The New York Times Cooking AI Assistant Is a Sous-Chef, Not a Chef
The NYT Cooking app, subscribed to by millions of The New York Times readers, has rolled out a built-in artificial intelligence assistant on recipe pages. The feature, initially available only on iOS for a portion of subscribers, appears as a text box on each recipe page with a prompt inviting users to customize the recipe.
The key difference from asking a generic chatbot to "create a cinnamon roll recipe" is substantial. The NYT Cooking assistant works on specific recipes from the catalog — thousands of dishes tested by professional editors, drawn from decades of food columnists, restaurant archives, and cookbooks. Users don't receive a new recipe out of thin air; they receive modifications applied to a tested recipe, with changes highlighted in orange on a separate tab.
The stress test: dinners work, baking doesn't
The most detailed review, published by Jennifer Pattison Tuohy in The Verge on November 7, tested the assistant in two distinct scenarios that reveal the pattern of successes and failures. For a typical weeknight dinner — a chicken and rice dish that gets repeated every Friday — the assistant demonstrated immediate practical utility. It suggested ingredient swaps based on what was available, adjusted cooking times for different chicken cuts, and adapted the steps when a herb was missing. The result, according to the tester, "saved dinner".
In baking, the situation was quite different. When asked to convert a Cinnamon Roll Bundt Cake into muffins — a common request for someone who doesn't own the right pans — the assistant correctly converted the cooking time (from three hours to 50 minutes), divided the batter into two 12-unit tins, and converted metric weights. So far, everything was coherent.
But here lies the structural problem: when switching between the original and modified recipe, the assistant confused the butter quantity, suggesting only half of what was needed. The result was a dry cake, as documented by the tester in the comparative photos published in the article.
Why baking and cooking are fundamentally different domains
The distinction is not accidental. Cooking — the kind of preparation where you swap an herb for another, where you adjust a seasoning — is a domain where variations are tolerated. Baking, in particular, depends on precise ratios: fat, sugar, flour, and leavening must be in exact proportions. A 20% deviation in butter quantity alters the protein structure of the cake, the melting point of the fat, and the hydration of the dough.
Language models like the one powering the NYT Cooking assistant are trained on millions of recipes, comments, and culinary texts. They learn language patterns, not chemistry. When the model "knows" a cake recipe calls for butter, it can infer a reasonable quantity; but it doesn't truly "know" that the butter needs to be creamed rather than melted, or that the order of incorporating ingredients changes the final result. As The Observer noted in broad coverage of the topic, the model "lacks intuition about texture, seasoning balance, or how a sauce should feel when properly reduced".
Training: human comments as supervision
A detail worth noting is what the assistant learned, beyond the general corpus of recipes. According to Genevieve Ko, senior editor at NYT Cooking, the tool was partially trained on users' own recipe comments — that space where readers have long discussed ingredient swaps, technique fixes, and diet adjustments. It's a curious datum: the assistant was trained not only by professionals, but also by the collective wisdom of readers who, for years, solved the same problems the chatbot now resolves automatically.
This explains why the model could "know when to say no" — when asked to use minced beef in a dessert trifle, it refused. The system has filters against bad results, and Ko confirmed the launch was made gradually, with continuous performance monitoring, as reported in the original NYT article.
The competition with Alexa Plus and Google Gemini
The Verge tester compared the NYT Cooking assistant with Alexa Plus on an Echo Show 11 smart display — Amazon's voice AI assistant — and the Samsung Food app on a smart refrigerator. In ingredient swap tasks for the chicken and rice dish, both offered similar guidance but with slightly different cooking times. The NYT app's advantage was displaying the modified recipe visually; Alexa, being voice-only, "couldn't show the adapted recipe on screen," limiting itself to narrating changes while keeping the original recipe visible.
Google's Gemini also offers a kitchen assistant, with the promise of "expert advice and personalized help" while cooking. But Google's approach is more generic: instead of working on specific recipes tested by professionals, it generates free-form responses. NYT Cooking goes the middle ground: it keeps the original recipe intact, generates a customized version on a separate tab, and lets the user review changes before starting.
What this means for the future of cooking apps
NYT Cooking is not a niche experiment. It has a substantial base of paying subscribers, and this is the first major recipe app to integrate generative AI natively — not as a separate feature, but as an integral part of the user experience. The decision to start on iOS, gradually and with monitoring, reflects strategic caution: the company wants to see how users interact before expanding.
The pattern emerging from reviews is clear. For everyday tasks — swapping ingredients, scaling portions, adapting for different pan types — the assistant works as a "sous-chef": useful, efficient, and with good contextual judgment. For baking and preparations where chemical precision is critical, it still demands active human supervision. This is not a product flaw that will be fixed with more tokens; it's a fundamental limitation of language model architecture for tasks requiring physical world understanding.
If the NYT Cooking assistant represents the leading edge, what will be the next major consumer app to integrate generative AI — and how far will users trust it before discovering the difference between a good sous-chef and a bad chef?
Sources: The Verge, The New York Times, The Observer
✓ Independent sources cross-checked and verified before publishing
