idw – Informationsdienst Wissenschaft
Nachrichten, Termine, Experten
Nachrichten, Termine, Experten
idw - Informationsdienst
Wissenschaft
CISPA researcher Tejumade Àfọ̀njá co-authored a new international study that uses food as a starting point to reveal significant cultural blind spots in today’s AI systems. The study also introduces a new participatory research approach to create more inclusive datasets and evaluate biases in AI models. The paper “The World Wide Recipe: A Community-Centred Framework for Fine-Grained Data Collection and Regional Bias Operationalisation” was presented at the ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25) in Athens in June 2025 and won the Best Paper Honorable Mention award.
“Food is an important gateway to culture,” explains CISPA researcher Tẹjúmádé Àfọ̀njá, a PhD student on the team of CISPA-Faculty Dr. Mario Fritz. “We wanted to explore how generative AI represents people’s food cultures in generated images.” Behind it was the desire to explore possible cultural biases in AI models. “The lead project coordinator of our paper, Siobhan Mackenzie Hall, had found in previous studies that many models are biased in one way or another,” Àfọ̀njá continues. “We asked ourselves which lens we could use to look at this problem. Food turned out to be a good one, because it’s a universal language.” Specifically, the team investigated how certain dishes are depicted in AI-generated images. To do this, a new reference dataset was developed in a first step, which was then used to test existing models in a second step.
A New Dataset Featuring Dishes from Around the World
The author team chose a community-oriented research approach and called it the World Wide Recipe. People from all over the world were invited to contribute their knowledge. “We wanted to give people a say in how their cultures are represented in AI systems,” says Àfọ̀njá. As a first case study, they built the World Wide Dishes (WWD) dataset. It is a collection of 765 dishes from 106 countries, described in 131 local languages. Each entry was contributed directly by community participants all over the world, who shared the cultural, linguistic, and culinary context behind each dish and provided images. “We compared WWD with existing datasets whose data had been collected from the internet,” explains Àfọ̀njá. “More than half of the dishes in the dataset don’t appear there, which gives it its unique character.” The dataset and all code have been released under an open license to encourage transparency and collaboration.
Misrepresentation in Existing Models
In a second step, Àfọ̀njá and her colleagues used WWD to compare the images of the dishes included in it with AI-generated images of those dishes. The comparative analysis was again carried out by members of the communities. “We found that many of the models were stereotypical in their outputs. For example, when we prompted the models to generate an image of a Nigerian dish like Amala, the results were often unappealing or inaccurate,” explains Àfọ̀njá. “In contrast, when we asked for something like a hot dog from the United States, the generated images were much closer to the real thing.” This applied to all models tested: DALL·E 2, DALL·E 3, and Stable Diffusion. “The image quality was generally poor, and there were clear misrepresentations of the culture,” she continues. “The reason is that many models are trained on internet data, and if foods from certain regions aren’t represented online, those regions will be overlooked.”
A Global Tool Requires Global Input
Àfọ̀njá and her colleagues conclude from this finding that the companies behind the models need to invest more in long-tail training and data collection for large language models. “Our argument is that the companies must prioritize all regions if they want to build models that truly represent global culture,” says Àfọ̀njá. “It’s not enough to design a model in Silicon Valley or Germany and expect it to work everywhere. Collecting more data is key—but it has to be done in collaboration with communities, not just extracting data from them.” An important keyword here is ownership of the data. “When you collect data from communities, the question is always: Who owns it—the community or the organization that funded the data collection?” says the CISPA researcher.
Data Collection and the Fight Against Cultural Bias
Àfọ̀njá would love to scale World Wide Dishes, but it’s very expensive. Until now, the whole project was entirely volunteer-driven. “None of the contributors were paid,” she explains. “With proper funding, we could pay community contributors to collect even more local data—asking families for recipes that aren’t online, for instance. That kind of data is invaluable but costly to obtain.” Because the method of data collection was so important for the project, another paper was produced as a follow-up product. “We published a paper called ‘The Human Labour of Data Work,’ which documents how we collected the dataset and the challenges involved. It focuses on the human effort, cultural trust, and lessons for anyone building similar datasets in the future.” Anyone who listens to Àfọ̀njá knows that this issue is close to her heart and that she will continue to campaign for AI models to lose their cultural bias.
Magomere, Jabez; Ishida, Shu; Afonja, Tejumade; Salama, Aya; Kochin, Daniel; Foutse, Yuehgoh; Hamzaoui, Imane; Sefala, Raesetje; Alaagib, Aisha; Dalal, Samantha; Marchegiani, Beatrice; Semenova, Elizaveta; Crais, Lauren; Mackenzie Hall, Siobhan (2025): The World Wide Recipe: A Community-Centred Framework for Fine-Grained Data Collection and Regional Bias Operationalisation, In: FAccT ’25, 23-26 June, Athens, Greece, Conference: ACM Conference on Fairness, Accountability, and Transparency
https://github.com/oxai/world-wide-dishes
Visualization to the paper "The World Wide Recipe: A Community-Centred Framework for Fine-Grained Da ...
Merkmale dieser Pressemitteilung:
Journalisten
Informationstechnik
überregional
Forschungsergebnisse
Englisch
Sie können Suchbegriffe mit und, oder und / oder nicht verknüpfen, z. B. Philo nicht logie.
Verknüpfungen können Sie mit Klammern voneinander trennen, z. B. (Philo nicht logie) oder (Psycho und logie).
Zusammenhängende Worte werden als Wortgruppe gesucht, wenn Sie sie in Anführungsstriche setzen, z. B. „Bundesrepublik Deutschland“.
Die Erweiterte Suche können Sie auch nutzen, ohne Suchbegriffe einzugeben. Sie orientiert sich dann an den Kriterien, die Sie ausgewählt haben (z. B. nach dem Land oder dem Sachgebiet).
Haben Sie in einer Kategorie kein Kriterium ausgewählt, wird die gesamte Kategorie durchsucht (z.B. alle Sachgebiete oder alle Länder).
