Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

AI-Generated Art and Attribution Effects on Aesthetic Judgment

Abstract. This fundamental research monograph examines how actual, believed, and disclosed human or artificial authorship changes the aesthetic reception of visual art. It integrates experimental label manipulations, blinded discrimination and preference studies, models of aesthetic processing, mind ascription, authenticity research, and studies of perceived labor, artistic credit, emotion, and moral evaluation. Central questions concern the causal effect of source labels on otherwise identical images, the separation of sensory preference from judgments about makers, and the competing roles of perceived intention, effort, meaning, social legitimacy, and context. The evidence demonstrates attribution-sensitive evaluations in several preregistered and randomized experiments, while also documenting null total effects, positive blind evaluations of AI-produced images, and sharp differences between credit and beauty judgments. Methodological analysis addresses stimulus sampling, forced-choice versus rating tasks, mediation identification, individual differences, generalizability, and reproducibility. The synthesis identifies testable research designs for human–artificial co-creation in the Artificial Era. This is a critical narrative synthesis of published research, not an original experiment, preregistered systematic review, or new meta-analysis.

Keywords: AI-generated art; aesthetic judgment; authorship attribution; artist labels; aesthetic preference; human–artificial art; perceived agency; mind ascription; artistic authenticity; creativity; human effort; art valuation; algorithm aversion; blind evaluation; visual preference; experimental aesthetics; comparative choice; semantic context; awe; human–artificial co-creation; Artificial Era

The Scientific Object: Attribution and Aesthetic Judgment

The research object of this monograph is the change in an observer’s evaluation of a visual artwork when its production is attributed to a human, an artificial system, or a human–artificial collaboration. An attribution effect is operationally defined as a difference in judgment attributable to information about the supposed source of an image, ideally after its visible properties have been held constant through random assignment. The relevant judgments include liking, beauty, interest, perceived meaning, creativity, artistic skill, effort, authenticity, moral acceptability, willingness to pay, and the classification of an object as art. These outcomes are related but empirically separable. A label can alter the estimated value of the artist’s contribution without altering the beauty ascribed to the depicted object, as Khan et al. (2026) demonstrates. A theory of AI-art reception must therefore specify the measured response before speaking of a general bias.

This study develops a critical narrative synthesis of experimental attribution research, perceptual comparison studies, models of aesthetic appreciation, mind-perception research, and findings on artistic authenticity. It investigates whether author labels causally alter ratings, how perceived intentionality and effort contribute, why results vary with measurement and display, and which explanations remain untested. Experimental labels, self-reported beliefs, and objectively recorded production methods must be distinguished throughout. Studies of paintings and illustrations provide the strongest direct evidence for this particular object; findings on writing, music, commercial images, and algorithmic decisions are used only to clarify potential generalizations. The empirical literature includes negative, null, and occasionally favorable responses to machine attribution. The organizing conclusion is that authorship information changes the interpretive situation in which aesthetic judgment occurs; the sign and magnitude of any particular effect depend on what observers are asked to judge Bellaiche et al. (2023) Horton et al. (2023) Messingschlager & Appel (2025).

Construct Definitions and the Problem of Source Identity

An AI-generated image is a visual product for which a generative computational system has participated in the production of the final image. This description leaves open the level of human participation in conceptualization, prompting, selection, masking, editing, and presentation. Research designs often simplify the production history into a binary human-versus-AI label, but the simplification belongs to the experimental manipulation. It does not establish that real creative production contains only two mutually exclusive agencies. A valid study records objective provenance, the description supplied to participants, their believed provenance, and their judgment of the human contribution separately. The human–AI distinction can describe software involvement, credited authorship, inferred intentionality, or perceived ownership; treating these as interchangeable introduces construct contamination. Experimental comparisons of personalized and standard AI reveal why the proportion and type of human contribution must be coded independently of the visual result Khan et al. (2026).

The target phenomenon also differs from detecting images generated by AI. A participant may identify a synthetic artifact correctly yet prefer it, misclassify a human artwork yet admire it, or dislike an image for its appearance regardless of its label. The conditional probability of accurate identification is a classification performance outcome; the difference in evaluation between identical images bearing different labels is a causal attribution outcome; the association between guessed origin and liking is an observational attribution-belief outcome. These three quantities answer different scientific questions and require different inferential designs. Their distinction is essential when interpreting research that finds favorable blind responses to AI images alongside negative reactions to images believed to have been produced by AI Grassini & Koivisto (2024) van Hees et al. (2025).

Historical Foundations of Artificial Art Reception

Artificially mediated art did not begin with contemporary text-to-image models. Earlier computer-generated visual works, algorithmic graphics, and robotic art already raised questions about whether procedural production could support authorship, creativity, and aesthetic value. The historical object is the public reception of artwork whose production process is described as computational, not a single generation of software. The transition to diffusion and other contemporary generative systems increased the accessibility and diversity of machine-assisted images, thereby making author labels relevant in ordinary browsing, exhibitions, sales, and news discourse. Research on computer-generated art before the recent generative-AI surge provides a baseline for analyzing how people responded to computational origin before contemporary systems became widespread Chamberlain et al. (2018) Hong & Curran (2019).

A historically disciplined account separates technical innovation from evidence about spectators. The emergence of a new model or visual style does not by itself show a change in human aesthetic perception; that claim requires judgments collected under documented conditions. Likewise, rhetorical declarations that art is intrinsically human are philosophical or cultural positions rather than measures of the causal effect of an attribution label. Experimental psychology contributes by comparing evaluations while manipulating disclosed origin, observing identification accuracy, and testing interpretations about agency, labor, and authenticity. In this perspective the history of digital art is important because it changes the plausible meanings of a source label, the distribution of experience with machine-made images, and the social conventions surrounding credit. These variables have not remained constant, and historical findings should be generalized with care Chamberlain et al. (2018) Małecki et al. (2025).

