The New York Times is rolling out an AI tool in its Cooking app that lets subscribers modify recipes, and the reaction shows how touchy AI features are inside paid, trust-based products. The Verge reported that the app is testing AI recipe modifications and that it has "sparked some drama among users." Nieman Lab's headline says users are not thrilled.
We read the full Verge report but could not open the Nieman Lab article, so for that outlet this post relies on its headline and the Times quote it carries, plus an AI-training news summary of the launch and the public Hacker News discussion. Where a detail comes from a secondary summary we say so.
TL;DR: the questions people ask first
| Question | Answer |
|---|---|
| What does it do? | Applies modifications to NYT recipes, mainly ingredient substitutions and dietary changes |
| Are recipes AI-written? | No; base recipes stay human-written and tested, per reporting and commenters |
| How is output labeled? | Summaries say results are framed as generated, not editorially tested |
| Who has it? | Per The Verge, all Cooking subscribers on iOS, with wider release "soon" |
| Earlier AI features | Converting volume to weight and scaling servings |
| Model or vendor | Not found in our sources |
| Reaction | Many commenting subscribers object; others call it useful |
What exactly launched?
According to The Verge, NYT Cooking senior editor Genevieve Ko announced the feature on the Friday before publication. It can recommend substitutions for ingredients and cooking equipment, scale recipes to any number of servings, and "generally make our recipes work for what a cook needs at that moment." Customized recipes are private and cannot be published or shared through the app. The Verge says the tool adds a chatbot-style text box at the bottom of recipes and asks you to agree to terms and conditions first. It is being rolled out gradually while the Times monitors how it performs: it is currently available to all Cooking subscribers on iOS, with broader availability coming "soon." A separate summary of the launch says the tool lets subscribers make dietary changes, and that results are labeled to separate them from editorially tested ones; the Verge excerpt we read does not mention that label. Nieman Lab, via a quote in the search summary we could access, has the Times saying it has "just launched our latest tool, which gives cooks the ability to apply modifications to our recipes." The Verge says Ko described earlier AI use in the app: converting imperial volume measurements to metric weights two years ago, and a feature last year that halves or doubles a recipe's servings.
The Verge's example is a user asking to use baby spinach instead of kale; the tool advises that because baby spinach wilts faster, it should be added in a later step. (A Hacker News commenter quoted a similar chard-to-spinach example from the article.) That is the kind of question the tool is meant to answer, and it is a sensible use: a modification is a small, well-bounded task with a vetted recipe as context.
Why is it controversial?
The concern is not that a chatbot exists; it is where it sits. People subscribe to NYT Cooking for recipes that staff have developed and tested. The Verge reports that readers pushed back in the announcement's comments, citing the environmental impact of widespread AI and its effect on the user experience; one asked whether the feature can be disabled, and Ko answered "No, not at this time," adding it is "completely optional to use." Another reader called it "unbelievably disappointing." Times spokesperson Jordan Cohen told The Verge that comments stayed open four days with Cooking editors answering questions, and that many members of the community are trying the tool on iOS and "reporting good results" (a company statement). One subscriber quote, repeated in the HN thread from the article, says the internet is full of AI cooking content and they pay the Times because they prefer theirs done by people. That fear connects to a broader pattern: Bloomberg reporting relayed by Fortune last year described food creators complaining that AI-generated recipe "slop" was distorting how people find cooking advice online.
Others in the thread argue the opposite. Several describe using ChatGPT or Claude to adjust recipes already, and say a built-in tool anchored to a vetted base recipe is better than pasting text into a chatbot. Some note that vocal commenters are power users and not the whole subscriber base. Another commenter worries about hallucinated ingredients or unsafe instructions, which is a real concern for food. These are individual opinions, not measurements of how subscribers as a group feel.
What developers and users are saying
Reading the thread, four views stand out, each attributed to the commenter and not verified.
- One commenter, tristanj, argued that large engineering teams in the AI era have spare capacity for pet projects few users request. That is speculation about motives, not reporting.
- huitzitziltzin called requests to turn the feature off exaggerated because it is optional and only matters if you change a recipe.
- spl757 said AI can hallucinate ingredients or unsafe preparation instructions, and argued it is being pushed too hard and fast.
- JumpCrisscross said they already use AI to adjust recipes but choose the model and give it parameters, so a site-specific tool would not suit their habits.
