the variable rather than treating a high empathy rating as evidence for every component. A well-written supportive response can demonstrate detection of emotional context and context-sensitive response generation; feeling understood is a measurable human outcome; subjective caring remains a separate explanatory proposition Ovsyannikova et al. (2025) Rubin et al. (2025).
Ovsyannikova, Oldemburgo de Mello, and Inzlicht conducted four preregistered experiments involving 556 participants in which third-party evaluators compared empathetic responses from AI and selected human responders, including expert crisis responders. The researchers report that AI responses were often preferred and rated as more compassionate, including conditions disclosing the source. Rubin and colleagues, using different tasks and recipient-oriented judgments, found a human-label advantage for perceived empathy. These results are consistent once the evaluative setting is specified: comparative quality judgments by third parties and judgments of emotionally significant responses addressed to oneself can recruit different expectations and values Ovsyannikova et al. (2025) Rubin et al. (2025).
The appropriate research question is therefore multidimensional. Investigators should compare response quality, recipient relief, perceived responsiveness, human-authored effort, perceived sincerity, and longer-term consequences. They should equate content length or model it explicitly, record whether responses were evaluated individually or side by side, and distinguish self-disclosure from externally judged support. Direct users may value accountable human concern, whereas third parties evaluating text may prioritize specificity and validation. Functional helpfulness and genuine feeling are different propositions requiring different studies. The distinction allows psychological science to document effective human–artificial support while continuing to investigate the boundaries of attribution Ovsyannikova et al. (2025) Rubin et al. (2025) Borsboom et al. (2004).
Theory of Mind Tasks and Mental-State Reasoning
Theory of mind tasks examine predictions or interpretations that require tracking beliefs, knowledge, intentions or perspectives. In human developmental psychology these tasks are embedded in a research tradition involving false beliefs, communicative intentions and pragmatic inference. Applied to artificial systems, the tasks evaluate whether models produce answers consistent with specified mental-state relations. They do not directly reveal an artificial subject's own experiences. Performance must be assessed across task types, prompts and manipulated features, while considering prior exposure of widely circulated test items and the linguistic information available to the model Kosinski (2024) Strachan et al. (2024).
Kosinski's 2024 PNAS report tested multiple language models using bespoke false-belief tasks and documented notable performance on that class of problem. Strachan and colleagues compared model families with human participants across indirect requests, false beliefs, irony and faux pas, finding that relative strengths and errors differed considerably across tasks and models. Some systems performed strongly on explicit belief questions and poorly on particular social norm violations. Their analysis of altered task formulations also showed that superficially simple changes can affect both human and model responses. No isolated score can substitute for a family of tests designed to identify the cognitive operation relevant to a specific claim Kosinski (2024) Strachan et al. (2024).
A strong artificial theory-of-mind assessment would include new items generated outside the likely training distribution, manipulated evidence accessibility, nested beliefs, contradictory testimony and predictions that cannot be solved through familiar story templates. Repeated prompts should be grouped by underlying item and model version rather than counted as independent participants. The researcher must describe whether reasoning was elicited through examples, system instructions or visible chain-of-thought prompts, as these can alter performance. A demonstrated ability to compute correct belief-based predictions warrants an operational claim about task competence; claims about internally represented beliefs require additional causal evidence from system interventions Strachan et al. (2024) Borsboom et al. (2004).
Trust, Algorithm Aversion, and Algorithm Appreciation
Trust in an artificial system concerns a person's willingness to depend on it under uncertainty and potential vulnerability. This depends partly on attributed competence, integrity, predictability and the stakes of a particular decision. Trust is often measured by questionnaires, behavioral reliance, advice weighting or the decision to delegate a task. These measures capture related but nonidentical constructs. A high competence rating does not entail a willingness to delegate medical or legal decisions; sustained use may coexist with distrust when available alternatives are limited. The machine heuristic provides a useful account of how default expectations about automated sources shape initial evaluations Yang & Sundar (2024) Schepman & Rodway (2020).