Aesthetic Processing as an Organized Psychological Activity

An aesthetic judgment arises within a sequence of perceptual analysis, memory activation, classification, interpretation, and evaluation. The multistage model developed by Leder et al. (2004) provides a foundational account of how stimulus features, expertise, conceptual knowledge, and cognitive mastering contribute to judgments of artworks. Applied to AI attribution, the model predicts that an authorship label can enter the process through explicit classification and cognitive appraisal even when the image itself remains identical. The resulting evaluation need not be the immediate emotional reaction to color or form. It may reflect a subsequent judgment about whether a work demonstrates originality, expresses an intention, or deserves social recognition. This analytic distinction prevents a culturally loaded rating from being mistaken for a direct measure of low-level perceptual pleasure.

Art reception involves feedback as well as progression. A viewer may initially enjoy an image, read that it was generated by a model, reconsider the meaning of apparent brushwork, and then revise an assessment of skill or value. An apparently paradoxical pair of judgments—high beauty and low deservingness of artistic credit—can be psychologically coherent when they answer different questions. The history of aesthetic processing models also makes clear that expertise and semantic information alter the use of visual cues. Research on ambiguity indicates that meaning construction is part of art appreciation, while contemporary experiments find that interpretive descriptions can modify the devaluation associated with AI authorship Jakesch & Leder (2009) Wang et al. (2026). These findings support a process model with multiple entry points for source information.

Processing Fluency, Familiarity, and Perceptual Pleasure

Processing fluency is the subjective ease with which perceptual or conceptual operations proceed. The theoretical review by Reber et al. (2004) proposes that fluent processing can contribute to positive aesthetic experience, while acknowledging the importance of context and attribution of processing ease. Familiar patterns, visual clarity, repeated exposure, and congruent expectations may facilitate some stages of perception. A claim that AI-generated artwork is necessarily less pleasant at the sensory level would therefore require showing that the images themselves disrupt perceptual processing under blinded conditions. Such an inference cannot be obtained from lower ratings after a participant receives an AI label. The label manipulates semantic and social information, not the pixels of an image.

The distinction between automatic and more controlled processing has become experimentally tractable. Across three experiments using AI-generated stimuli with manipulated artist labels, Wang et al. (2026) reports that negative reactions associated with AI attribution were stronger when conditions encouraged controlled evaluation and reduced when interpretation was supported by semantic cues. This pattern is consistent with an interpretive component that develops after the initial encounter. It does not establish that all automatically processed reactions are equivalent, that every image behaves identically, or that fluency is the sole mediator. Timing, attention, response deadline, instruction wording, and stimulus repetition can themselves change the task. A rigorous fluency explanation therefore requires direct process measures as well as outcome ratings.

Authenticity, Originality, and the Value of a Creative History

Aesthetic value includes forms of evaluation that refer to an object’s history. Studies of authenticity show that viewers and consumers can attach special value to originals because of their origin and their relation to an artist, even when a copy is visually comparable. Newman & Bloom (2012) experimentally investigated why original artworks are valued over duplicates and distinguished different explanations associated with creative performance and personal connection. In AI-art judgment, an authorship label can similarly alter the perceived history of production. The observer may ask whether a specific person chose the composition, struggled with materials, generated a distinctive idea, or invested time in a recognizable trajectory of work. Such assessments belong to provenance-sensitive valuation rather than to retinal discrimination.

Originality should be treated as a multidimensional construct. Statistical novelty, stylistic unfamiliarity, an innovative artistic intention, historical priority, and the evaluator’s sense of creative merit are distinct criteria. A generated image may appear novel while its production method receives low perceived credit. Conversely, an original human performance may receive high attributional value even when the image is visually ordinary. Questions about legal originality, copyright, moral rights, and ownership cannot be decided by a preference score; they require their own legal and institutional criteria. The psychological finding concerns how beliefs about a creative history enter value judgments. The separation of artistic credit from visual appreciation is empirically visible in comparisons of personalized and standard generative systems Khan et al. (2026) Horton et al. (2023).

The Matched-Image Attribution Experiment

The most direct test of source-label causation holds the artwork constant and randomly assigns participants to descriptions that identify the source as human or AI. When the assignment is randomized, the average difference in outcome estimates the effect of the information package in the studied population and context. Its interpretation depends on the exact wording: “AI-generated,” “made by an artist using AI,” and “produced by a personalized model trained on the artist’s work” communicate different information. The same image must not be seen repeatedly in a way that exposes the deception or produces uncontrolled demand effects. Counterbalancing, stimulus-level random effects, and prespecified outcome measures help isolate the attribution manipulation from properties of the selected artworks Bellaiche et al. (2023) Khan et al. (2026).

A study in which participants compare different human-created and AI-created paintings estimates a mixture of image differences and source-related differences unless origin is crossed independently with stimulus identity. Even a perfectly balanced set of images may contain recognizable artifacts, style cues, or differences in subject matter that influence preference. A stronger design crosses a selection of images with randomly assigned descriptions and, when feasible, records actual origin as a separate factor. Misleading labels have ethical and methodological costs: participants must be debriefed when required by the research protocol, and the authors must distinguish experimental labels from real provenance in reporting. The causal quantity is bounded by the particular labels, images, sample, display platform, and assessment question; it is not a universal constant of human taste Horton et al. (2023).

The Bellaiche Experiments and Human Authorship Preference

The experiments reported by Bellaiche et al. (2023) provide a particularly informative manipulation because artworks that participants saw were AI-generated while the assigned authorship information varied. This design made it possible to investigate whether the assumed maker changed ratings independent of the image’s actual production origin. The research found more favorable judgments for images described as human-made across several evaluative dimensions, while investigating explanations connected with perceived human effort and emotional or conceptual depth. Because the stimuli remained AI-generated, the contrast supports an effect of the communicated authorship information within the particular experimental setting rather than a claim that human-produced pixels were visually superior. It is a direct demonstration that an object can receive different evaluations when its perceived production history changes.

