Some of the competing chefs must have felt something change beneath them at one point, between the judges’ deliberations and the trophy handover. Not the anxiety that comes with losing; all professional cooks have that. Strange. They were informed that an AI had created the winning recipe. During development, there is no tasting. No instinct developed over many years in a hot kitchen. It is merely a language model that draws from a vast array of pre-existing recipes and produces something that just so happens to be perfect.
Whether this will turn into a pivotal moment in the history of competitive cooking or just an odd anecdote remains to be seen. Professional chefs, however, have responded sharply, which is not wholly unexpected. Sweat, flame, and the kind of muscle memory you can’t download seem to have encroached upon a space that seemed to be shielded by its very physicality.
Here, the larger context is important. AI has been subtly entering kitchens for a few years now in ways most diners wouldn’t notice, such as suggesting wine pairings, identifying food waste, and streamlining supply chains. The industry had essentially come to terms with that version of AI. This is not the same. Winning when competing against humans on terms created by humans touches a nerve that operational efficiency could never.
The fact that the recipe itself wasn’t absurd is what really complicates the situation. This one worked, unlike the crockpot mojito cookbook that turned out to be almost entirely mojito recipes or the bratwurst ice cream that an early GPT-3 model famously recommended as a side dish. The dish appeared to be tasty, technically sound, and cohesive. And no one is quite sure how to handle that. Because the model had never tasted anything.
When food writer Rukmini Iyer noted that an AI hasn’t prepared and tested six iterations of a dish before arriving at its best attempt, she put it succinctly. Human recipe development is often messy, sensory, and iterative. AI ignores all of that and uses pattern recognition alone to arrive at a plausible conclusion. Occasionally, the pattern is incorrect. Apparently, it isn’t always the case.
Beneath all of this, the issue of intellectual property remains largely unresolved. Cookbooks, food blogs, and culinary databases are used to train AI recipe models. When that training results in a dish that wins a competition, it’s fair to wonder whose knowledge produced it. Joanne Lee Molinaro, a Korean vegan cookbook author, discovered firsthand what it’s like when AI gets too close to your work. Her book was almost exactly duplicated, cover and all, by a fictional author whose vegan recipes somehow included chicken. The victory in this competition seems like a different take on the same issue: the extraction, recombination, and faceless presentation of creativity.
It’s difficult not to feel sorry for the chefs in that room as you watch this happen. They spent years training. They have a deeper understanding of texture, heat, and acid than any model, no matter how advanced, can. And yet, here we are. AI’s ability to cook isn’t really the question at hand. What matters is whether the industry will make the decision.