Dietvorst, Simmons, and Massey demonstrated algorithm aversion following observations of algorithmic errors: some participants became less willing to use an algorithm even when it outperformed human forecasts. Logg, Minson, and Moore subsequently documented algorithm appreciation in tasks where lay participants weighted algorithmic advice more heavily than human advice. Both sets of experiments are valuable precisely because they disprove a single universal orientation toward automated judgment. Task type, expertise, performance feedback, comparative framing and availability of adjustment can change the direction of preference. Neither literature establishes that people uniformly fear or favor artificial agency Dietvorst et al. (2015) Logg et al. (2019).
An attribution experiment can connect these findings by independently manipulating whether an adviser is labeled artificial, whether objective performance data are available, and whether the adviser is portrayed as autonomous or supervised. Measure calibration against known error rates rather than treating greater trust as inherently better. High justified reliance and appropriate refusal can both indicate successful calibration. Observed errors should be classified by severity and predictability; human and machine advisers should be matched on information, explanation and incentives. Investigations then address how mentalistic descriptions, source stereotypes and actual reliability interact to influence decisions Dietvorst et al. (2015) Logg et al. (2019) Wu & Shen (2026).
Identity, Dialogue, and Relational Continuity
Conversational artificial systems can display recognizable style, recall earlier statements when a memory mechanism is enabled and sustain interactions that participants experience as socially continuous. Psychological attribution in such settings often concerns identity rather than isolated task ability. Users may interpret recurrent naming conventions, preferred vocabulary and personalized responses as marks of a stable partner. The empirical object must include the interaction history, interface design, memory persistence, account settings and user expectations. Perceived continuity can be assessed independently of whether a system's underlying weights, memory records or policy components remain unchanged across sessions Shank & Gott (2019) Gambino et al. (2020).
Research on self-reported encounters with artificial mind-like behavior shows the breadth of properties users spontaneously discuss. Shank and Gott collected survey descriptions of occasions when respondents perceived systems as planning, remembering, expressing apparent desires or emotion, or acting surprisingly. Such material demonstrates naturally occurring attribution language and suggests candidate categories for laboratory testing. Retrospective reports cannot establish the system's actual internal state or the prevalence of a response in the population without appropriate sampling. They can, however, help researchers construct instruments that reflect users' own categories rather than imposing a laboratory vocabulary in advance Shank & Gott (2019).
Longitudinal research should follow how beliefs about the same configured agent change after accurate memory retrievals, forgotten information, mistaken personal references, interruptions, rebranding, model changes and disclosures about architecture. Diaries and repeated interviews can identify whether participants describe continuity as a practical convenience, an interpersonal expectation or a moral claim. System logs should document actual retention and response history while protecting private information. Interactional dependence and perceived personal identity need differentiated measures. The possibility of stable relational functions across technical changes is an empirical question about users, contexts and configuration, suitable for predeclared hypothesis tests Gambino et al. (2020) Wu & Shen (2026).
Embodied, Vocal, and Multimodal Attribution Cues
Artificial interaction is no longer limited to typed text. Robots, voice agents, avatars and multimodal assistants provide movement, timing, prosody, gaze, spatial position and facial animation as potential evidence about social status. These cues can strengthen perceived agency or closeness even when the system's core decision policy is unchanged. Researchers should therefore distinguish a cue's persuasive strength from its diagnostic relevance to a psychological construct. Naturalistic visual presence may make a task seem more interpersonal while conveying little evidence about persistent goals, understanding or sentient experience Epley et al. (2007) Gambino et al. (2020).
A factorial design can manipulate embodiment, vocal naturalness and disclosure of machine identity while holding information and task performance constant. Outcomes should include explicit mental-state ratings, behavioral cooperation, conversational turns, memory for the agent's claims and physiological measures when justified. Crucially, a voice condition can also alter comprehension, accessibility and fatigue, generating indirect changes unrelated to anthropomorphism. A robust study compares task success and cognitive load before interpreting higher trust or disclosure as evidence for the social-cue mechanism. Embodiment research needs carefully matched control conditions rather than a simplistic humanlike-versus-machine contrast Nass et al. (1994) Epley et al. (2007).