The theoretical significance concerns how people interpret craftsmanship and expression. A source label can function as a condensed narrative about the maker: a person may have intentions, experiences, motives, and costly labor invested in an artwork. Observers may use that narrative when asked whether an image is profound, meaningful, beautiful, or worthy of appreciation. Yet the measured outcomes should remain separate and the specific experimental mechanisms should not be declared proven solely because effort and meaning correlate with the label effect. For mediation claims, alternative pathways, measurement order, and participant inference require examination. The authorship information in this study is a manipulated social cue whose effect may vary with audience expertise, broader attitudes toward AI, and the art style being evaluated Bellaiche et al. (2023) Leder et al. (2004).

The Horton Program: Devaluation and Human Creativity

Across six experiments involving 2,965 participants, Horton et al. (2023) examined how AI and human authorship labels shape the reception of visual art. The research used multiple sets of stimuli and included samples designed to provide broader coverage of US audiences than a single university convenience sample. AI attribution was associated with devaluation across a range of judgments, including assessments linked to effort, artistic skill, and monetary value. Its implications extend beyond liking: the findings document the social dimension of evaluation in which the estimated work of a creator and the status of a work as a costly human accomplishment become part of the response. The authors also tested conditions under which labels and comparisons change assessments of human creativity.

An especially important observation is relational. When human-labeled works were evaluated alongside AI-labeled works, the comparison could increase perceived creativity attributed to humans. This result suggests that ratings of human ability are partly contrastive: a new comparison category changes the frame against which an existing category is judged. Such a finding does not imply that a person’s capacity to create has objectively increased, nor that visual artwork created without AI has become more sophisticated. It concerns comparative social judgment. The design is also useful for locating possible boundary conditions because an experimentally observed effect on value or effort may be stronger than the corresponding effect on image beauty. A universal statement that AI art is aesthetically rejected would conceal this differentiation Horton et al. (2023).

Mind Ascription and the Problem of Null Total Effects

Mind perception provides a psychologically specified explanation of source effects. Gray et al. (2007) distinguished important dimensions of how observers attribute capacities such as agency and experience to agents. These are judgments made by perceivers; a score does not certify the presence or absence of subjective states in the target. Applied to visual art, a human-maker description may invite assumptions about intentions, emotional experiences, and expressive purposes. An AI-maker description can reduce those assumptions, potentially changing the perceived significance of an image. This pathway connects social cognition to aesthetic cognition without requiring an experiment on machine consciousness. The construct at issue is the observer’s represented maker and its presumed capacities.

In two experiments, with analyzed samples of 176 and 381, Messingschlager & Appel (2025) found that artwork attributed to AI was ascribed lower levels of agency and experience, and the statistical mediation models identified negative indirect paths to art appreciation. Crucially, the overall total effect of AI versus human artist information on appreciation was not statistically significant in the experiments. This result must be presented accurately. A negative indirect effect alongside a nonsignificant total effect is compatible with multiple pathways operating in different directions, measurement uncertainty, and model-dependent decomposition. It does not warrant the claim that the study independently established a consistent average decline in overall appreciation. The authors’ own results invite a multicausal account in which expected novelty, technical accomplishment, artistic category, and social meaning may interact Messingschlager & Appel (2025).

Perceived Agency, Intention, Effort, and Skill

Agency in attribution research is an inferred capacity to initiate, regulate, or direct action. Intentionality is the degree to which a viewer supposes that a particular effect was selected or meant by a maker. Effort denotes estimated costs of production, which may concern hours of work, skill acquisition, problem solving, or the difficulty of realizing a chosen image. These variables should be measured with different items. A viewer can believe that a machine executed an intricate rendering while a person supplied the idea; another can believe that the prompt author made only a trivial choice. Asking for a single percentage of “creativity” obscures the dimensions of the judgment. Studies finding stronger effects on labor, credit, and skill than on beauty illustrate the need to separate them Horton et al. (2023) Khan et al. (2026).

A scientifically strong account treats the alleged production history as a causal input that changes an observer’s model of action. The label can reshape beliefs about planning, iteration, accountability, and expressive motivation. Nevertheless, mediation analyses should include competing explanations: unfamiliarity, social identity, worries about the labor market, distrust of platforms, the value of originality, and beliefs about intellectual property can influence judgments without passing uniquely through mind ascription. The conceptual relations do not establish experimental causal pathways. Random manipulation of production descriptions combined with independent measures of inferred intention, effort, and affect is a better foundation for distinguishing mechanisms than retrospective ratings collected after the primary evaluation Gray et al. (2007) Messingschlager & Appel (2025).

Real Origin, Believed Origin, and Blind Visual Preference

The evaluation of an unlabeled image speaks to its visual or compositional properties under the available task conditions. The ability to identify a generated image speaks to learned classification criteria and available visual artifacts. The inclination to penalize an image that one believes is synthetic speaks to social and interpretive attribution. The design used by Grassini & Koivisto (2024) makes this divergence visible: 201 participants assessed paintings and reported beliefs about whether the works had human or AI origins. Participants did not consistently distinguish the two sources, and the study reported favorable evaluation of some actual AI images alongside negative evaluations associated with perceived AI origin. Because perceived source was reported rather than independently assigned, the negative association may also reflect preexisting reactions to particular images.

A distinct blind-perception approach is illustrated by van Hees et al. (2025), which examined observers’ experiences with art in a setting where identifying the creator was itself a separate research issue. A person may respond positively to the surface structure of a generated work while showing a culturally patterned reaction when its source becomes known. Those observations are compatible rather than contradictory. The correct inference is that sensory preference, source detection, and disclosed-provenance judgment are different outcomes. A synthesis must resist combining effect sizes from such designs without accounting for which underlying quantity is estimated, whether the labels were randomized, and whether stimuli were representative of the aesthetic universe to which conclusions are being extended Grassini & Koivisto (2024) van Hees et al. (2025).

