NYT Cooking Rolls Out an AI Assistant for Subscribers: Great at Ingredient Swaps, Almost Ruined One Reviewer's Cake

The New York Times has quietly shipped one of the more interesting consumer AI features of the season: a built-in assistant inside NYT Cooking, its recipe subscription app. The Verge's Jennifer Pattison Tuohy spent several days with the tool and came away with a verdict that captures where consumer AI actually is in late 2026 — genuinely useful at most everyday tasks, still capable of a failure that ruins dessert.

The assistant appears as a small text box on each recipe page, prompted with "Customize this recipe..." From there, subscribers can ask for ingredient substitutions, scale a recipe to any number of servings, request vegan or gluten-free versions, get equipment swaps when they lack a Bundt pan, and convert measurements. Notably, it never edits the original recipe: it creates a new, private customized version, so the tested original stays untouched — a design choice that addresses much of what recipe purists feared.

Two details stand out in how it was built. First, the model was trained in part on NYT Cooking's own reader comments, the sprawling community knowledge base where subscribers have long answered each other's substitution questions — a clever way of turning fifteen years of comment-section expertise into training signal. Second, in Tuohy's testing it declined to blindly follow absurd requests, a guardrail many consumer chatbots still lack.

It is not flawless. The assistant nearly talked her into using half the required butter in a cake — the kind of ratio error that is fatal in baking, where chemistry does not forgive hallucinated quantities. The failure mode is familiar: language models are strong at the flexible, judgment-based swaps that define savory cooking, and weak at the exact numerical constraints of baking, where a halved butter quantity collapses the crumb.

The rollout has not been friction-free. NYT Cooking first tested AI recipe modification with subscribers earlier, drawing vocal pushback from users who objected to AI features on principle and asked, unsuccessfully, for a way to disable the tool; the Times says it is completely optional to use. The feature is available to subscribers on iOS at launch, with broader availability promised soon, and the Times has framed it as an extension of the app's job of making tested recipes work in real kitchens.

Strategically, the move is a sign of how legacy publishers are approaching AI in 2026: not by licensing their archives to model builders, but by wrapping AI around their own trusted, human-tested content and keeping the subscriber relationship intact. The Times, currently in litigation against AI data pipelines it says used its work without permission, is making a pointed argument — the value was never the text, it was the testing, the community and the trust. An assistant trained on reader comments, guarded by human editors, is that argument rendered in product form.