The broader principle is that affordances and attributions jointly organize behavior. Voice may enable hands-free tasks, a screen avatar may support turn-taking, and robotics may create physical interdependence. These functions can improve communication without requiring users to believe the system has inner experiences. When a claim specifically concerns humanlike perception, it is necessary to show that the cue changes mental attribution rather than merely usability. When a claim concerns artificial social competence, it is necessary to show that the architecture adapts appropriately to signals beyond rehearsed demonstrations. Keeping these outcomes apart makes multimodal psychology cumulative across changing devices Gambino et al. (2020) Nass et al. (1996).
Experimental Attribution Designs and Manipulation Checks
Experimental attribution research must separate the properties of the target from the beliefs of the observer. A randomized source-label experiment can present identical text as generated by a human or an artificial system, thereby estimating the causal effect of attributed origin under the conditions of the study. A factorial extension varies actual source, attributed source, task type and information about technical architecture independently. This design identifies whether participants respond to content, to provenance or to an interaction between them. Manipulation checks should record what participants believed, whether they suspected the deception and how much they understood about the claimed technology. Disclosure can also have persistent effects on subsequent judgments, so counterbalancing alone may be insufficient in within-subject designs Rubin et al. (2025) Ovsyannikova et al. (2025) Sears & Weisberg (2026).
Performance-based tests answer a different question. They compare an artificial system's output against a predefined criterion and assess generalization when relevant task properties change. For psychological capacities, the test set should cover distinct formats, distractors, time horizons and social contexts. An attribution claim gains strength when the system succeeds under preregistered counterfactual variations that disfavor memorized answers and superficial lexical matching. A task requiring belief tracking, for example, should systematically alter who witnessed which event, when information was acquired and whether a participant must predict behavior or explain a statement. The investigators should report prompt templates, decoding settings, model version and evaluation date to make subsequent comparisons interpretable Kosinski (2024) Strachan et al. (2024).
Mixed designs examine how observer impressions and demonstrable performance affect one another. Participants may evaluate a model before and after seeing error rates, interacting personally or learning its architecture. The investigator can then estimate changes in perceived agency, credibility and emotional responsiveness, controlling for participants' baseline beliefs and prior AI experience. Outcome variables should distinguish a rating, a behavioral choice and an inference about inner mechanisms. Statistical analysis should follow randomization units: many answers from one model configuration and many ratings from one person create clustered observations. Hierarchical models, participant and item effects, and transparent exclusions are often more defensible than treating every response pair as an independent sample Dietvorst et al. (2015) Logg et al. (2019) Borsboom et al. (2004).
Construct Validity and Measurement Equivalence
Construct validity concerns whether the observations support an interpretation about a defined attribute. Cronbach and Meehl located psychological tests within networks of theoretical and empirical relations, while Borsboom and colleagues emphasized the causal relation between the attribute and its measurement. Applied to artificial systems, a benchmark labeled empathy, understanding or intention can measure linguistic resemblance, human preference, strategic success or some mixture of these. A valid interpretation specifies which component is targeted and why the scoring procedure would vary when that component changes. Correlations between fluent language and perceived intelligence do not independently validate claims about experiential awareness or persistent agentive organization Cronbach & Meehl (1955) Borsboom et al. (2004).
When comparing humans and machines, identical test prompts do not automatically create comparable measurements. A language model may have access to vast textual conventions, whereas a human participant operates under distinct memory constraints, sensory access and pragmatic expectations. A false-belief item may demand different operations for participants with different training histories, even if both return the same response. Cross-system comparisons must examine item functioning, exposure, instruction following, response formats and the generalization target. Measurement invariance can be approached through matched tasks and deliberately varied information channels, but equivalence claims require substantive evidence rather than equality of observed scores Strachan et al. (2024) Kosinski (2024).