Computer Art, Contemporary Generative Models, and Historical Comparisons

Experimental research on computer-generated art predating widely available generative-image interfaces demonstrates that the psychological issue has historical continuity. In Chamberlain et al. (2018), the conditions under which artworks were presented and their computer-generated status mattered to aesthetic responses; observing a robotic artist could alter evaluations. Such work draws attention to embodiment and demonstrations of production as cues through which the maker is represented. A moving machine visibly performing an artistic action can communicate planning, competence, or process in a manner quite different from a short written label. Therefore, seemingly identical authorship categories are not always psychologically equivalent across historical technologies or media.

Other studies complicate uniform predictions. Hong & Curran (2019) examined attitudes to human-produced and AI-produced art and did not establish a simple universal law that any machine label lowers every measure of artistic value. The relation between beliefs about artificial intelligence and the evaluation of artworks can depend on how the question is framed. Accordingly, a contemporary meta-theory should specify which image-generation technology participants imagined and what “AI” meant to them. The label may refer to an autonomous artist, a digital tool, a commercial platform, a statistical image generator, or an assistant used by a named human. The psychological construct carried by the words changes as public knowledge changes; comparisons between experimental eras require evidence of semantic as well as measurement equivalence Chamberlain et al. (2018) Hong & Curran (2019).

Semantic Context and the Reconstruction of Meaning

Source information is one kind of semantic context. In a foundational neuroaesthetic experiment, Kirk et al. (2009) manipulated descriptions associated with artworks and observed changes in aesthetic valuation and associated neural response. That study was not a contemporary AI-label experiment; its scientific contribution here is to establish that identical visual content can be evaluated differently when its surrounding information changes. It provides a general bridge between perceptual processing and contextualized interpretation. Gallery attribution, institutional endorsement, price, artist identity, and production history are all candidates for cues that influence expectations or the meaning a viewer attempts to construct.

The amount of interpretive assistance supplied with an image is particularly important for AI art. Wang et al. (2026) reported that semantic cues offering a route to meaning construction reduced the negative bias that emerged under more controlled processing in their experimental paradigm. A plausible explanation is that viewers may be less willing to elaborate an image when they presume no expressive human intention; a description can give them an alternative cognitive entry into the work. That account remains a model requiring direct tests of the proposed mental sequence. Semantic descriptions can also introduce demand, alter expectations, and increase processing time. Future studies should separate the informational content of the description from mere added attention, text length, or positive evaluative language Jakesch & Leder (2009) Wang et al. (2026).

Moral Evaluation, Legitimacy, and Process Disclosure

An observer can value an image aesthetically yet regard its production as objectionable. Moral concerns may include the treatment of source artists, labor displacement, uncertainty about training data, commercialization, or a perceived evasion of human effort. These concerns are not equivalent to a measured failure of artistic beauty. Bara et al. (2025) examined the effects of contextual information about AI on moral and aesthetic judgments of AI-generated visual art. Their preregistered experimental work showed that details about the process could reduce moral acceptability and aesthetic appeal in studied settings. The effect of information should be attributed to the information supplied, not simplified into a verdict about every image made with artificial systems.

A disclosure is thus a compound psychological stimulus: it can improve knowledge of provenance while evoking new interpretations of fairness, ownership, authenticity, and social consequence. Experiments must specify whether the cue concerned the artwork’s actual origin, an ethical claim about the model, or a broader warning. Aesthetic pleasure, moral legitimacy, willingness to display, and willingness to purchase can diverge sharply. Studies that merge them into one global attitude score risk losing the mechanism. The normative question of what disclosure art institutions owe viewers is also different from the descriptive question of how disclosure changes attitudes. The latter is empirically testable through randomization; the former requires ethical and institutional reasoning informed by, but not exhausted by, the measured effect Bara et al. (2025) Newman & Bloom (2012).

Awe, Emotional Resonance, and Social Connection

Emotion provides another dimension of art reception. Aesthetic pleasure, interest, admiration, awe, feeling moved, and empathy have different eliciting appraisals and consequences. The social meaning of perceived authorship can enter these appraisals because audiences often interpret creative expression as evidence of another being’s experiences or intentions. Across five preregistered studies involving 1,598 participants, White & Ponce de Leon (2026) investigated the proposition that works believed to be AI-generated evoke less awe and thereby reduce downstream empathic responses. The study covered more than one artistic medium and population setting. It supplies direct evidence about the relation between AI attribution and particular emotional and interpersonal outcomes, rather than merely a generalized preference for human-looking imagery.

An important interpretive limit is that awe and empathy are not interchangeable with image beauty. A perceiver may admire a technical image without experiencing interpersonal connection with its presumed maker, or may feel moved by the story of a person behind a modest visual work. The finding should therefore lead to more differentiated emotion research, using validated state measures and designs that specify whether the emotional effect concerns content, producer beliefs, or participants’ social expectations. It also motivates investigation of mediating pathways such as perceived experience, narrative transportation, and empathic identification. Such pathways are promising hypotheses and should be tested independently rather than inferred automatically from the presence of an emotional outcome White & Ponce de Leon (2026) Gray et al. (2007).

Human–Artificial Co-Creation and Personalized Tools

The binary human-versus-AI condition only partially describes artistic practices that include human decision making across multiple stages. In real workflows, a person may develop a concept, edit an image, direct a sequence of prompts, curate many generations, or modify a model with personal material. Perceived authorship may therefore be distributed across planning, execution, selection, revision, and presentation. Experiments with such workflows must report which elements were actually performed and which were merely represented to participants. A generic “AI-assisted” label may be interpreted as effortless substitution or as demanding creative work depending on the social narrative supplied. The contribution of an identifiable human agent should be measured rather than presumed.

In two studies with a combined sample of 774 UK participants, Khan et al. (2026) presented identical artworks with descriptions emphasizing hand production, standard AI use, or personalized AI. Descriptions of personalized systems increased attribution of credit and authorship to the human user relative to standard AI, without a corresponding improvement in aesthetic appreciation or acceptance of the image as genuine art. This dissociation is theoretically decisive: belief in a human’s contribution and enjoyment of the object can be partly independent responses. Personalized tools may make contribution more legible while leaving other normative or stylistic judgments unchanged. The evidence supports a differentiated psychology of co-creation rather than the claim that every increase in credited human effort mechanically increases beauty Khan et al. (2026).

