The theoretical scaffolding of the study combines three well-established psychological frameworks. Social comparison theory, originating with Leon Festinger in 1954, holds that people evaluate themselves against others and that comparisons with superior targets tend to depress self-evaluation. Objectification theory adds that appearance-saturated environments encourage people to adopt an observer’s perspective on their own bodies, producing shame and appearance anxiety. The tripartite influence model identifies media, peers and family as the channels through which appearance ideals are internalised, and digital platforms compress all three into a single, continuously refreshed interface. What the researchers argue is that AI-mediated imagery is not merely another instance of media exposure but a theoretically distinct case, for three reasons: the comparison set is algorithmically curated and personalised, generative models produce bodies with no real referent, and users tend to grant machine judgements a heuristic objectivity that human judgements do not receive.

That last point deserves emphasis. A retouched magazine photograph can be dismissed as an editorial choice, but an algorithmically assembled feed may be read as an impartial description of which bodies are valued. The authors identify three candidate mechanisms through which this distinctness could operate: comparison-set curation, in which personalised ranking increases the frequency and homogeneity of upward appearance comparisons; automated self-alteration, in which filters return an optimised version of the user’s own body and generate an intrapersonal comparison between the biographical and the computationally corrected self; and attributed algorithmic evaluation, in which the mere belief that a system is ranking one’s appearance generates anxiety and perceived exclusion. The study measures perceptions relevant to the second and third mechanisms; observing the first would require platform-level data on what each participant was actually shown, which no survey can supply.

The empirical backdrop is a body of audit research showing that computer-vision and text-to-image systems skew towards thin and athletic phenotypes. Studies cited by the authors report that generative models produce thin or athletic figures even in response to appearance-neutral prompts, that images of thin bodies receive higher levels of algorithmic promotion, and that moderation systems have been reported to conflate higher-weight bodies with sexualised content, reducing the reach of fat activist accounts. Critical scholars describe this as an emergent property of systems optimised for visual uniformity and user retention rather than a deliberate design goal. The present study deliberately does not test any of these claims about system behaviour; its contribution lies at the perceptual level, documenting how users interpret environments that the audit literature describes.

Methodologically, the research is a quantitative, cross-sectional correlational survey administered through Microsoft Forms between October 2025 and February 2026. The 20-item instrument was piloted with 43 participants and content-validated by five senior academics, yielding an overall Cronbach’s alpha of 0.80. The final sample of 284 respondents comprised 198 women, 82 men and 4 non-binary participants, with 93 percent residing in Ecuador. Sampling was non-probability and self-selected, so no claim of representativeness is made. Because the variables were ordinal or categorical and departed significantly from normality on every test, the analysis relied on non-parametric procedures: Spearman’s rho for bivariate associations, Mann-Whitney U tests for two-group comparisons, Kruskal-Wallis tests with Dunn post hoc comparisons for multiple groups, and exploratory ordinal logistic regressions to check whether the main associations survived adjustment for gender, age and education.

The headline findings are numerically modest but theoretically pointed. Perceived AI-based photographic enhancement was positively associated with self-reported feelings of body-related exclusion (rho = 0.24, p < .001), and comparison of one’s own body with AI-generated representations showed the strongest association in the study, with perceived automated enhancement (rho = 0.32, p < .001). Notably, this association held both for participants who reported feeling better after comparing themselves with a generated image and for those who felt worse, suggesting that the variable indexes engagement with AI-mediated imagery rather than the affective outcome of that engagement. Participants who compared themselves with AI-generated images, and those reporting greater exclusion, were also the most likely to identify these systems as generating predominantly thin or athletic bodies, reflected in negative correlations (rho = -0.20 and rho = -0.17) on a scale whose lowest code corresponds to that response.