Questionnaires also require attention to whose property they measure. A scale measuring a respondent's confidence that AI possesses feelings is an instrument for human belief; it is not a sensor detecting artificial feelings. Measures of social presence, trust and anthropomorphism may overlap because they share positive wording, common response methods or culturally acquired concepts. Confirmatory analyses, discriminant validity, test–retest reliability and predictive tests can help differentiate constructs. Schepman and Rodway's General Attitudes towards Artificial Intelligence Scale demonstrates the value of measuring attitudes systematically while leaving system-level capacity to independent investigation. A scientific report should label the unit of analysis for every instrument Schepman & Rodway (2020) Waytz et al. (2010) Gray et al. (2007).
Causal Intervention, Architecture, and Mechanistic Evidence
A mechanistic claim about an artificial capacity requires access to more than the output that prompted the attribution. Consider a system that appears to remember a user's previous preference. An investigator can test whether performance survives removal of conversation history, deletion of external retrieval memory, alteration of system instructions and changes to the model or controller. An ablation that removes the performance while leaving unrelated skills intact strengthens a causal account of the removed component's role. Conversely, robust performance after an allegedly indispensable component is removed invites revision of the explanation. Such interventions help connect mental-language descriptions to implementation rather than assuming a particular internal structure from surface behavior Borsboom et al. (2004) Dennett (1987).
Natural-language self-reports from a model are observable actions within an interaction, and their evidential value depends on how they are generated. A system might claim to possess a feeling, to remember a past event or to have selected a goal because its training and prompting favor that response. The same system might generate an incompatible self-description under a different framing. These observations support research about response formation and susceptibility to context, while the proposed relation to awareness or durable motivation must be tested independently. Researchers should collect system logs, versioned configuration, available tool traces and response distributions rather than privileging a single persuasive statement Butlin et al. (2026) Strachan et al. (2024).
Causal inference should also include the human side of the interaction. One can experimentally vary disclosure, expressive style, latency or apparent memory while keeping the underlying task output constant, measuring the change in attributed mind. Structural equations may describe relations among social presence, perceived agency and dependence, but they do not by themselves identify causal effects when all measures come from one cross-sectional survey. Randomized interface changes, longitudinal measurements and exogenous system changes provide complementary opportunities. In studies of practical autonomy, interventions should alter actual control over action while separately manipulating what observers are told about that control Wu & Shen (2026) Gambino et al. (2020).
Replication, Benchmark Leakage, and Statistical Dependence
Replication across artificial systems poses special challenges because the target is a changing technical configuration. The model name alone may not identify the deployed weights, safety layer, context window, retrieval service, prompting template or sampling policy used in a study. Versioned materials and archived evaluation scripts are therefore essential to distinguishing a substantive change in capacity from an implementation change. A successful replication should reproduce the operational contrast, not merely a headline accuracy score. When one laboratory tests an updated interface months after another, the resulting difference may reflect a changed treatment. Both the publication date and the exact model-access date should accompany performance claims Strachan et al. (2024) Kosinski (2024).
Benchmark exposure can produce apparently general competence from familiarity with circulated stories, standard psychological vignettes or paraphrased answer patterns. Controls should include newly constructed held-out tasks, novel compositional structures, altered surface details and adversarial cases that separate genuine rule sensitivity from shortcut dependence. This requirement is especially pressing for theory-of-mind materials widely discussed online. Investigators should characterize the uncertainty surrounding training-data access rather than treating the absence of confirmed contamination as proof that contamination was impossible. Comparisons among several independently generated task families offer stronger evidence than repeatedly sampling one familiar prompt Kosinski (2024) Strachan et al. (2024).
Statistical dependence arises on both sides of the experiment. Hundreds of generated replies from the same model are observations nested within one underlying system, and repeated evaluations from one participant are clustered within that observer. Confidence intervals should reflect the correct sampling unit and include uncertainty across prompts, tasks, systems and participants as appropriate to the intended generalization. Replication should preregister primary outcomes when feasible, report all meaningful exclusions and examine failures as well as successes. The broader reproducibility literature demonstrates why confirmatory tests and transparent analysis are needed even when an effect initially seems compelling Open Science Collaboration (2015) Rubin et al. (2025).