The disagreement itself is the useful signal: optional and bounded features still trigger a values reaction when the product's brand is human craft.
How does this compare with other publishers?
Publishers are splitting into two camps: those licensing content to AI companies and those building AI features into their own products. We covered how OpenAI's partner-publisher pipeline works, how Canadian news publishers sued Anthropic and how publishers are weighing blocking crawlers as search traffic drops. The NYT Cooking tool is the third path: using AI inside the product as a utility, while keeping the core content human.
For the broader consumer side of AI and food, our guide to AI for cooking, nutrition and meal planning covers what general chatbots do well and badly, and the social-platform side of the AI label debate appears in our post on Adam Mosseri's comments about AI negativity and slop labels.
Why recipes are a hard test for AI assistance
Cooking looks like an easy domain for a language model because recipes are everywhere in training data. It is harder than it appears for three reasons.
Chemistry hides in the details. Swapping spinach for chard is forgiving; swapping baking powder for baking soda, or butter for oil in a cake, changes structure, browning and rise. A model can give a fluent answer that ignores the reaction underneath, and the failure only shows up after an hour of cooking and a wasted meal.
Context is physical. Pan size, oven calibration, altitude and ingredient moisture all change results, and none of them appear in the text of a recipe. Human recipe testers cover these by repeating a dish many times. A modification tool has no equivalent feedback.
Safety matters. Food has allergens, undercooked-meat risks and storage rules. A hallucinated instruction here is not just embarrassing. That is why commenters raised hallucination as a concern, and why a visible label separating generated output from tested output is the minimum responsible design.
On the other side, the use case is genuinely common. Home cooks substitute ingredients constantly, and a quick, grounded answer beats searching forums. The question is not whether the feature can be useful but whether the interface makes its limits obvious at the moment of use, for example by flagging when a substitution is likely to change texture or when baking chemistry is involved.
What good design looks like here
If you build a similar feature, the NYT approach contains some reasonable guardrails, as far as we can tell from the reports:
- Ground the model in a vetted source. A modification starts from a tested recipe, not from nothing.
- Label generated output clearly, so a reader can tell tested from generated.
- Keep it optional to use. The Times says it is optional but, for now, cannot be switched off; consider offering a real off switch.
- Limit the task. Substitution and scaling are narrower than writing a recipe from scratch.
- Add a safety layer for food-specific risks, such as allergens, cooking temperatures and food-safety timings. We found no statement about how the Times handles these.
What we could not verify is how well it works. Substituting an ingredient changes chemistry as well as flavor, especially in baking, and a plausible answer can still ruin a dish. Anyone using the tool should treat it like advice from a knowledgeable friend, and check allergens and safe temperatures independently.
How to use a recipe-modification AI safely
Whether you use the Times tool or a general chatbot, a short routine reduces the risk of a bad result.
- Start from a tested recipe and ask for one change at a time, so you can tell which change caused a problem.
- Ask what will change, not only what to swap: texture, cooking time, moisture and seasoning.
- Be careful in baking, where ratios matter, and prefer swaps that the original recipe's author or a trusted source also recommends.
- Check allergens yourself. Never rely on an AI answer to confirm a dish is safe for someone with an allergy.
- Verify temperatures and times for meat, poultry and eggs against an official food-safety source.
- Note what worked, so the next change starts from experience rather than a fresh guess.
These habits apply to any assistant that edits trusted material, which is why the label distinguishing generated from tested content matters so much.
What is not confirmed
- The model and vendor behind the tool.
- The timing of wider availability beyond iOS ("soon," per the Times).
- How the Times tests accuracy or handles allergen and safety questions.
- How many subscribers object. The reports describe vocal comments, not a survey.
What this means for what you build
If you run a subscription product built on human expertise, expect AI features to be judged on trust before utility. Label clearly, explain the grounding, and make the opt-out easy. If you are a reader, the practical takeaway is simple: ask for a substitution, then sanity-check it against what you know about the dish.
Related reading
- AI for cooking, nutrition and meal planning
- ChatGPT partner publisher content pipeline
- Canadian publishers sue Anthropic
- Google Zero and publishers blocking crawlers
- Mosseri on AI negativity and slop labels
- Spotify removes AI slop tracks
- Sources: The Verge and Nieman Lab
Details are accurate as of October 29, 2026 and rely on secondary reporting; we will update once the Times' own announcement is confirmed.