A new study has revealed how artificial intelligence can both reflect and help uncover society’s hidden prejudices against people with obesity. Researchers have used AI-generated human faces to better measure unconscious weight bias, highlighting how negative stereotypes remain deeply embedded in the way people perceive appearance.
The research addresses a critical flaw in current psychological testing, where tools such as the Implicit Association Test (IAT) rely on low-quality or unrealistic images that fail to reflect the diversity of the real world. Existing image sets used to assess implicit bias are often limited in ethnicity, age, and realism, particularly when representing individuals with higher body weight. This has raised concerns that results from these tests may be skewed or unreliable.
In response, researchers created a new library of 48 AI-generated portraits featuring people of different ethnicities, ages, and genders, shown at either average or higher body weight. These digital images were designed to appear as realistic as possible, and they were tested on a group of 210 adult participants who were asked to rate the faces based on various traits, such as competence, friendliness, and attractiveness.
The findings were stark. Faces perceived as overweight were consistently rated lower in terms of attractiveness and competence. This suggests that implicit weight bias remains widespread, even among people who may not be aware of holding such views. Importantly, these biases were present despite the controlled and standardised nature of the AI-generated images, showing that it was not clothing or facial expressions driving the negative impressions but weight alone.
Matilde Tassinari, PhD, a postdoctoral researcher from the University of Helsinki, noted one of the more surprising patterns to emerge. “One of the clearest findings was that participants rated overweight faces as less realistic than average-weight faces, even though both were generated using the same AI model, parameters, and standardisation. This discrepancy likely reflects not a flaw in the images, but a lack of cultural familiarity with high-quality, neutral representations of people with obesity.”
Many participants were unable to tell whether an image was AI-generated or a real photo, yet heavier faces were more likely to be seen as fake. This suggests that societal expectations around body size may be shaping how authentic or trustworthy a face appears.
Tassinari also pointed out potential limitations within the technology itself. “We must also consider that AI models themselves may be less capable of generating realistic images of people with obesity. This stems from the fact that training datasets may underrepresent larger body types or include them in biased or stereotypical ways. In this sense, the reduced realism ratings could be a signal not just of participant bias, but of bias within the AI itself.”
By using AI to control variables such as lighting, facial positioning, and background, the researchers isolated weight as the main factor under scrutiny. This level of consistency is difficult to achieve using traditional photography, which often introduces confounding differences in posture, clothing, or image quality.
“This dual aspect of bias has broader social implications,” Tassinari added, “suggesting that increased exposure to diverse and realistic representations of body types could help recalibrate perceptions.”
The study’s lead author emphasised the importance of using validated and realistic stimuli when studying unconscious bias. Without accurate representations of people with obesity, previous research may have underestimated the scale or nature of the problem. By improving the tools used to measure bias, scientists hope to better understand how prejudice forms and how it can be reduced.
Tassinari underscored the inclusive potential of the technology. “This research also demonstrates how generative AI can be harnessed for inclusion when thoughtfully applied. In this context, AI helped us overcome practical challenges in creating diverse and high-quality images of individuals with obesity, something that has historically been limited or misrepresented in existing datasets.”
The implications are particularly relevant in clinical and educational settings. Past studies have shown that healthcare professionals and students can hold implicit biases against people with obesity, potentially affecting the quality of care and interactions.
Tassinari highlighted the broader utility of the dataset. “While designed for use in Implicit Association Tests, the stimuli set can also be used more widely in research. While validated sets of faces for psychological research are present, they often do not systematically represent people with obesity, hindering the opportunity to reliably study weight stigma. To fill this gap, all stimuli and their ratings have been made freely available for researchers and practitioners to use.”