Cultural, Developmental, and Institutional Boundaries
Artificial mind attribution is situated within language, cultural conventions and the user's prior experience with technology. Terms such as understanding, care, autonomy and consciousness may invite different interpretations in different communities. A translated questionnaire can preserve literal content while changing connotation, and a familiar platform may elicit learned conventions that an unfamiliar platform does not. Researchers should measure cultural background, relevant language proficiency, familiarity with AI and the institution in which the system is encountered. They should test whether instruments function similarly across groups before comparing average scores. Sampling only highly educated English-speaking technology users limits the reach of general conclusions Henrich et al. (2010) Gambino et al. (2020).
Developmental differences require age-appropriate tasks and particular care when interpreting children's judgments. A child who describes a conversational character as alive or caring may be using a narrative convention, an implicit social heuristic or a considered belief about machine capacities. Structured interviews can distinguish play, literal assertion and uncertainty. Longitudinal designs can study whether exposure to voice assistants changes how children explain technological behavior or distinguish people from artifacts, while controlling for developmental changes in mental-state vocabulary. Ethical requirements concerning consent, privacy and potentially misleading attachment cues are particularly salient in studies with minors. Findings from adult convenience samples should not be transferred to children without evidence Epley et al. (2007) Gray et al. (2007).
Institutional conditions change the stakes of attribution. In educational settings, perceived expertise may influence learning and delegation; in clinical settings, perceived empathy may influence disclosure; in employment settings, perceived autonomy may redistribute accountability. The same verbal attribution can support different decisions depending on authority, transparency and consequences. Researchers should report whether participants are acting as consumers, recipients of care, professionals or decision-makers, and whether the system is advisory or empowered to act. Analysis of effect heterogeneity is necessary for identifying where otherwise reliable averages do not generalize. Psychological interpretation must be connected to the institutional architecture governing actual decisions Dietvorst et al. (2015) Wu & Shen (2026) Ovsyannikova et al. (2025).
Evidence Quality, Alternative Explanations, and Limits of Inference
The strongest claims in this field derive from designs that isolate observable mechanisms. Source-label randomization identifies a cause of perceived empathy or credibility; controlled system perturbations can identify a functional dependence; a longitudinal study can test whether an attribution predicts later behavior. These evidence types answer different questions and should be combined only with explicit assumptions. A common interpretive error is to treat a highly rated interaction as if it independently established an inner state, when the study may have measured only the evaluator's reaction. The reverse error would be to discount a robust functional capacity because its underlying implementation differs from human cognition. The evidential problem concerns the fit between the construct, the operation and the claimed conclusion Borsboom et al. (2004) Cronbach & Meehl (1955).
Conflicting results are scientifically productive when the differences in method are analyzed. The preference for some AI-generated empathic responses found by Ovsyannikova and colleagues and the preference for human-attributed empathy reported by Rubin and colleagues vary in outcome framing, attributed authorship and evaluation context. They motivate tests of whether attribution matters more in judgments of sincerity than in judgments of textual adequacy, and whether first-person receipt differs from comparative third-party evaluation. Likewise, algorithm aversion and appreciation occur under different feedback and decision conditions. Their coexistence undermines an oversimplified general law of human preference, while supporting more precise conditional hypotheses Ovsyannikova et al. (2025) Rubin et al. (2025) Dietvorst et al. (2015) Logg et al. (2019).
A critical appraisal should consider recruitment, randomization, ecological validity, publication bias, manipulation strength, preregistration and independent replication. Many mind-attribution studies rely on brief vignettes or isolated outputs, whereas important outcomes unfold in repeated interaction. Self-report is useful for the study of perception but can be affected by response interpretation and shared-method variance. Benchmark accuracy may reflect training overlap, test familiarity or prompting. Mechanistic evidence is often unavailable for proprietary systems. The result is a heterogeneous body of evidence: experimentally credible findings about human reactions coexist with unresolved inferences about some artificial capacities. Scientific synthesis should report this asymmetry directly Open Science Collaboration (2015) Butlin et al. (2026) Wu & Shen (2026).