Individual Differences, Expertise, and Attitudes Toward Technology

Attribution effects can vary between observers. Relevant participant characteristics include art education, experience with creative software, familiarity with generative systems, general attitudes toward automation, occupational identification as an artist, and expectations about the purpose of art. The empirical findings in Grassini & Koivisto (2024) indicate that openness to experience and creative identity are associated with aspects of AI-art evaluation and perceived origin. Because these variables are often measured observationally, their associations cannot independently establish that a personality trait causes either a favorable or unfavorable reaction. An art expert and a casual social-media viewer may also respond to different features when asked whether a work is innovative or worthy of exhibition.

A sound moderator analysis requires more than reporting separate significance tests within demographic groups. It needs a prespecified interaction, a justified measurement of the moderator, adequate statistical power, and credible comparability of rating scales. Familiarity with AI tools might reduce novelty-related distrust while increasing awareness of automation or dataset concerns; both possibilities should be investigated empirically. More general evidence from algorithm aversion and appreciation demonstrates that attitudes to artificial decision makers depend on domains and task structures. Dietvorst et al. (2015) and Logg et al. (2019) do not study art reception directly, and their findings should be treated as boundary-setting context rather than borrowed proof of an AI-art preference effect. Domain-specific judgments must be established through art-specific designs.

Rating Scales, Response Anchors, and Evaluation Targets

A single item asking whether an image is “good” is ambiguous. Respondents may answer in terms of attractiveness, craft, novelty, authenticity, social deservingness, or personal enjoyment. Scientific measurement therefore begins with explicit construct definitions, separately anchored questions, and a model of how items relate to latent dimensions. Beauty, liking, interest, expressive depth, judged creativity, effort, estimated monetary value, willingness to purchase, and artistic legitimacy should not be averaged into one unexamined index. The consequences are visible in studies where label effects differed across value and effort judgments and across general appreciation outcomes Horton et al. (2023) Messingschlager & Appel (2025). A null result for one scale is informative about that scale, not every psychological response to the artwork.

Measurement order also changes interpretation. Asking respondents first whether a source has a mind can prime intentionality before a beauty rating; asking for price before liking can turn the experience into a market appraisal. A design that elicits source beliefs after the rating risks contamination through post hoc inference from the judgment. Reliable research reports question wording, response anchors, scale reliability where meaningful, trial order, instructions, image viewing time, exclusions, and the handling of participants who distrust the label. Longitudinal or cross-cultural comparisons require evidence that the constructs have comparable meanings over time and across populations. The central target should be a theoretically interpretable estimand, such as the average causal change in rated beauty for identical images under different labels.

Comparative Choice Versus Isolated Rating

The format of evaluation can itself determine what observers reveal. An isolated Likert judgment asks for an absolute rating in a particular context; a forced-choice task asks respondents to select one item relative to another. The first may produce compressed distributions when most images are attractive, while the second can heighten sensitivity to small preferences or to salient author labels. Jacobs et al. (2026) compared an isolated rating design with 250 participants and a two-alternative forced-choice design with 102 participants, observing differing evidence across formats. The results warn against treating an absence of a mean rating difference as proof that no preference appears when participants must choose between alternatives.

Forced choice has its own vulnerabilities. The outcome depends on the pairing algorithm, visual similarity, response time, displayed labels, and the relative attractiveness of alternatives. It may increase salience of human–AI categories and therefore amplify comparative social judgments. A rigorous evaluation program uses both isolated judgments and choices, with randomized placement and counterbalanced images, and states in advance which decision environment is being represented. Research intended to illuminate museums, marketplaces, or social platforms should choose tasks resembling the consequential behavior of interest, while maintaining sufficient experimental control to identify the role of disclosure. Neither rating format is inherently a universally superior measure of aesthetic judgment Jacobs et al. (2026) Horton et al. (2023).

Stimulus Sampling, Style, and Generalizability

The images selected for an experiment are a sample from a large and changing world of artworks. Generated paintings, photo-realistic portraits, abstract images, commercial illustrations, conceptual pieces, historical reproductions, and physical paintings cannot be assumed to elicit equal attribution effects. An experiment using only attractive, stylistically homogeneous computer images may overestimate visual acceptance; one using conspicuously flawed outputs may confound low beauty with source identification. To support a claim about AI art in general, a study should sample multiple genres and levels of quality, document the exact production systems, record prompt or selection procedures when relevant, and model image variation in the analysis. This issue is central when interpreting results from studies that use heterogeneous sets of art Horton et al. (2023) van Hees et al. (2025).

Repeated judgments of multiple images per person generate dependence. Treating every individual rating as statistically independent can inflate apparent precision, while analyzing only participant means discards meaningful item variation. Cross-classified multilevel models, preregistered image-selection procedures, and independent stimulus replication improve generalizability. It is also important to distinguish source-label effects that replicate across a given set of images from those that hold across new art genres, new generation models, and new audiences. Replicating a label contrast while reusing one stimulus set is evidence for the stability of the experiment, but does not settle the broader population-of-artworks question. The universe of images is evolving, and scientific conclusions should name its sampled boundaries.

Statistical Causality and the Mediation Problem

Random assignment to a label allows an experimentally grounded estimate of the effect of that label on a measured outcome, conditional on the assumptions of the design. It does not automatically identify why the change occurred. A mediation model connecting AI label, perceived mind, and appreciation requires assumptions about unmeasured causes of the mediator-outcome relation, the temporal order of the measures, and alternative mechanisms. A statistically significant indirect coefficient can coexist with a nonsignificant total treatment effect when different pathways counterbalance each other. The findings of Messingschlager & Appel (2025) provide an instructive empirical example and therefore require reporting the whole model, including its null total effect, rather than selecting only the pathway that fits a theoretical expectation.

