NYT Cooking AI Features: What They Do and Why Readers Object

The app already had other AI-assisted tools

This isn't NYT Cooking's first brush with AI. Earlier features helped with conversions between volume and weight - like translating a cup of bok choy into grams - and with scaling recipes up or down. Those tasks are more constrained: the input is fairly clear, and the expected output is pretty direct.

Recipe modification is broader. "Make this serve six" is often mostly about quantities, but cooking time and pan size can change too. "Replace this ingredient" can involve flavor, texture, moisture, cooking time, allergies, and the role the ingredient plays in the dish. Some swaps are straightforward. Others quietly change what the recipe is.

A gradual launch is also a product test

Genevieve Ko, a senior editor and columnist at NYT Cooking, said the new tool is rolling out gradually so the team can monitor performance - response speed, for example - and learn how people use it. At the time of the announcement, it was available to Cooking subscribers through the iOS app, with broader access coming later.

Ko also said the Cooking team spent time reviewing sample answers so the assistant's responses would match what culinary experts would advise. That editorial work matters, but it doesn't turn the feature into a guarantee. It's about setting expectations and checking outputs - not ensuring every AI-generated adjustment will work the way you want in every kitchen.

For cooks, the takeaway is simple: this is a suggested adaptation, not a newly tested recipe. If the change affects food safety, a serious allergy, or a technique you can't easily undo, you're going to want to scrutinize the answer more than you would for a routine conversion.

The feature may solve a problem users do not have

One recurring reaction was that ingredient substitutions don't need AI. People have adapted recipes through experience, cookbooks, search, and plain old common sense for years. If the tool feels like a technology layer added onto something people already know how to do, it can come off as extra complexity rather than help.

There's also the question of what people think they're paying for. Some commenters said they subscribe because they want work created and edited by humans. For them, the issue isn't whether an AI suggestion is sometimes correct. It's whether AI belongs in a product they associate with human editorial judgment.

Trust and control are part of the feature

Ko addressed concerns readers raised about AI's environmental impact, the concentration of power among AI companies, and broader risks tied to the technology. Those worries go beyond any single cooking workflow. And a feature can be optional to use while still changing how customers feel about the service that introduced it.

That's where the control question shows up. Readers asked for a way to turn the feature off. Ko said there was no setting to disable it at the time, even while noting that using the feature was optional. In practice, both can be true. But from a user's point of view, "I can ignore it" doesn't always feel the same as "I can choose not to have it in my product."

The model can be wrong in a way that looks plausible

Another objection focused on quality. A confident substitution can still lead to a disappointing dish if it misses how an ingredient functions - or how changing timing and moisture affects the outcome. The answer may sound reasonable, and still be hard for a less experienced cook to catch before dinner is on the table.

That's a real trust problem: you may not find out the recommendation is wrong until you've already acted on it. The more a feature presents itself like expert guidance, the more important it is to show uncertainty and be clear about the limits of the advice.

Treat substitution as a spectrum, not one task

If you're evaluating a substitution assistant, a useful approach is to think in terms of risk and complexity:

• Low-complexity swaps: Swapping one similar leafy green for another, with a note about cooking time. This is often helpful when the recipe method stays mostly the same.
• Recipe-structure changes: Swapping a key ingredient that affects moisture, fat, acidity, or texture. Here, the assistant may need to adjust other quantities - or at least make it clear the result won't be the same.
• Dietary restrictions: Removing gluten, dairy, or another ingredient can involve several linked substitutions, not a single swap. The answer should explain what changed and how the recipe might behave differently.
• Safety-sensitive requests: Allergy advice, food handling, and uncertain ingredient substitutions shouldn't be treated like casual improvisation. The product needs clear boundaries and a cautious tone.

This isn't saying every suggestion will fall neatly into the buckets above. It's just a way to match the amount of trust the feature asks for with the consequences of being wrong.

Preserve the learning path as well as the shortcut

A good cooking assistant should explain why a change makes sense when that explanation helps. "Add spinach later because it wilts faster" gives the cook both an instruction and a useful bit of technique. A plain swap might be faster, but it leaves you without a way to judge whether the advice fits your ingredients - or your preferences.

At the same time, explanation shouldn't turn every request into a lecture. If someone just needs a quick adaptation, they should get one. The product challenge is to provide enough reasoning to support trust without turning a simple decision about dinner into a long back-and-forth.

So the best version of this use case isn't "AI makes cooking easier." It's more specific: AI can reduce the effort of adapting a recipe when the user wants help - while keeping the original recipe and the cook's judgment intact.

1. Define the job before choosing the technology

Start with the user's actual problem. Are they trying to use ingredients they already have? Accommodate a dietary restriction? Avoid arithmetic? Learn why a recipe works? Those needs overlap, but they aren't interchangeable.

Then compare the AI approach to simpler options. A conversion table, a tested substitution guide, or a well-designed set of recipe filters might solve a narrower problem more predictably. AI makes sense when the range of requests is too varied for static rules - and when the flexibility is worth both the uncertainty and the operational cost.

2. Test consequences, not just fluent answers

A response that sounds good isn't automatically one that cooks well. Testing should use real recipe scenarios and judge the outcome against culinary criteria: texture, timing, flavor balance, and whether the instructions still make sense after a change.

Include easy cases, edge cases, and situations where the right move is to ask a follow-up question or admit uncertainty. Editorial review can help, but teams also need a way to spot recurring failure patterns after launch - and to update guidance as those failures show up.

3. Make user control explicit

Optional use is a meaningful choice, but it isn't the only control users might want. Teams should separate opting out of an AI interaction from hiding or disabling the feature in the interface. Whether to offer both depends on the product, but it's the kind of decision you want to think through before launch, not hear about later in comment threads.

The interface should also make it clear what the assistant changed and what it didn't verify. If it proposes an ingredient swap, users should be able to see that the original recipe remains the tested baseline and the updated version is a suggestion.

4. Measure trust and retention alongside engagement

High usage doesn't automatically mean high value. A tool can drive curiosity clicks while weakening confidence in the editorial product. Teams should look at repeat use, user feedback, abandonment, support requests, and whether people understand the status of the modifications the assistant suggests.

And track who benefits. A substitution tool may be especially valuable to people managing dietary needs or working with limited ingredients - even if many experienced cooks don't want it. The goal isn't to satisfy everyone with one default experience. It's to make the feature useful for the people it's meant to serve without making the broader promise of the product feel less credible.