AI-generated food images make some people uneasy—here’s why

By Priyanjana Pramanik, MSc.Reviewed by Susha Cheriyedath, M.Sc.Feb 24 2025

Imperfect AI-generated food images are unsettling, falling into the uncanny valley—where realism meets discomfort. Could food neophobia be driving our aversion to these digital dishes?

Study: Eerie edibles: Realism and food neophobia predict an uncanny valley in AI-generated food images. Image Credit: Shutterstock AI

In a study soon to be published in the journal Appetite, researchers explored what it is about images of food generated by artificial intelligence (AI) that evoke feelings of uncanniness in the people who see them.

Their findings suggest that viewers find AI-generated images less pleasant and uncannier than both realistic and unrealistic (cartoonish or abstract) images. Further, perceived pleasantness and affinity initially declined before increasing in a cubic rather than a simple U-shaped curve. The research team noted that these feelings could be driven by food neophobia rather than food disgust, which causes people to avoid novel foods due to safety concerns. The study’s results have implications for clinical research, marketing, and advertising using AI-generated food images.

Background

Text-to-image programs can generate pictures based on detailed prompts, and artificially produced images with varying levels of realism are used in numerous applications. AI-generated food images have significantly replaced food photography in marketing and advertising, but consumers' responses have been mixed. AI foods can appear off-putting, strange, or wrong, and viewers often express their preference for images of real food.

These feelings could be due to an effect first described by Masahiro Mori in his seminal 1970 essay, in which he discussed how people feel unnerved by humanoid robots or realistic prosthetics. Mori noted that as an entity becomes more realistic-looking, we initially experience more affinity towards it, but eventually, it comes across as unnatural or eerie; this decline in the viewer’s perceived affinity enters the uncanny valley.

While the uncanny valley phenomenon has not been studied for food, it has been noted for buildings, other-than-human species, and various inanimate objects. The researchers in this study sought to determine if AI-generated food images would elicit a similar effect, particularly when they contained slight distortions or deviations from expected realism. People could show aversion to AI-generated food because the image evokes disgust in the same way that moldy, contaminated, or spoiled foods do, preventing them from ingesting harmful substances. However, the study found that food disgust sensitivity did not significantly predict perceptions of uncanniness—rather, food neophobia played a key role.

These are the disease avoidance drivers of uncanniness, and humans may have evolved to discriminate subtle differences in the appearance of food to recognize that they could be a threat to their survival; however, this discrimination could also lead to false positives, where edible food is discarded due to visual imperfections or AI-generated images are perceived negatively. The study also explored whether other mechanisms, such as violations of internalized food norms or specialized visual processing, could contribute to the negative perceptions of AI food.

About the Study

The research team explored whether the uncanny valley effect applies to AI-generated food images and tried to identify the individual differences or mechanisms that contribute to negative perceptions.

They hypothesized that the uncanniness of images would be explained by a cubic (rather than simple quadratic or linear) relationship and that people would see unrealistic or real foods as less uncanny than AI-generated imagery.

To explore the underlying mechanisms, researchers hypothesized that individual differences in food neophobia, rather than food disgust, could predict food uncanniness. They also tested whether images of rotten or moldy food show a similar pattern of uncanniness to AI-generated food but for potentially different reasons.

A pilot study involving 12 people at a German university was used to select AI-generated images from a set of 99, grouped as realistic, imperfect (for example, distorted), and unrealistic (for example, cartoonish, abstract, stylistic). They also rated the images out of 100 on six scales: lifelike to artistic, photorealistic to abstract, pleasant to repulsive, ugly to pretty, uncanny to plain, and boring to eerie.

Based on the results of the pilot study, the research team recruited 95 participants from online survey platforms and asked them to rate 38 images based on their perceived eeriness, warmth, and realism. The images included six with high levels of realism, six with low levels, 20 with intermediate levels, and six images of rotten food. They also filled out questionnaires to assess their food neophobia (but not food disgust, which was later found to be insignificant), each on a scale of 1 to 100.

The data were analyzed using linear mixed-effects models with cubic realism terms as fixed effects and participant-level random effects to predict uncanniness and test the initial hypotheses. A post-hoc analysis also examined whether body mass index (BMI) influenced responses to AI-generated food.

Findings

The participants comprised 95 German speakers (including 76 who identified as White) who were, on average, 31.28 years old; 27 identified as female, 64 as male, and the remaining 4 as other.

Correlational analysis indicated a slight positive correlation between realism and uncanniness, a strong negative correlation between pleasantness and uncanniness, and a slight positive correlation between realism and pleasantness. In the linear models, the cubic term of realism performed the best as a predictor of uncanniness. However, the quadratic term of realism performed the best as a predictor of pleasantness.

The statistical comparisons showed that participants considered imperfect AI-generated images to be significantly less pleasant and more uncanny than both realistic and unrealistic AI-generated images. Crucially, the unrealistic (cartoonish or highly stylized) AI-generated images were not perceived as particularly uncanny—only the imperfect, distorted ones were.

Food disgust did not emerge as a significant moderator of the effect when considering individual drivers of uncanniness, but food neophobia did. The supplementary analysis found that participant BMI also appeared to moderate the impact, suggesting that individuals with higher BMIs may view AI-generated food images more positively.

Conclusions

While the study did not control for how familiar participants were with AI-generated images or for factors such as food savoriness and dietary preferences, the findings indicate that imperfect AI food images are subject to the uncanny valley effect, while highly unrealistic images are not.

Imperfect but realistic images were seen as less pleasant and more eerie, and the impact of realism on uncanniness and pleasantness was primarily moderated by aversion to novel foods (food neophobia), not food disgust.

The findings have implications for AI-generated food in marketing, advertising, and clinical research. Because BMI appeared to influence reactions to AI-generated food, future studies could explore how individuals with obesity or eating disorders perceive AI food, particularly in relation to calorie estimation and restrictive eating behaviors.

Cultural influences and exposure to AI-generated images should also be considered in future studies, as increased familiarity with AI imagery could potentially reduce the uncanny valley effect over time.