At the request of a Verge reader, Jennifer Pattison Tuohy tested the new AI assistant built into the NYT Cooking app — and found that while it is effective at solving everyday kitchen problems, it fails completely when it comes to baking, where ingredient ratios don't forgive.
Originally launched in October 2026 by senior Cooking editor Genevieve Ko, the tool works as a chatbot within the recipe page itself. Users receive a text box with the prompt "Customize this recipe…" and can request everything from ingredient swaps to scaling for different serving sizes. The system was trained on years of community user comments and the Times' own editorial guidelines.
The NYT approach is deliberate: rather than rewriting the original recipe, the assistant generates a modified version in a separate tab, keeping the tested and approved recipe intact. The application's own help page warns that "this feature is in beta and AI can make mistakes" — which sounds like boilerplate, but is actually the most honest description of a language model applied to tasks requiring precision.
For a weeknight dinner, the assistant performed exceptionally well. The tester asked to adapt a ginger chicken recipe that required ingredients she didn't have on hand — chard, cilantro, and scallions. In seconds, the assistant generated a new recipe with a new title, incorporating chicken tenders, red onion, and green beans.
The system also recalculated cooking times and steps to ensure the substitute ingredients cooked evenly. And it doesn't just suggest obvious swaps: when asked about using baby spinach instead of kale, it didn't just suggest the substitution but indicated that the spinach should be added in a later step because it wilts faster.
In this category of task — flexible cooking with room for approximation — the AI shines. Rather than scrolling through hundreds of comments to find the right answer, the assistant synthesizes culinary guidance into immediate, actionable steps. The tester compared the results with Amazon's Alexa Plus on an Echo Show 11: the NYT Cooking assistant provided similar guidance but with the advantage of displaying the modified recipe on screen.
The cake. Specifically, a Cinnamon Roll Bundt Cake. The tester didn't have a Bundt pan and asked for an alternative. The assistant suggested a standard loaf pan, generously buttered. The result: the cake stuck to the pan and crumbled. The AI didn't account for the fact that the pan changes the entire baking time and texture.
In another test, the assistant nearly convinced the tester to use half the required butter in the cake — a context error that, in savory recipes, would be merely inconvenient, but in baking is catastrophic.
This is the fundamental trap of language models applied to precise tasks: they can discuss food fluently without truly understanding the chemistry behind it. Cakes depend on precise ratios of fat, sugar, flour, and leavening. A 10% variation can mean the difference between a fluffy cake and a dry brick.
NYT Cooking has a substantial paid subscriber base. This isn't a niche experiment — it's generative AI in production, with thousands of real users testing it daily. The problem is that user confidence tends to increase after early wins, making later failures more damaging.
This pattern has been observed in other AI integrations: confident output and accurate output are not the same thing, and the gap tends to appear at the edges of the task, not the center. An assistant that is helpful 80% of the time but silently wrong the other 20% is not the same as one that is reliable.
The lesson for any company integrating AI into consumer products is to scope carefully. Define where AI is genuinely reliable before putting it in front of customers. And when the task involves baking, perhaps the best approach is to explicitly signal when the AI is operating outside its comfort zone.
The question remains: will the next generation of culinary assistants be able to distinguish between "can adjust" and "don't touch this" — or will we continue to witness costly mistakes as more AI tools enter the kitchen?
Sources: The Verge, New York Times, AIToolly
✓ Independent sources cross-checked and verified before publishing