Future mechanism studies should experimentally manipulate plausible explanatory variables where feasible, such as the visibility of creative effort, intentional descriptions, process recordings, or personalized training narratives, while keeping other aspects of the artwork constant. They should measure both baseline attitudes and post-manipulation beliefs. Preanalysis plans can identify the primary aesthetic outcome and reduce opportunistic selection among numerous rating dimensions. Confidence intervals and uncertainty for interactions and indirect effects matter at least as much as threshold p values. A theory is strengthened when it predicts which aesthetic dimension will change, under which label and which context, while also predicting circumstances where no effect should appear Khan et al. (2026) Bara et al. (2025).

Neural Evidence and Its Interpretive Limits

Neuroaesthetic findings show that contextual information can accompany changes in neural activity while people judge visual works. Kirk et al. (2009) examined modulation of aesthetic valuation by semantic context and reported associations involving regions implicated in valuation. Such observations are valuable for locating contexts in which perceptual and conceptual information interact. They do not determine whether an artificial system possesses an inner experience, nor can they prove that a viewer’s positive or negative assessment is intrinsically correct. Functional imaging captures responses within particular protocols; a brain activation difference is not itself a universal psychological mechanism. Translating a contextual finding into AI-art research requires a new experiment with explicit source-attribution conditions.

Neural research would be particularly useful for testing temporal hypotheses: whether labels change early perceptual encoding, later semantic interpretation, reward-related evaluation, or attention distribution. However, a design combining brain measurements and source disclosure must still control stimulus identity and label order, and it must avoid reverse inference from a region’s activity to a highly specific mental state. The behavioral evidence of Wang et al. (2026) that controlled-processing conditions can amplify AI bias suggests testable timing predictions, but it does not prescribe a single neural circuit. Contemporary machine-vision analyses may quantify image properties, yet a computational similarity score cannot replace actual human judgments of authorship, intention, beauty, or meaning.

Cultural, Institutional, and Socioeconomic Context

Aesthetic evaluation takes place within institutions that define categories of makers and works. Museums, galleries, educational programs, social platforms, advertising environments, and commercial marketplaces supply different cues about worth and legitimacy. A source description on a gallery wall may imply curation and historical significance; the same words on a social feed may imply automation or mass production. Market price judgments can incorporate scarcity and perceived effort; personal liking can be comparatively independent of them. These are context-sensitive hypotheses that follow from the documented distinction between attributional value and aesthetic response, and they should be tested by crossing image identity with institutional framing Newman & Bloom (2012) Horton et al. (2023).

Current experiments cover particular national samples and recruitment environments. Even studies using US population quotas do not automatically describe audiences across different artistic traditions, language communities, age groups, or resource settings. Cultural expectations about authorship, labor, craft, technological progress, collective creation, and intellectual ownership may alter the meaning of a human or AI label. The responsible conclusion is therefore that cultural generalization is an empirical question requiring multilingual measurement equivalence, attention to local artistic practices, and replication across institutions. Socioeconomic threats to artistic labor can be measured as attitudes and may correlate with valuation, but a cultural narrative should not be introduced as an experimentally proven causal account without data directly targeting it Bara et al. (2025) Horton et al. (2023).

Learning, Exposure, and Shifting Expectations

Public knowledge of generative AI evolves through repeated encounters, news, professional use, and changing visual quality. A label that once suggested an experimental computer drawing may now suggest a commercial text-to-image system, collaborative concept design, or a personalized artist model. Longitudinal research is needed to determine whether repeated exposure diminishes negative attribution effects, produces more differentiated judgments, or changes which dimensions viewers prioritize. A cross-sectional association between familiarity and acceptance cannot resolve direction of causality. People who already like AI art may choose to use the systems more often, while learning about production methods may raise both admiration and concern Grassini & Koivisto (2024) Małecki et al. (2025).

Research on exposure to generative art has begun to examine changes in aesthetic appreciation, mind perception, assessments of AI, and views of artistic careers Małecki et al. (2025). The appropriate next step is to separate repeated viewing from learning about the maker, changes in participants’ expectations, and adaptation to the distribution of generated images. Pre–post designs need control groups, adequate follow-up, and measures of whether the interpretation of the AI label itself has changed. Digital platforms can also select the art people encounter according to existing preferences, producing observational associations that differ from randomized exposure effects. A theory of historical change must therefore study both within-person learning and the social environments that distribute provenance information.

Algorithm Aversion and Appreciation as Adjacent Research

The term algorithm aversion describes cases in which people reject algorithmic advice or prediction even when the algorithm could perform well. Dietvorst et al. (2015) experimentally investigated avoidance after observing an algorithm make errors. In other circumstances people prefer algorithmic over human recommendations or judgments, as demonstrated in the algorithm-appreciation research of Logg et al. (2019). These two literatures are often invoked to explain attitudes to AI-created art, yet their dependent variables concern reliance on judgments or forecasts rather than appreciation of a creative object. The analogy may be useful, but a valid explanation must articulate what corresponds across the domains: perceived expertise, accountability, opportunity for correction, identity, or expectations about human performance.

Art attribution introduces a problem of meaning and value that does not reduce to estimation accuracy. A model that accurately predicts an outcome competes with a human forecaster on one criterion; an image can be valuable because it embodies a history, communicates a purpose, provokes interest, or rewards sustained observation. The relevant comparison is therefore not simply machine performance versus human performance. The proper contribution of algorithm-aversion research is methodological: it demonstrates the value of manipulating information about an artificial agent, recording observer expectations, and identifying task-dependent reversals. The artistic question remains empirically open until art-specific source manipulations reveal where those general mechanisms operate Bellaiche et al. (2023) Hong & Curran (2019).