The sociodemographic results add nuance. Women reported more frequent use of appearance-modifying filters than men (74.4 percent of men never used them, against 53.0 percent of women) and were more likely to agree that AI promotes thin bodies as the aesthetic norm, though both effects were small, with r values of .20 and .16. Educational attainment showed a clearer gradient: perceived aesthetic bias rose across educational levels (H(3) = 17.70, p < .001), and support for training AI systems with diverse bodies climbed monotonically from 36.6 percent among those with complete secondary education to 72.3 percent among the postgraduate group. Postgraduate respondents were also the most likely to identify AI output as thin or athletic, with 57.8 percent doing so. An apparent association with age, however, dissolved once education was held constant, indicating that the age pattern was an artefact of the strong correlation between age and educational attainment in this sample.

Perhaps the most intriguing result concerns uncertainty. Reported exclusion rose monotonically across the three response categories of the perceived-enhancement item, from a mean rank of 131.44 among those who denied that AI enhances photographs, through 145.78 among the unsure, to 169.04 among those who affirmed it. Uncertainty, in other words, occupied an intermediate rather than a neutral position. The authors offer a post hoc interpretation consistent with work on the algorithmic imaginary: not knowing whether one’s image is being algorithmically adjusted may itself be a condition of appearance-related unease, because it removes the possibility of attributing a perceived discrepancy either to oneself or to the system. They flag this as a hypothesis for future designs that measure algorithmic uncertainty directly, not as a demonstrated effect.

The authors are careful about what the data cannot show. All variables were measured at a single point in time, so the central association between perceived enhancement and exclusion admits at least three readings: perceiving AI as interventionist may heighten sensitivity to bodily standards, feeling excluded may increase attention to and detection of such intervention, or both may follow from an unmeasured third variable, most plausibly prior body dissatisfaction. The adjusted ordinal models showed that the two principal associations survived control for gender and education, with perceived enhancement returning an odds ratio of 1.47 and body comparison an odds ratio of 2.30 for reporting exclusion, but the models explained only a modest share of variance, with McFadden pseudo-R-squared values between 0.026 and 0.063. The authors also note an asymmetry with real consequences: higher education predicted greater recognition of aesthetic bias, yet nothing in the data suggests that those who recognised it reported less exclusion. Recognition and protection, they write, appear to be separable, a testable proposition rather than a demonstrated one.

The study closes with a balanced assessment of the technology itself. Generative systems can be prompted or fine-tuned to represent body types that conventional media pipelines have excluded, at marginal cost far below that of commissioning photography; the same ranking machinery that concentrates attention on normative appearance can, under different objective functions, surface body-diverse content; and the auditability of automated systems is a genuine asset, since bias in a model can be measured and contested in ways that bias in a magazine picture desk could not. But the authors caution that the evidence for current skew is far stronger than the evidence for redirection. What their data establish is covariation among perceptions, comparisons and feelings within one self-selected sample, not causal influence, and not a verdict on what AI systems actually do. The most informative next step, they argue, would combine algorithm audits with perceptual surveys, validated multi-item instruments and multi-wave designs, to determine whether users’ perceptions track real system behaviour and to identify the conditions under which the two diverge. Until then, the study stands as a first quantitative sketch of how a Latin American sample of users experiences the quiet aesthetic authority of the algorithm.

Subject of Research: Perceived algorithmic aesthetic bias, body comparison with AI-generated images, and weight-based digital exclusion among social media users

Article Title: Artificial intelligence, body image comparison, and perceived aesthetic bias: a correlational study of weight-based digital exclusion

Article References: Torres-Toukoumidis, A., León-Alberca, T., & Lituma, J. P. C. (2026). Artificial intelligence, body image comparison, and perceived aesthetic bias: a correlational study of weight-based digital exclusion. Current Psychology, 45(19), Article 1576. https://doi.org/10.1007/s12144-026-10121-9

Image Credits: AI Generated

DOI: 10.1007/s12144-026-10121-9

Keywords: artificial intelligence, algorithmic bias, body image, social comparison, weight stigma, digital exclusion, social media, objectification theory, generative models, Ecuador, cross-sectional survey, algorithmic imaginary