Affective Science has its stock characters: the snarling dog, the smiling baby, the wrecked car, the plate of delicious food. For decades, these images have done quiet work in laboratories, helping researchers elicit emotions on demand. A participant sits down in front of a screen, a picture appears, and the mind and body respond with a little fear, a little disgust, a little joy. The problem is that many of these images have started to show their age. Some of the most widely used sets of affective stimuli were developed in a very different media era, and sometimes it shows immediately: in the colours, the clothing, the background technology.
A stimulus meant to evoke fear may also evoke a trip back to the 1990s. There is also a deeper issue: emotional images do not work equally well everywhere. A scene that reads as comforting in one culture may look strange or emotionally flat in another. Even when an image is understandable, its emotional meaning may be far less universal than psychologists once assumed.
This is part of a broader problem in psychology. For years, researchers have pointed to the field’s reliance on “WEIRD” samples and materials—Western, Educated, Industrialised, Rich, and Democratic. The tools of emotion research are no exception. A picture designed to evoke appetite, comfort, or shame in one setting may miss the mark entirely in another.
Emotion science needed an update. Now an international team led by Maciej Behnke has proposed one. Using generative artificial intelligence, the researchers developed the Library of AI-Generated Affective Images (LAI-GAI), a project designed to modernise affective stimulus sets and make them more adaptable across cultures, age groups, and genders.
From old images to a new library
The idea behind the project, published in Advances in Methods and Practices in Psychological Science, was straightforward: could the same technology used to generate eye-catching digital art also be used to build scientifically useful emotional stimuli?
To answer that question, the researchers used a human-in-the-loop approach, meaning that AI did not operate on its own. Instead of simply asking a model to “make something sad,” the team began with existing validated images from older stimulus databases. Large language models such as ChatGPT were used to generate detailed descriptions of those scenes. Those descriptions were then fed into image generators such as Midjourney and Freepik to create updated, high-resolution versions. Finally, the outputs were manually reviewed and revised to remove artifacts and improve accuracy.
The result was a set of 847 images covering 12 discrete emotions, from nurturant love and awe to disgust and anger. But the key innovation was not just visual quality. It was flexibility.
In a validation study involving 2,470 participants from 58 countries, the team found that AI-generated images were, on average, at least as effective as traditional image sets in eliciting the intended emotional responses. That finding alone is notable: it suggests that carefully developed AI-generated stimuli can function as legitimate tools for psychological research rather than as a novelty.
Why cultural matching matters
The more important advance, however, may be that these images can be adapted. One longstanding weakness of traditional affective image sets is that they assume emotional meaning travels easily across cultural contexts. In practice, it often does not.
To address this, the LAI-GAI team created culturally matched image sets for six broad cultural contexts: African, Arab, Asian, Indian, Latin American, and European/North American. This required much more than automatic editing. Left on its own, generative AI often reproduced stereotypes or culturally inaccurate details. In some cases, for example, prompts related to African contexts defaulted toward imagery of poverty. To counter this, the researchers worked with cultural experts who reviewed the outputs and rejected or revised images that felt inauthentic. Roughly one-third of the initial images required this kind of intervention.
The results supported the effort. Culturally matched images were somewhat better at hitting their intended emotional target than unmatched ones. The effect was not enormous, but it was meaningful: emotional stimuli appear to work better when the depicted world feels familiar and socially plausible to the people viewing it.
The researchers also extended this logic beyond culture. They created versions of the same emotional scenes for children, adults, and older adults, as well as for female and male targets. These variations preserved emotional effectiveness while giving researchers much greater control over who is represented in a stimulus.
A more flexible future for emotion research
For psychologists, this could change the practical logic of designing studies. Instead of building experiments around whatever fixed image sets happen to exist, researchers may increasingly be able to build stimuli around their actual research questions. They could create more up-to-date materials, tailor them to the populations they study, and reduce their reliance on narrow, aging databases.
This does not mean the problem is solved. Generative AI can amplify stereotypes, produce subtle visual errors, and create a false impression of cultural authenticity if left unchecked. The LAI-GAI project is valuable partly because it makes those limitations visible. Its contribution is not just a library of images, but a transparent pipeline showing how AI can be used responsibly in psychological science.
That may be the most important point. The project does not suggest that machines can replace human judgment in the study of emotion. Instead, it shows that AI can become part of a rigorous scientific workflow when it is paired with careful validation, expert review, and cultural oversight.
In that sense, the real contribution of LAI-GAI is methodological. It offers a practical way to update one of the most basic tools in affective science: the emotional image. And in doing so, it suggests that the future of psychological research may be not only more technologically advanced, but also more flexible, more representative, and better aligned with the people it aims to understand.
The entire article can be found here.
We also invite you to visit the project website: https://www.affectdatabases.amu.edu.pl/
photo: Władysław Gardasz
