The Verge reviewed the NYT Cooking AI assistant: useful at dinnertime, dangerous in the oven
The New York Times Cooking app has launched a built-in artificial intelligence assistant. The promise is straightforward: ask a question mid-cook, swap out an ingredient you don't have, adapt a recipe to what's in your pantry. It should be as basic as reaching for a measuring cup. And in theory, the most capable generative AI system in the world should be able to handle exactly that.
The problem is that AI doesn't actually know how to cook — it just talks about food. And the difference, when your hands are covered in flour and the oven is at 425°F, is enormous.
What the assistant does (and where it excels)
The feature was announced in October 2026 by Genevieve Ko, senior cooking editor at the New York Times, in an article titled "How, and Why, We're Adding A.I. Tools to NYT Cooking." The product and engineering team recognized that users had been asking for years for a way to adjust recipes without having to recalculate or scroll through comments at the bottom of the page: measuring ingredients in grams, doubling or halving recipes, swapping ingredients for alternatives. The AI assistant, accessed via a star icon (✦) on the recipe page, was designed to solve all of this conversationally.
In practice, it really does work for flexible tasks. According to Jennifer Pattison Tuohy's hands-on review in The Verge, when dealing with a standard weeknight chicken recipe — that universal dilemma of "I don't want the same protein again" — the assistant suggests alternatives, adjusts spices, and reorganizes the recipe in real time. The same goes for simple substitutions: you don't have basil? It suggests mint. You need to double the recipe? It recalculates everything. For the messy world of evening cooking, where room for error is the norm, the assistant performs well enough to become a daily habit.
What sets NYT Cooking apart from generic tools is its ecosystem of tested recipes backed by professionals. While voice assistants like Amazon Alexa Plus or the Samsung Food app try to extract recipes from random blogs across the internet, NYT Cooking operates on an editorial catalog with expert testing behind every recipe. This means the AI's suggestions have a quality floor that a standalone ChatGPT doesn't offer.
The problem: when recipes become chemistry formulas
That's where everything fell apart.
Tuohy took Genevieve Ko's Cinnamon Roll Bundt Cake recipe and asked the assistant to convert it into muffins. Anyone who's tried knows that the question "who owns a Bundt pan?" is genuinely realistic — the specific pan is expensive, takes up space, and you use it once every two years. The assistant adjusted the cooking time from three hours to 50 minutes and divided the batter across two 12-unit tins. It sounds reasonable. The result, according to the review, was a disaster.
The problem isn't specific to NYT Cooking's assistant — it's the fundamental limitation of any language-based system applied to baking recipes. Baking isn't cooking. Cooking is art: you adjust seasoning to taste, swap herbs, improvise with what you have in the pantry. Baking is chemistry: the ratio of fat, sugar, flour, and leavening agent is non-negotiable. Get the ratio wrong and it doesn't matter how good a cook you are — the cake won't rise, stays raw in the middle, or turns into a flat, dry disk.
A 2026 study by Recipy tested 12 AI-powered ingredient substitutions and found that popular models frequently recommended dangerous conversions without flagging context. ChatGPT and Claude suggested swapping baking powder for baking soda in a 3:1 ratio without warning that this only works if you eliminate the recipe's acid — in recipes with buttermilk or yogurt, the result is flat and metallic-tasting. Only Gemini and Perplexity flagged this. What Recipy concluded, in summary: "The honest answer to this question in most recipes is 'pick a different recipe tonight,'" and only Gemini said it plainly.
Why this matters for the entire industry
NYT Cooking isn't a niche app — it's a platform with thousands of paying subscribers. When an AI assistant gives bad baking advice to even a fraction of that base, that's a real-world problem, not just a benchmark failure. The pattern that emerges is the same one seen across any product that adds AI to its user interface: the assistant works well 80% of the time, but is quietly wrong the other 20%. The problem is that users tend to trust the system more after early wins, making later failures much more damaging.
Food scientist Rachel Routhenstein warned in a Seattle Times piece that "AI may swap whole-wheat flour for gluten-free oat or almond flour, but may not take into account how to adjust liquids and binding agents, which can offset the texture and palatability." In other words: AI confuses fluency with understanding.
What to watch
NYT Cooking should be more honest about when AI is pushing its limits. A system that detects the user is asking to adjust a baking recipe could display a warning before applying the modification — something like "This recipe relies on precise ratios. Your suggestions may require manual adjustment." The same logic applies to any product integrating AI for tasks requiring precision: define where the AI is reliable before putting it in front of customers.
The question isn't whether AI assistants will enter the kitchen — they already have. The question is how much we'll trust them for recipes where a mistake isn't just a wrong dinner, but wasted ingredients, time, and patience. And what happens when the assistant that helped you make Sunday dinner also convinced you to swap leavening for baking soda and the cake came out tasting like soap? The technology removes friction, yes. But at some point, the friction was what kept you from breaking everything.
Sources: The Verge, The New York Times, Lumien
✓ Independent sources cross-checked and verified before publishing
