The most honest line in NYT Cooking’s new AI experiment is the label on the version WIRED’s Joe Ray actually cooked, not the box that invites you to change dinner: “Joseph’s version made with AI. Not tested or endorsed by the original recipe’s creators.”
That disclaimer is a product strategy wearing an apron. A brand built on tested recipes is letting a model touch the work, and telling you the human cook is still the thing you paid for.
That’s the business lesson under the sauce. Put AI automation on the chores. Keep a named person as the thing a summary engine cannot lift. The rest of this launch is a test of whether a premium product can hold that line.
A thread, an off-switch, and a values argument
The Times announced the modifier on October 23, 2026, in a Times Insider piece. In the Cooking iOS app, a “customize this recipe” field sits just above the ingredient list. Edits show up in red-orange, including changed words in the title, and you can flip back to the original. Ray reviewed it on November 8 as an iOS-only beta. In the announcement thread, the company’s line for everywhere else was “soon.” Nobody has pinned down a date for Android or the web.
The comments did not go quietly. Ray counted 172 over four days, almost all of them angry, including threats to cancel. Then commenting stopped. Readers wanted to know who had asked for this. They also objected to data centers and the energy cost of adopting AI at all. That second grievance is easy to treat as atmosphere. It is an adoption risk. Some subscribers treated the feature as a values problem, not a button they could argue with.
Jordan Cohen, an executive communications director at the Times, told The Verge a different institutional story. Comments “remained open for four days, with editors from Cooking taking care to answer questions, address concerns and engage directly with readers.” Shutting older threads, he said, is “standard practice at The Times so that our moderators can focus on new discussions.” Many Cooking members, he added, were already trying the tool on iOS and “reporting good results.”
Both accounts can be partly true. A loud thread is not a cancellation study, and “good results” is not an error rate. The Times does not break Cooking subscribers out of its wider digital business. Digiday has cited a Times digital base of 11.88 million without dating that figure to this launch, so nobody outside the building can say whether 172 comments were a revolt or a loud corner of the room.
Hide and disable are not the same promise, and nothing public shows a hide control already live. In the announcement thread, a Times respondent identified only as Ko told a reader there was no way to disable the feature “at this time, ” and that using it was optional. By Ray’s November 8 review, Camilla Velasquez, general manager of NYT Cooking, said users would be able to hide it. Different people, different moments, different words. If you ship software, that gap is the fact. Do not sand it down.
A field, not an agent
The product is a single request field that rewrites the method. The box looks like an open assistant. It is not one. Coverage of the rollout describes a chatbot you have to agree to terms before using. Then the person who runs Cooking tells you not to treat it like company.
“You shouldn’t be chatting with it.”
That is Velasquez. When the industry sells AI agents, it ships chat. This product is trying to be a bounded helper. The tension is the point.
This was not Cooking’s first quiet use of AI. In that same announcement thread, Ko put metric volume-to-weight conversion about two years before the October 2026 launch, and a halve-or-double scaler the year after that. Relative to the announcement, that points roughly at 2024 and 2025, but those are inferences from a comment, not a dated product history. They are also chores. Cups to grams does not threaten the relationship between Melissa Clark and a reader. A box that rewrites her method does.
The public tool is aimed at the requests recipe sites have always fielded: an ingredient swap, a different pan, a dietary constraint, or a recipe scaled to the oil on hand. Velasquez told Ray that current use clusters there. In one example from the announcement thread, asking for baby spinach instead of kale pushed the greens later in the method, because spinach wilts faster.
Ray reported the lookup order the Times described: the recipe itself, then the Times’ ingredient substitution guide, then user comments, then what the company calls AI “drawing on our expertise.” The vendor, the training setup, and what that last phrase actually contains are unpublished. Recipe editors, he reported, helped write the original rules and later checked how they behaved.
Guardrails are part of the pitch. “We put a lot of effort into trying to tell you when not to do something if it’s no longer in the spirit of the recipe, ” Velasquez said. Ray hit that wall on purpose. Asked to turn Eric Kim’s gochujang buttered noodles into a cookie, the system refused and pointed him at Kim’s gochujang caramel cookies. A hard redirect to a real recipe is a better failure than a confident, inedible invention.
Customized versions stay private. Reporting on the feature says they cannot be published or shared in the app. Easy to miss, and it matters. It is a personal overlay, with a path back to the dish a human signed, not a second, machine-written cookbook beside the tested canon.
The miss that should worry you
Ray expected to hate this. He has spent years watching “smart” kitchen gear miss jobs a cheap toaster already understands. His verdict, after several real cooks, splits cleanly: useful as a constrained editor, not an improvement on the signed recipe.
