The most revealing part of a new AI burger experiment happened after the model had finished. Chef Justin Schneider refined the quantities and supplied cooking instructions. In an October 11 Stanford preprint, that collaboration produced a mushroom burger that 101 restaurant diners rated 5.3 out of 7 for overall liking.[1]
From ingredient weights to cooking
The method, described in the team’s earlier June paper, divides recipe generation into two decisions. One diffusion model chooses which ingredients appear. A second proposes their quantities. During training, recipes are deliberately corrupted; the models learn how to reverse that damage. Generating a recipe then runs the learned process from noise towards a plausible combination. The training collection contained 2,216 recipes represented through 146 ingredients.[2]
Splitting presence from quantity makes intuitive sense. Whether a burger contains cheese is a different decision from how much cheese it contains. An ingredient can have a small mass and a large effect on flavour; another may supply most of the bulk. Treating a recipe as one undifferentiated list would obscure those different jobs.
The earlier system learned ingredient weights, without cooking instructions. Its kitchen workflow therefore assigned an executive chef the job of converting candidate formulations into preparation protocols, which other chefs could follow. This is a consequential division of labour: the model supplies a structured proposal, while a human specifies how it becomes food.[2]
Consider a hypothetical list containing mushrooms, flour and oil. It leaves plenty undecided. Are the mushrooms whole or chopped? Does the cook remove moisture before combining them with the flour? Is the oil used inside the mixture or on the pan? Those choices can produce very different things to bite. A promising list still contains many possible meals.
The preparation protocol therefore belongs with the ingredient weights. Changing a cooking step can change the result even when the formulation stays the same.
What the diners actually judged
The six meat-free burgers were served in random order at a San Francisco restaurant, with ingredients undisclosed. Participants ate burgers regularly; vegetarians and vegans were excluded. The mushroom version scored 5.5 for flavour and 5.0 for texture. A Big Mac scored 5.3 overall, 5.4 for flavour and 5.2 for texture in an earlier survey. Different diners make those figures a historical reference, not a controlled tie.[1]
A favourable tasting identifies a candidate worth developing. Establishing equal appeal would require a study designed for that comparison, with the products judged under the same conditions.
The mushroom burger’s estimated environmental score was below one-third of the Big Mac’s. This weighted ingredient-level index combines greenhouse gases, land use, water scarcity and eutrophication; it is not a carbon-footprint ratio. The recipe contains provolone, so it is meat-free, but not vegan.[1]
A composite score also makes a design choice visible. Combining environmental pressures into one number requires deciding how much weight each receives. A kitchen concerned chiefly with scarce water might choose differently from one focused on emissions. A generator can offer alternatives, but someone still has to decide what a better burger means.
The next persuasive test would take this collaboration beyond a tasting session: the same preparation repeated across kitchens, at a workable cost, with customers choosing to order it again. Those are different questions from whether an algorithm can propose a surprising recipe. The interesting destination is an ordinary one—a meal good enough to become somebody’s usual order.
Sources & further reading
- Generative AI enables human-centered design of sustainable plant-based foods arXiv / Stanford University · 11 Oct 2026 · Accessed 17 Oct 2026
- Generative artificial intelligence creates delicious, sustainable, and nutritious burgers npj Science of Food · 26 Jun 2026 · Accessed 17 Oct 2026