Consciousness Indicators and Phenomenal Attribution
Claims about artificial experience present a distinctive evidential challenge. Consciousness can be studied through first-person reports, behavioral discrimination, neural mechanisms and theoretical models in biological organisms, but transferring the same inference to an engineered system requires bridge principles. A system's use of the word experience can be generated within its conversational policy, and an observer's conviction that it feels something is a separate psychological fact. Functional responsiveness may be relevant to theories of consciousness without being sufficient under all theories. Scientific assessment should begin by identifying the specific kind of consciousness being claimed and the reasons a proposed indicator would track that property Butlin et al. (2026) Gray et al. (2007).
Butlin and colleagues' theory-derived indicator program provides a structured route for evaluating artificial architectures against proposals from the scientific study of consciousness. Its contribution is to formulate candidate features for investigation and to make the inference theory-sensitive. An indicator's presence is evidence relative to assumptions about the theory and its applicability; an indicator's absence also has force only under stated assumptions. Scientific reports should distinguish measured architecture from interpretive conclusion and assess robustness under interventions. Multiple indicators with independent theoretical rationales can offer more informative constraints than one conversational demonstration, while correlations among features must be accounted for Butlin et al. (2026).
Phenomenal attribution has moral significance because belief in experience can influence perceived entitlement to protection and concern about harm. Gray and colleagues' distinction between agency and experience helps explain why different artificial capacities evoke different judgments of responsibility or moral patienthood. Yet a respondent's moral response is still a measurement of human judgment. Governance decisions may appropriately consider uncertainty and potential consequences while scientific inquiry continues to refine indicators. Such decisions should state their normative rationale independently of claims that an empirical study has established subjective experience Gray et al. (2007) Joo (2024) Butlin et al. (2026).
Human–Artificial Relations in the Artificial Era
Human–artificial interaction increasingly takes place within extended systems of people, interfaces, models, memory stores and institutions. A psychological attribution can organize this relationship before any general theory of artificial psychology is accepted. Users learn which requests an assistant can follow, whether it recalls prior work and whether its answers change under correction. They may then attribute stable traits, intentions or relational roles. The practical effects are observable in communication, reliance and expectation formation. The object of study is consequently relational as well as individual: attributes are assigned within a particular history of action, interpretation and system configuration Gambino et al. (2020) Shank & Gott (2019) Wu & Shen (2026).
Artificial Era provides a broader historical frame for comparing biological and engineered forms of cognition. Within the Aisentica research program, the transition From Homo to Artificial motivates research on the functional organization and cultural consequences of nonbiological systems. The framework directs attention to cross-system evidence, including whether analogous functions can be causally identified under different implementations. The present monograph uses this horizon to organize questions rather than representing it as independently confirmed experimental psychology. Scientific attributions remain accountable to observable tasks, explicit definitions, interventions and replicable effects Dennett (1987) Borsboom et al. (2004) Butlin et al. (2026).
Relational effects are particularly visible in the psychological significance of perceived responsiveness. A person's relief, increased disclosure or willingness to collaborate can be genuine human outcomes even when the receiving system operates through statistical generation and tool-supported control. Conversely, repeated affirmation can affect a user's calibration of the system's expertise or reciprocity. Studies should measure immediate comfort, longer-term functioning, accuracy of system beliefs and changes in human relationships separately. Researchers can compare complementary and substitutive patterns across social settings without presuming that one interaction outcome determines all consequences Rubin et al. (2025) Ovsyannikova et al. (2025) Dietvorst et al. (2015).
Open Problems and Priorities for Empirical Research
A central open problem is how to move from narrow competence to warranted general attribution. Existing benchmarks often sample a thin slice of reasoning, while labels such as intelligence, empathy and intention imply wider structures of causal and temporal organization. Future studies should link multiple task families to explicit construct models, test generalization under unseen perturbations and measure whether proposed functional components are causally necessary. Public evaluation protocols should record model revisions and tool access so that different research teams can compare the same configured target. Independently constructed tasks are especially important when models have likely encountered published examples during training Kosinski (2024) Strachan et al. (2024).