The miss is the evidence. On Alison Roman’s chicken and dumplings (five stars, more than 6, 000 ratings), he swapped bone-in, skin-on thighs for boneless, skinless breasts. The tool shortened the simmer and added butter to the roux, on the reasonable theory that breasts render less fat. It then had him return the shredded breast to the pot for about another 30 minutes. That stretch, he wrote, was the largest problem in his testing: less succulent than the recipe deserved. He also wondered whether that swap should have been refused rather than accommodated. The Times has not said it will change the case.
Changing a finished recipe is a subtractive process. It is good at constraints (the ingredient you have, the diet you keep, the pan you own). It is a poor candidate for beating Melissa Clark at being Melissa Clark. Nobody has published how often these versions are cooked, rated, or pulled. Ray’s notebook is not that study.
Velasquez told him the newsroom has been “hiring more than ever, ” with “zero intention of publishing fewer recipes.” Hold that as what it is: a general manager’s statement to a reporter, not an audited headcount. Ray also noted that his own food-writing career has already been knocked around by AI. A brand can keep hiring recipe talent and still sit inside an industry squeezing writers. Both fit in the same quarter.
What a summary cannot steal
NYT Cooking is running two bets that look contradictory until you see them as a portfolio. AI goes inside the recipe. People get the spend, because a naked method card is exactly what a summary engine can lift.
Velasquez has described video, imagery, and human connection as “becoming the language of the internet.” The line is from Digiday, in a piece about audience growth, not about this bot. Digiday reported that in 2024 people watched over 4.3 million hours of NYT Cooking video on YouTube, according to the Times, and that the channel had more than a million subscribers. In that same look at Cooking’s audience strategy, Digiday reported that 70 percent of subscribers return to Cooking week over week. That is a claim about repeat use, and it is an extraordinary number if “return week over week” means what it says. It is not a satisfaction score for this beta, and the report is not dated to the launch. Series such as “The Veggie” and “Cooking 101” are part of the bet: a person, not a method card. So is Clark on camera, asking viewers what they do with kale stems.
Melissa Chowning, founder and CEO of Twenty-First Digital, told Digiday the strategic version of that bet. Pair the recipe with a personality and the experience is “far more difficult for AI summaries to replicate” than a recipe alone. Steve Paine of Sistrix, looking at keywords Cooking ranked for in Sistrix’s dataset, found Google AI Overviews triggered on about 11 percent of them. He called that rate average. The reporting does not say average against which other publishers, so do not carry “average” into a board deck as if it were a benchmark. The threat is ordinary. Ordinary disruption is how subscription products get nibbled.
The companies that get this wrong bolt an open chat agent onto a trusted workflow and call the engagement a win. The sturdier pattern is duller: do the unit conversions first, show every change, keep the original one tap away, refuse the request that leaves the point of the product, and never let a private overlay pass as a new public canon.
Readers who recoiled were defending a relationship: tested food, named cooks, a comment section that still felt like a kitchen. America’s Test Kitchen sits in the same category of “we cooked this until it worked.” A bad automated variant does not just ruin Tuesday. It taxes a brand that spent real money learning how long a thigh should simmer. A saved grocery trip is a real product. A subscriber who feels the cook they pay for got quietly cheaper is a real loss. The disclaimer is how you tell those outcomes apart.
Questions worth asking before you copy the pattern
-
What has to stay human?
The signed recipe, the named cook, and the judgment that a change has left the spirit of the dish. Personality and video are the parts a summary engine cannot cheaply copy. The modifier is an overlay on work a person already tested, not a second canon.
-
What is allowed to fail in private?
A substitution that makes one dinner worse, if it stays marked, reversible, and off the public recipe. Ray’s chicken breast in a thigh stew is that kind of miss. A failure the product publishes, or a chat box that becomes the thing subscribers think they bought, is not.
-
What should you publish as an error rate?
How often a modified recipe is cooked, rated, or pulled for a bad result. This launch has not published that. Early users “reporting good results, ” and a four-day comment thread, do not size the error. Engagement is the wrong denominator.
-
Did the versions work?
In Ray’s kitchen, as a constrained editor: mostly, with a clear texture failure when boneless breast stood in for bone-in thighs and went back into the pot for about another 30 minutes. One careful review is not a success rate, and he did not find the model improved on the cooks.
-
Is recipe staff being cut to pay for it?
Velasquez told WIRED that NYT Cooking has been “hiring more than ever” and has “zero intention of publishing fewer recipes.” That claim is not independently verified, and it does not erase wider pressure on food writing.
-
When does it leave the iPhone?
It is on iOS for Cooking subscribers now, and Ray reviewed it as a beta. The public timing for everywhere else is “soon, ” with no date for Android or the web.