NYT Cooking’s AI saved chicken, then nearly ruined the cake

By James Calloway

At about 7 p.m. on a weeknight, the new AI assistant inside the NYT Cooking app helped me get chicken on the table, then nearly cost me the cake. That split outcome matters more than a general promise of faster meal planning: one meal worked, while the dessert nearly failed.

The routine starts at the refrigerator, where I try to choose dinner without triggering my husband’s complaint: “Chicken again?” I do not meal plan, so the same small decision recurs. Usually, chicken wins before I have compared recipes or checked ingredients.

NYT Cooking now includes a built-in AI assistant, putting help at the point where I decide what to cook. Before this test, my kitchen already had three AI touchpoints: Alexa Plus on an Echo smart display across the room, the Samsung Food app on the refrigerator, and an AI-powered cooking robot.

I had already described the arrangement: “AI has become my go-to sous chef — albeit one I need to keep a close eye on.” Across the setup, AI helped me swap ingredients, generate dinner ideas, and fix things after I made a mistake. The usefulness was obvious; so was the need for judgment.

What does the built-in assistant change?

The meaningful interface change is placement. Alexa Plus answers from a countertop display, Samsung Food appears on the refrigerator, and the robot offers another device in the kitchen. NYT Cooking now attaches the assistant to the place where recipes and the evening’s decision meet.

That places AI assistance in the same application as recipe selection and preparation. It also raises the cost of a poor suggestion. Advice delivered at the planning stage may redirect dinner; advice delivered during preparation can affect the result of a dish already underway.

The reported AI use maps to three specific jobs:

• Swap ingredients: revise a dish when the available item differs.
• Generate ideas: create options when the refrigerator offers no clear answer.
• Fix mistakes: help correct the cook’s work after something goes wrong.

Those functions address three points where a household cooking routine can stall. The NYT assistant’s added value is its location. It puts AI help inside the recipe application while I am already choosing and preparing food; the other tools place assistance elsewhere.

For this test, success had a blunt household measure. The chicken became dinner, so the assistant helped remove the nightly choice. The cake came close to failure, so the same system also demonstrated why an answer cannot substitute for inspection at the stove.

Why did the cake expose the risk?

The near-failure matters because the assistant did not merely supply an optional idea. It helped direct a task with a visible outcome. A successful dinner removes friction from a routine that repeats around 7 p.m.; a ruined cake removes the result the cook was trying to make.

The account does not identify the exact cake error, the number of prompts used, or the assistant’s response time, so it cannot establish an accuracy rate for NYT Cooking. It establishes a more limited but useful finding: in one home kitchen, the assistant produced one rescue and one near-miss.

That is not a product benchmark. It is also not an excuse to dismiss the workflow. The built-in assistant reached a decision that had defeated a tired cook and supplied help during a live mistake, but it also placed enough trust in the cook to preserve the outcome.

What should the assistant make clearer?

The design task is to make proposed changes inspectable before they become physical actions. If the assistant substitutes an ingredient, revises a quantity, or suggests a corrective step, the interface should identify that change instead of burying it inside a conversational reply.

The cook also needs a clear point to pause, compare the suggestion with the pan or bowl, and reject it. That need follows directly from the test’s result: a plausible instruction can support the chicken dinner and still produce the wrong direction for the cake.

Those checks preserve agency without slowing the routine with a formal approval process for every question. The assistant can handle the initial narrowing of choices; the cook should retain control over whether the proposed dish, substitution, or repair should proceed.

The next test for NYT Cooking’s assistant is simple: repeat enough weeknight meals and cakes to learn whether its rescues outnumber its near-misses, with prompts and outcomes recorded rather than judged by one successful dinner.