The project

Trust, by the senses

What it takes for people to follow an AI’s advice in the kitchen, and whether it helps them waste less food.

The problem

Households account for around 54% of the roughly 59 million tonnes of food wasted each year in the EU (Eurostat). Most people say they feel responsible for this, yet everyday habits change little: time is short, the fridge is half empty, and it is hard to think of something to cook with what is left.

AI assistants can suggest recipes from leftovers in seconds. The obstacle is trust. People accept AI for technical tasks but are more sceptical when it comes to food. Earlier work from our group found that recipes are trusted less when presented as AI-generated, especially when they are unusual (Califano et al., 2024), and that AI food images lose their appeal once people learn they were made by AI (Califano & Spence, 2024).

The idea: crossmodal nudging

Gastrophysics has shown that flavour is built by all the senses together. In a classic study at the Crossmodal Research Laboratory, making the sound of a crisp louder made people rate it as crisper and fresher (Zampini & Spence, 2004), work that later earned an Ig Nobel Prize. A recipe suggestion can be correct and still feel wrong if it breaks these implicit expectations.

A crossmodal nudge builds such cues into what an AI assistant shows: the words it uses, the image of the dish, the sound of it cooking. It does not argue or persuade. It makes the suggestion feel closer to real food. GASTRO-X tests whether this works, and for whom.

Four courses

I.Benchmark2026–27

AI models against home cooks

Focus groups in Oxford will build 20 realistic “fridge-and-pantry” scenarios, five for each of four European culinary regions. Home cooks across Europe and several leading AI models will then propose meals for the same scenarios, and we will compare how safe the recipes are and how much food they save.

II.The senses2027

Teaching AI what food should feel like

A Gastrophysics Prompt Handbook will turn findings on crossmodal correspondences into instructions an AI model can follow. The model will produce recipe bundles with a description, an image and a short sound, with and without the handbook, and multisensory researchers will rate them blind.

III.The nudge2027–28

Does sensory realism build trust?

An online experiment in four countries will show people an AI recipe as plain text, as text with an image, or as the full sensory bundle, with or without disclosure that it came from AI. We will measure trust and willingness to try the recipe.

IV.The demo2028

An open kitchen assistant

The findings will go into an open-source web prototype, released on GitHub and available to try on this site.

Open science

All studies will be preregistered and approved by the University of Oxford’s research ethics committee before data collection. Data, materials and code will be shared in open repositories following the FAIR principles, and publications will be open access.

References

Califano, G., & Spence, C. (2024). Assessing the visual appeal of real/AI-generated food images. Food Quality and Preference, 116, 105149. https://doi.org/10.1016/j.foodqual.2024.105149
Califano, G., Zhang, T., & Spence, C. (2024). Would you trust an AI chef? Examining what people think when AI becomes creative with food. International Journal of Gastronomy and Food Science, 37, 100973. https://doi.org/10.1016/j.ijgfs.2024.100973
Zampini, M., & Spence, C. (2004). The role of auditory cues in modulating the perceived crispness and staleness of potato chips. Journal of Sensory Studies, 19(5), 347–363.