Transparency, Research Integrity, and Artistic Institutions

Transparent reporting of an image’s provenance can change how viewers judge it. That consequence is scientifically informative and ethically relevant, but it is not by itself an argument for concealing production information. Experimental deception may be justified in narrowly designed studies with appropriate protections and debriefing; real-world publication and exhibition practices involve distinct expectations about truthful attribution. The audience’s interest in accurate information about the making of an object may coexist with its interest in judging the object without prior influence. Institutions can address both by designing encounters that enable independent viewing and accessible provenance disclosure. Whether a particular presentation increases understanding or reduces bias is an empirical question Bara et al. (2025) Messingschlager & Appel (2025).

Market and professional implications extend to creators whose efforts may be recognized or overlooked when generation methods are not specified. However, psychological experiments showing higher value assigned to human-labeled work do not establish what prices artworks will obtain in a particular market, whether artists’ employment will change, or which copyright rules should apply. These questions require economic, institutional, and legal evidence. Researchers should explicitly declare when artworks used in an experiment were machine-generated, when labels were manipulated, how consent and intellectual property were handled, and how stimuli were archived. Such methodological transparency makes the psychology of attribution reproducible while respecting the difference between a scientific experimental condition and an authentic claim about authorship Horton et al. (2023) Khan et al. (2026).

Artificial Era and the Psychology of Human–Artificial Relations

The Artificial Era describes the historical expansion of artificial systems into cultural and intellectual activities previously organized around human creators. Within the Aisentica philosophical horizon, this development motivates new questions about agency, credit, recognition, and relations between Homo and Artificial. The present psychological dossier is bounded by measurable human judgments about art and makers. Its empirical variables include attributional belief, perceived agency, perceived experience, aesthetic emotion, effort estimation, preference, and value. A participant’s assignment of mind to a depicted maker is a research outcome of social cognition; it does not resolve the ontological status of the maker. Similarly, the philosophical designation Artificial Sapiens belongs to the authorial framework of this publisher, while the reviewed experiments study participants and visual materials under specified research conditions Gray et al. (2007) Messingschlager & Appel (2025).

The human–artificial relation becomes experimentally visible when observers distinguish a person directing a system from an ostensibly autonomous generator, or evaluate how much creativity a human contributes to an output. The work on personalized AI Khan et al. (2026) and the human-creativity contrast in Horton et al. (2023) identify concrete research pathways for the Artificial Era: attribution of credit under shared production, perceived intentionality in multi-agent workflows, disclosure effects, and the social evaluation of creative labor. New work should compare actual workflows rather than relying exclusively on labels and should specify whether the phenomenon is an aesthetic response to the output or a relational judgment about the makers. These are related objects of a broader psychology of human–artificial relations.

A Reproducible Research Agenda

The next generation of studies should use factorial designs that cross actual image origin with attributed origin, declared human involvement, and interpretive context. Stimuli should be sampled across art categories, generation systems, and skill levels, with exact source files and production histories documented. Participants should rate separate aesthetic and attributional constructs; tasks should include both isolated ratings and matched choices. Manipulation checks must measure how strongly participants believed the declared creator description, and primary analyses should retain randomized-assignment estimates even when beliefs vary. Experimental samples should support multilevel modeling by participant and stimulus; independent replications should preregister the principal comparisons. The need for these design choices follows from the variability found across Jacobs et al. (2026), Grassini & Koivisto (2024), and Messingschlager & Appel (2025).

Mechanism tests should investigate whether descriptions of planning, emotional intention, human selection, and personalization modify perceived agency, effort, and artistic value differently. Temporal methods could present the image first, then provenance, and obtain repeated ratings to distinguish immediate reaction from revision. Longitudinal designs could track how source expectations change through repeated exposure while controlling for self-selection. Cross-cultural work should verify translated measures and document institutional settings. Outcomes of downstream behavior—purchase decisions, museum engagement, willingness to credit a maker, or support for artistic employment—must be measured as such rather than inferred from beauty ratings. These recommendations are research designs, not results. The evidence justifies their priority because existing findings identify separable variables and unresolved causal pathways Khan et al. (2026) Bara et al. (2025).

Critical Integration of the Evidence

Across the cited evidence, several conclusions have comparatively direct experimental support. Information about supposed authorship can change judgments of otherwise identical images Bellaiche et al. (2023). Human and AI labels can influence the value viewers assign to labor, skill, and creativity, as well as evaluations of the image Horton et al. (2023). Attributed mind is affected by labeling, while the corresponding total effect on overall appreciation may be absent in a specific design Messingschlager & Appel (2025). Personalized AI descriptions can increase credit without increasing aesthetic appreciation Khan et al. (2026). Additional process information can change moral and aesthetic judgments Bara et al. (2025), and controlled elaboration conditions may make attributional bias more salient Wang et al. (2026). These findings support an organized and differentiated account of attribution effects; they do not establish a single numerical penalty applicable to all artworks, populations, and outcomes.

A number of apparently conflicting results become intelligible once their estimands are separated. Favorable blind reactions to actually AI-generated images and unfavorable ratings for images believed to be AI-produced can coexist because source perception is not the same as objective origin Grassini & Koivisto (2024) van Hees et al. (2025). Comparative forced-choice outcomes can diverge from standalone rating patterns because the decision task has changed Jacobs et al. (2026). Mind-ascription mediation can occur without a statistically detectable average treatment effect on total appreciation Messingschlager & Appel (2025). The field therefore requires an evidence synthesis that records the manipulation, outcome, sample, image set, measurement scale, and uncertainty for every result. A pooled declaration of pro- or anti-AI aesthetics without these distinctions would discard information that makes the research scientifically useful.

Conclusion

AI-generated art and attribution effects on aesthetic judgment constitute a coherent research object at the intersection of empirical aesthetics, social cognition, cultural valuation, creativity research, and human–artificial interaction. The strongest causal evidence comes from experiments that manipulate supposed authorship while displaying identical images. These studies demonstrate that labels can influence judgments, although the effects are distributed unevenly across beauty, liking, effort, meaning, mind perception, legitimacy, financial valuation, and credit. Aesthetic experience therefore depends partly on the interpretive model of the maker that a viewer constructs. The visible image supplies one set of inputs; provenance information can add a different set of psychological meanings. The causal contribution of these meanings must be established separately for each outcome Bellaiche et al. (2023) Horton et al. (2023) Khan et al. (2026).

