The machine had never tasted anything. It had never smelled a grill, never felt the give of a bun, never known whether salt belonged anywhere near beef. What it had was 2,216 burger recipes, scraped and tidied from the sprawl of a public cooking website, and a question its makers had bolted on top: not which burger was likeliest, but which burger was best. From that, with no rules about flavor written into it anywhere, it taught itself the rough shape of what humans want to eat. Then it went looking for something better.
That something turned out, in a blind test in a San Francisco restaurant, to outscore a Big Mac. One hundred and one volunteers chewed their way through six burgers without knowing which was which, and rated an AI-invented recipe higher on flavor than the burger McDonald’s has sold in more than a hundred countries.
From Predicting to Inventing
The system is called BurgerAI, and it comes out of the Living Matter Lab at Stanford, run by Ellen Kuhl, a mechanical engineer who now directs the university’s Bio-X life sciences institute. Kuhl is blunt about why this matters beyond lunch. Most AI systems are trained to predict what already exists. We wanted AI to invent what should exist next,
she says. The distinction sounds small and isn’t. A predictive model finishes your sentence; a generative design model is asked to solve for an outcome you specify and then hand you something that did not previously exist. As Kuhl puts it, BurgerAI does not ask which burger is most likely. It asks which one best satisfies a tangle of competing objectives.
Those objectives are the catch. A burger has to taste good, which fights against being good for you, which fights against being kind to the planet. Pulling on one rope tends to drag the others the wrong way.
Underneath the appetising name sits a fairly austere piece of machinery: a diffusion model, the same broad family of AI that powers image generators, rather than the large language models that produce text. It works in two stages, first deciding which of 146 possible ingredients go in, then working out how much of each. The team trained it on burgers filtered from a haul of more than half a million recipes, and the numbers it was navigating are faintly absurd. There are, on the lab’s reckoning, more than 10^43 ways to combine those ingredients, which is to say more possible burgers than there are stars in the observable universe, by a margin that makes the comparison feel almost rude.
To check the thing actually understood burgers rather than just memorising them, the researchers set it a peculiar test: rediscover the Big Mac. The recipe was deliberately kept out of its training data (McDonald’s keeps the real one proprietary, so the team stitched together a reference from four open-source copycats). On average, across ten runs, the model had to generate 7.3 million burgers before it stumbled back onto that exact combination. A reassuring result, oddly. It means the famous recipe sits where it should, in a high-probability pocket of the design space, recognisable but not trivially easy to land on.
The Taste Test Was the Real Exam
Then came the part no equation could settle. Recipes are not food, and an ingredient list is not a meal, so the lab brought in an executive chef to turn the AI’s cold inventories into actual cooking instructions, then handed those to a separate kitchen to prepare. Diners scored everything on a seven-point scale. Two of the AI’s “delicious” burgers matched or beat the Big Mac on overall liking, flavor and texture, with one drawing notably more votes for tasting meaty, moist and frankly fatty. The AI did not just generate plausible recipes, Kuhl says; it created burgers that real people enjoy.
The greener results are where it gets interesting. A mushroom burger the model designed carried an environmental footprint more than ten times lighter than the Big Mac’s, by a score combining land use, water, emissions and pollution. Postdoctoral fellow Vahidullah Tac, the paper’s first author, had braced for that to taste like a compromise. We expected some trade-off between sustainability and consumer acceptance,
he says. The mushroom version did slip on the ratings, earthy where testers wanted savoury. But a beef-mushroom blend held the line, landing on par with the Big Mac while still cutting its impact. But we found a burger with dramatically lower environmental impact could still compete with one of the world’s most successful burgers,
Tac says.
The nutrition story is honest about its limits. The model’s healthiest creation, a bean burger, scored nearly twice as well as the Big Mac on a standard dietary index and used a sixth of the environmental resources, but diners were not fooled into loving it: bland, dry, grainy, they said. There is no free lunch in here, only a more clearly drawn map of where the trade-offs actually lie, which is arguably the more useful thing to hand a food company.
And it cuts both ways, this honesty. The training data leaned heavily Western, so the model knows burgers and not much else. It captures ingredients and amounts but nothing about how you cook them, which is a fair chunk of why food tastes the way it does. The environmental and nutritional figures lean on global averages, so they are best read as comparisons rather than verdicts. The researchers say all of this plainly, which is more than a lot of AI papers manage.
Why Burgers Were Never the Point
For Kuhl and Tac, the burger is bait. The lab has put out a companion paper showing that the same mathematics behind BurgerAI underpins generative design more widely, the kind used to invent new materials. Food choices are some of the most consequential decisions humans make every day,
Tac says, and food happens to be a problem with every hard feature you might want to practise on: huge design space, clashing goals, a verdict delivered by actual human senses rather than a benchmark. If a model can balance taste against carbon against protein, the argument runs, it might balance efficacy against toxicity in a drug, or strength against weight in an alloy. Kuhl describes food as a model system for AI as a partner in discovery rather than an autocomplete with ambitions.
Whether any of this reaches your plate is a separate question, and a harder one. A model can hand a company a burger that is cheaper for the planet and roughly as nice to eat, but it cannot make anyone order it. The burger is just the beginning,
Kuhl says, and on the evidence she may be right, though the beginning of what is the part nobody has tasted yet.
DOI / Source: Tac, Gardner & Kuhl, npj Science of Food 10, 199 (2026)
Frequently Asked Questions
Is it true an AI burger actually beat a Big Mac in a real taste test?
In one specific sense, yes. In a blind tasting with 101 people at a San Francisco restaurant, two AI-designed “delicious” burgers matched or outscored a Big Mac on overall liking, flavor and texture, with one rated higher on flavor. It is worth knowing that the same project also produced healthier and greener burgers that testers liked less, so the win was real but not the whole picture.
How does an AI design a recipe without ever tasting anything?
It uses a diffusion model, the same kind of AI behind image generators, trained on 2,216 real burger recipes to learn which ingredients and amounts tend to go together in food people actually make. It treats popularity as a stand-in for tastiness, then searches for new combinations that hit a target like low environmental impact. The real test is still the kitchen, which is why the team cooked the recipes and fed them to real diners.
Why does a project about burgers matter for anything beyond food?
Because the hard part is not the burger but the method: balancing several competing goals at once inside an enormous space of possibilities. The researchers argue the same approach could help design drugs, materials, or other products where taste, in effect, is replaced by efficacy, strength or cost. A companion paper shows the underlying mathematics is shared with generative design used in engineering.
Could one of these burgers actually show up on a menu?
Technically there is little stopping it, since the healthier and lower-impact recipes use ordinary whole-food ingredients rather than exotic additives. The catch is that a model can design something cheap for the planet and pleasant to eat, but it cannot make anyone choose it, and adoption, not invention, has long been the real bottleneck for sustainable food. Whether the economics and the appetite line up is the question still waiting to be answered.