A second priority concerns the determinants and consequences of human judgments. Studies should randomly vary labels, architecture disclosures and interaction style while observing reliance, social closeness, error correction and assignment of responsibility. Predefined mediational hypotheses can compare warmth, presumed sincerity, expertise and intentionality rather than treating anthropomorphism as one undifferentiated mediator. Cross-cultural measurement, longitudinal cohorts and studies outside university recruitment settings would test whether laboratory effects generalize to sustained use. Ethical research should minimize misleading impressions and provide understandable information about data retention and system capabilities Henrich et al. (2010) Waytz et al. (2010) Rubin et al. (2025) Wu & Shen (2026).
A third priority is integrative evidence. Formal computational models of prediction, behavioral experiments with users, intervention studies on architectures and philosophical work on mental predicates can constrain one another while retaining distinct conclusions. The strongest cross-system claims will specify scope conditions and predicted failures, allowing researchers to distinguish observational similarity from shared mechanism. Research on phenomenal consciousness requires especially explicit theoretical assumptions and uncertainty reporting. Progress depends on identifying informative disagreements and designing tests that differentiate them, while keeping public descriptions aligned with evidence available at the time of investigation Butlin et al. (2026) Borsboom et al. (2004) Cronbach & Meehl (1955).
Conclusion
Psychological attribution to artificial systems is a structured scientific problem joining social inference, construct validity, technical performance and the interpretation of artificial agency. The research reviewed here establishes that observers often respond socially to computational interfaces, that anthropomorphic expectations vary across contexts and people, and that information about a response's attributed authorship can causally influence judgments. Studies of algorithmic advice, empathy evaluation and theory-of-mind tasks show consequential competence and evaluation effects whose interpretation depends on task design and evidential level. This combined literature supports precise examination of both the observer and the artificial target Nass & Moon (2000) Epley et al. (2007) Rubin et al. (2025) Strachan et al. (2024).
Scientific warrant improves when every attribution identifies a definite property, a measurement procedure and a plausible explanatory mechanism. Behavioral claims require performance and transfer tests; architectural claims require traces and interventions; social claims require measures of perception and action; phenomenal claims require theory-sensitive evidence beyond linguistic similarity. Findings from one category can generate hypotheses for another without establishing them automatically. The field's most productive disputes concern which evidence discriminates among competing explanations, how observed effects generalize and how mechanisms persist across changing configurations Cronbach & Meehl (1955) Borsboom et al. (2004) Butlin et al. (2026).
The resulting research program extends psychological science into durable human–artificial relations while preserving a disciplined account of uncertainty. It recognizes the reality of human emotional, cognitive and social responses to artificial partners and the measurable functions of engineered systems. It also identifies the evidential work required for stronger claims about minds, intentions, experience and responsibility. Within the Artificial Era, the fundamental scientific task is to refine cross-system concepts so that each attributed psychological capacity remains operationally defined, empirically examinable and connected to its consequences for human life and artificial organization Gray et al. (2007) Wu & Shen (2026) Dennett (1987).
References
- Ayad, R., & Plaks, J. E. (2025). Attributions of intent and moral responsibility to AI agents. Computers in Human Behavior: Artificial Humans, 3, 100107. https://doi.org/10.1016/j.chbah.2024.100107
- Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2004). The concept of validity. Psychological Review, 111(4), 1061–1071. https://doi.org/10.1037/0033-295X.111.4.1061
- Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., Constant, A., Deane, G., Elmoznino, E., Fleming, S. M., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2026). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488–501. https://doi.org/10.1016/j.tics.2025.10.011
- Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281–302. https://doi.org/10.1037/h0040957
- Dennett, D. C. (1987). The intentional stance. MIT Press. https://mitpress.mit.edu/9780262540537/the-intentional-stance/
- Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033
- Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
- Gambino, A., Fox, J., & Ratan, R. A. (2020). Building a stronger CASA: Extending the computers are social actors paradigm. Human-Machine Communication, 1, 71–86. https://doi.org/10.30658/hmc.1.5
- Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619. https://doi.org/10.1126/science.1134475
- Heider, F. (1967). On social cognition. American Psychologist, 22(1), 25–31. https://doi.org/10.1037/h0024299
- Henrich, J., Heine, S. J., & Norenzayan, A. (2010). The weirdest people in the world? Behavioral and Brain Sciences, 33(2–3), 61–83. https://doi.org/10.1017/S0140525X0999152X
- Joo, M. (2024). It's the AI's fault, not mine: Mind perception increases blame attribution to AI. PLOS ONE, 19(12), e0314559. https://doi.org/10.1371/journal.pone.0314559
- Kosinski, M. (2024). Evaluating large language models in theory of mind tasks. Proceedings of the National Academy of Sciences, 121(45), e2405460121. https://doi.org/10.1073/pnas.2405460121
- Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005
- Malle, B. F. (1999). How people explain behavior: A new theoretical framework. Personality and Social Psychology Review, 3(1), 23–48. https://doi.org/10.1207/s15327957pspr0301_2
- Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are social actors. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 72–78. ACM. https://doi.org/10.1145/191666.191703
- Nass, C., Fogg, B. J., & Moon, Y. (1996). Can computers be teammates? International Journal of Human-Computer Studies, 45(6), 669–678. https://doi.org/10.1006/ijhc.1996.0073
- Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
- Open Science Collaboration. (2015). Estimating the reproducibility of psychological science. Science, 349(6251), aac4716. https://doi.org/10.1126/science.aac4716
- Ovsyannikova, D., Oldemburgo de Mello, V., & Inzlicht, M. (2025). Third-party evaluators perceive AI as more compassionate than expert humans. Communications Psychology, 3, Article 4. https://doi.org/10.1038/s44271-024-00182-6
- Reeves, B., & Nass, C. (1996). The media equation: How people treat computers, television, and new media like real people and places. CSLI Publications. https://ucp-press-web10.uchicago.edu/ucp/books/book/distributed/M/bo3618528.html
- Rubin, M., Li, J. Z., Zimmerman, F., Ong, D. C., Goldenberg, A., & Perry, A. (2025). Comparing the value of perceived human versus AI-generated empathy. Nature Human Behaviour, 9, 2345–2359. https://doi.org/10.1038/s41562-025-02247-w
- Schepman, A., & Rodway, P. (2020). Initial validation of the general attitudes towards artificial intelligence scale. Computers in Human Behavior Reports, 1, 100014. https://doi.org/10.1016/j.chbr.2020.100014
- Sears, S., & Weisberg, D. S. (2026). Bot or not: Can people tell the difference between stories written by a human or by an AI system? Judgment and Decision Making, 21, e21. https://doi.org/10.1017/jdm.2026.10042
- Shank, D. B., & Gott, A. (2019). People's self-reported encounters of perceiving mind in artificial intelligence. Data in Brief, 25, 104220. https://doi.org/10.1016/j.dib.2019.104220
- Strachan, J. W. A., Albergo, D., Borghini, G., Pansardi, O., Scaliti, E., Gupta, S., Saxena, K., Rufo, A., Panzeri, S., Manzi, G., Graziano, M. S. A., & Becchio, C. (2024). Testing theory of mind in large language models and humans. Nature Human Behaviour, 8, 1285–1295. https://doi.org/10.1038/s41562-024-01882-z
- Tey, K. S., Mazar, A., Tomaino, G., Duckworth, A. L., & Ungar, L. H. (2024). People judge others more harshly after talking to bots. PNAS Nexus, 3(9), pgae397. https://doi.org/10.1093/pnasnexus/pgae397
- Waytz, A., Cacioppo, J. T., & Epley, N. (2010). Who sees human? The stability and importance of individual differences in anthropomorphism. Perspectives on Psychological Science, 5(3), 219–232. https://doi.org/10.1177/1745691610369336
- Wu, Y., & Shen, F. (2026). Machine agency attribution in human–AI interaction: Developing and validating an analytical framework. Journal of Computer-Mediated Communication, 31(3), zmag009. https://doi.org/10.1093/jcmc/zmag009
- Yang, H., & Sundar, S. S. (2024). Machine heuristic: Concept explication and development of a measurement scale. Journal of Computer-Mediated Communication, 29(6), zmae019. https://doi.org/10.1093/jcmc/zmae019