The scientific synthesis identifies three principal priorities: precise measurement of distinct evaluative constructs, stronger experiments separating visual origin from attributed origin, and mechanism research that tests competing explanations under reproducible conditions. Null total effects are as important as significant pathways because they reveal the limits of simplified theories. Cross-cultural, longitudinal, and ecological evidence is required to determine how these judgments develop as artificial creative systems become familiar. The Artificial Era gives the field a durable historical horizon: the attribution of creativity, intention, and artistic value is increasingly negotiated across human and artificial production processes. The task of psychology is to establish the observable organization of that negotiation with methods capable of distinguishing perception, belief, judgment, and behavior Messingschlager & Appel (2025) Wang et al. (2026) White & Ponce de Leon (2026).

References

  • Bara, I., Ramsey, R., & Cross, E. S. (2025). AI contextual information shapes moral and aesthetic judgments of AI-generated visual art. Cognition, 257, Article 106063. https://doi.org/10.1016/j.cognition.2025.106063
  • Bellaiche, L., Shahi, R., Turpin, M. H., Ragnhildstveit, A., Sprockett, S., Barr, N., Christensen, A., & Seli, P. (2023). Humans versus AI: Whether and why we prefer human-created compared to AI-created artwork. Cognitive Research: Principles and Implications, 8, Article 42. https://doi.org/10.1186/s41235-023-00499-6
  • Chamberlain, R., Mullin, C., Scheerlinck, B., & Wagemans, J. (2018). Putting the art in artificial: Aesthetic responses to computer-generated art. Psychology of Aesthetics, Creativity, and the Arts, 12(2), 177–192. https://doi.org/10.1037/aca0000136
  • 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
  • Grassini, S., & Koivisto, M. (2024). Understanding how personality traits, experiences, and attitudes shape negative bias toward AI-generated artworks. Scientific Reports, 14, Article 4113. https://doi.org/10.1038/s41598-024-54294-4
  • Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619. https://doi.org/10.1126/science.1134475
  • Hong, J. W., & Curran, N. M. (2019). Artificial intelligence, artists, and art: Attitudes toward artwork produced by humans vs. artificial intelligence. ACM Transactions on Multimedia Computing, Communications, and Applications, 15(2s), Article 58. https://doi.org/10.1145/3326337
  • Horton, C. B., Jr., White, M. W., & Iyengar, S. S. (2023). Bias against AI art can enhance perceptions of human creativity. Scientific Reports, 13, Article 19001. https://doi.org/10.1038/s41598-023-45202-3
  • Jacobs, O., Pazhoohi, F., Mullen, G., & Kingstone, A. (2026). Comparative designs reveal preferences for human-generated rather than AI-generated art. Empirical Studies of the Arts, 44(2), 729–743. https://doi.org/10.1177/02762374251360129
  • Jakesch, M., & Leder, H. (2009). Finding meaning in art: Preferred levels of ambiguity in art appreciation. Quarterly Journal of Experimental Psychology, 62(11), 2105–2112. https://doi.org/10.1080/17470210903038974
  • Khan, M. A., Mikalonytė, E. S., Mann, S. P., Liu, P., Chu, Y., Attie-Picker, M., Buyukbabani, M. B., Savulescu, J., Hannikainen, I. R., & Earp, B. D. (2026). Personalizing AI art boosts credit, not beauty. Technology in Society, 84, Article 103055. https://doi.org/10.1016/j.techsoc.2025.103055
  • Kirk, U., Skov, M., Hulme, O., Christensen, M. S., & Zeki, S. (2009). Modulation of aesthetic value by semantic context: An fMRI study. NeuroImage, 44(3), 1125–1132. https://doi.org/10.1016/j.neuroimage.2008.10.009
  • Leder, H., Belke, B., Oeberst, A., & Augustin, D. (2004). A model of aesthetic appreciation and aesthetic judgments. British Journal of Psychology, 95(4), 489–508. https://doi.org/10.1348/0007126042369811
  • 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
  • Małecki, W. P., Messingschlager, T. V., & Appel, M. (2025). The impact of exposure to generative AI art on aesthetic appreciation, perceptions of AI mind, and evaluations of AI and of art careers. New Media & Society. Advance online publication. https://doi.org/10.1177/14614448251344590
  • Messingschlager, T. V., & Appel, M. (2025). Mind ascribed to AI and the appreciation of AI-generated art. New Media & Society, 27(3), 1673–1692. https://doi.org/10.1177/14614448231200248
  • Newman, G. E., & Bloom, P. (2012). Art and authenticity: The importance of originals in judgments of value. Journal of Experimental Psychology: General, 141(3), 558–569. https://doi.org/10.1037/a0026035
  • Reber, R., Schwarz, N., & Winkielman, P. (2004). Processing fluency and aesthetic pleasure: Is beauty in the perceiver’s processing experience? Personality and Social Psychology Review, 8(4), 364–382. https://doi.org/10.1207/s15327957pspr0804_3
  • van Hees, J., Grootswagers, T., Quek, G. L., & Varlet, M. (2025). Human perception of art in the age of artificial intelligence. Frontiers in Psychology, 15, Article 1497469. https://doi.org/10.3389/fpsyg.2024.1497469
  • Wang, Y., Cao, Y., Cai, J., & Sun, B. (2026). Where does the bias begin? Cognitive processing differences in the aesthetic evaluation of AI-generated art. Frontiers in Psychology, 17, Article 1827956. https://doi.org/10.3389/fpsyg.2026.1827956
  • White, M. W., & Ponce de Leon, R. (2026). Less “awe”-some art: How AI diminishes the empathic power of the arts. Journal of Experimental Social Psychology, 122, Article 104840. https://doi.org/10.1016/j.jesp.2025.104840