The New York Times officially launched on November 7, 2026 an integrated artificial intelligence assistant within its NYT Cooking application, one of the most prestigious recipe platforms in the digital world. The feature, initially available for subscribers on iOS, places a language model directly alongside the recipe the user is preparing, enabling real-time ingredient adaptation, portion rescaling, and cooking instruction adjustment — all without modifying the original recipe text.
The hands-on test conducted by The Verge reveals both the potential and the limitations of a generative AI system applied to home cooking. The assistant excelled at flexible cooking — like rescuing a chicken dinner when ingredients were missing — but stumbled at tasks requiring chemical precision, such as baking a cake, where it suggested the wrong amount of butter by confusing recipe versions.
What the assistant does — and how
The assistant appears as a discreet text box on the recipe page, with a prompt that reads "Customize this recipe…". Users type questions like "Can I swap chicken for chicken tenders?", "I have spinach instead of kale, what changes?", or "How do I adapt this recipe for 4 people instead of 8?". The system processes the question in the context of the open recipe and returns adapted instructions in a separate tab, leaving the original recipe intact.
This "secondary version" design is intentional. The New York Times kept the original recipe authors — who are professional editors and gastronomy columnists — and pushed AI-generated modifications to an isolated space. The result is that users can compare the original and the adapted version side by side, something many other generative AI apps do not offer.
The assistant's knowledge base is not generic: it was trained on the thousands of recipes in the NYT Cooking library, which includes recipes from newspapers, restaurants, cookbooks, and columns, all tested by humans. The model also draws on the newspaper's editorial guidelines and years of accumulated reader comments — which partly explains why the assistant refused an absurd request to substitute ladyfingers with minced beef in a trifle.
Where AI shines: flexible cooking
The assistant's strength is flexible cooking, especially savory dishes. When the reviewer needed to rescue a chicken dinner at night at 7PM, the AI suggested timing and temperature adjustments that worked in practice. In another instance, it suggested adding spinach to the dish after the kale — because spinach wilts faster — a tip that reflects real kitchen empirical knowledge, not just database output.
The assistant also handled ingredient swaps (like using chicken tenders instead of whole chicken breast), converted metric measurements to imperial, and adapted recipes for dietary restrictions. The reviewer tested adapting a Bundt cake recipe into muffins and halving the yield — and the system managed to manipulate quantities coherently within those limits.
Where AI fails: the chemistry of baking
The cake was the weak point. During one test, the reviewer adapted a cake recipe into muffins and asked to halve the yield. While cooking the original version and asking the assistant how much butter was needed, the system responded with six tablespoons — the amount for the modified version, not the original, which called for three-quarter cup. The confusion between versions was serious enough to completely ruin a confectionery product.
This type of error is not new in language models: context confusion appears whenever the model needs to hold multiple instances of data in mind simultaneously. But in baking, where the exact proportions of fat, sugar, flour, and leavening agents are the difference between a fluffy cake and a stone, the margin for error is practically zero.
What this means for the future of AI kitchens
NYT Cooking is not the first app to integrate generative AI. Several competitors — Samsung Food (formerly Whisk), Pluck, SideChef, ChefGPT — already offer recipe assistants with artificial intelligence. The NYT Cooking differentiator lies in the combination of three factors: (1) a recipe library with centuries of integrated editorial expertise; (2) a design that preserves the original work of authors; and (3) a model trained on the specific context of cooking, not on a generic corpus.
The baking limitation, however, raises a broader question: current generative AI is good at adapting recipes approximately, but still cannot understand the chemical relationships underlying cooking processes. It can confidently say "swap milk for almond milk" but cannot predict that the absence of milk protein will alter the dough structure. This distinction between flexible cooking and precise baking will be the limit of all current generations of AI cooking assistants.
Meanwhile, the NYT Cooking case is an example of how digital publishers are integrating AI into existing products — not as a separate feature, but as a layer of contextual intelligence over content that already has authority. The question that remains is: if NYT Cooking takes this step, which other paid content services will follow?
Sources: The Verge, The New York Times, AIToolly
✓ Independent sources cross-checked and verified before publishing
