NYT Cooking's new AI assistant won't just say yes to your recipe swap
The Times built its recipe chatbot to check three other sources before it lets the model answer. A hands-on review found the gap between that and a clean experience.
NYT Cooking’s reader comments have long been the unofficial substitution guide for the app’s 23,000-plus recipes: someone has usually already asked “can I use spinach instead of chard” years ago, with the answer buried deep in the thread. So when I read how the paper built a chatbot to replace that scroll, the part that caught me wasn’t the AI itself. It was what the AI is only allowed to do last.
NYT Cooking’s new AI assistant launched on October 23 and is rolling out gradually to subscribers on iOS, with wider availability planned, according to Genevieve Ko, the senior editor who runs the Cooking section’s editorial side. Ko wrote the announcement herself, and it reads less like a product launch and more like someone talking herself into trusting the thing she built. “I had to take a lot of deep breaths to deal with inaccurate, florid responses,” she wrote, describing the training process before the Times rewrote the model’s prompt using “decades of culinary knowledge.”
A fixed order of sources, model last
Something else stands out: the order the tool checks before it ever generates a new answer. Per the Times’s own help page, a customization request goes through four sources in a fixed sequence: the recipe’s own text and notes first, then the Ingredient Substitution Guide, then matching reader comments, and only after all three come up empty does the model generate something new from what the help page calls “principles established by our editors.” That’s a deliberate hedge against the thing every recipe chatbot gets mocked for: making up a plausible-sounding swap with no backing. The model is the fallback, not the first move.
It also won’t agree with you by default. Ko says the team trained it specifically against the instinct chatbots have to flatter a bad idea. The help page backs that up in practice: it can decline a request that “would not work in a recipe” and redirect you to a closer existing recipe instead of forcing your version to fit.
Where the hands-on test found trouble
A hands-on review from The Verge’s Jennifer Pattison Tuohy tested that claim directly, and it mostly held. She asked the assistant to swap ladyfingers for minced beef in a trifle (a reference to the Friends episode where Rachel does exactly that), and it refused flatly: “minced beef cannot be used in place of the ladyfingers or anywhere in this sweet dessert trifle.” Good call. On a weeknight chicken and rice dish, it swapped in ingredients she had on hand, adjusted the cooking method for them and highlighted every change in orange against the original. She called the result “delicious” and said it beat guessing from the comments section, which is what she, and most regulars, did before.
| What she tried | What the assistant did | How it went |
|---|---|---|
| Swap in chicken tenders, green beans, no cilantro | Rewrote the recipe, adjusted cook steps, highlighted every change | Worked well, called “delicious” |
| Swap minced beef for ladyfingers in a trifle | Refused the request outright | Correct refusal, no bad dish made |
| Asked by voice for the Bundt cake’s butter amount mid-bake | Answered with the customized version’s 6 tablespoons, not the original’s three-quarters cup, without saying which | Would have produced a dry cake if followed as given |
Tuohy halved the Bundt cake recipe for 12 muffins, which the assistant handled fine, including rounding five halved eggs to a sensible three rather than an awkward two and a half. The butter mix-up, from the table above, happened later, toggling mid-bake between the original recipe and her saved customized version, and Tuohy wrote that the wrong amount would have produced a dry cake. My read: that’s not the model inventing a fact. It’s a chat interface that doesn’t reliably tell you which recipe it’s talking about, which is arguably a harder problem to catch, because it looks like confidence rather than error.
NYT Cooking’s own help page is upfront that this is a beta: outputs “have not been reviewed by NYT Cooking editors and are not endorsed by the original recipe’s creators,” and the feature “can make mistakes.” Customized versions can’t be saved to your Recipe Box, rated or shared, which keeps any AI-generated version from circulating as if it were a tested Times recipe.
My read, and what to watch for
My read: the source-priority design is the actual story here, more than the chatbot wrapper. A lot of AI cooking tools skip straight to generation and hope the model has seen enough recipes to be plausible. NYT Cooking is routing around its own model whenever a human-written answer already exists, which is a more conservative design than most consumer AI features ship with. The cost of that caution shows up exactly where Tuohy found it: not in bad advice, but in a toggle-state bug that a plain recipe app would never have, because a plain recipe app doesn’t let you ask it questions mid-cook.
If you use the feature, the practical takeaway from the review is to delete a customized version once you’re done with it rather than switching back and forth between “Original” and “Your Version,” since the chat doesn’t always announce which one it’s answering from.
Simple rule: when in doubt, check the original.
The Times has sued OpenAI and Microsoft over AI training on its news content, a case the companies are contesting. It separately sued Brave on November 3 over a search index Brave sells to AI labs, a dispute that shows the Times treating outside AI use of its work very differently from AI it builds and controls itself. Related: The Times also sued Brave over a paywall setting and the search index Brave sells to AI labs.
Note on sources: the nytimes.com page itself returned an access error on a direct fetch; the announcement text above is quoted from the DNYUZ syndicated reprint of Ko’s post, which carries the same byline and text.
Sources