Four years ago, the food reporter Priya Krishna of The Times invited me and my fellow New York Times Cooking columnists to a Thanksgiving feast — with recipes from A.I.
She had asked a precursor to ChatGPT, called GPT-3, for recipes that sounded like ones she would make, and she cooked them exactly as written. I was the most skeptical going in and the most critical taster at the table. Priya is an exceptional recipe developer; the recipes A.I. generated for her were atrocious, and the accompanying descriptions — supposedly in her voice — spewed ridiculously written lies about her life and tastes.
I naïvely assumed A.I. wouldn’t affect my work as a recipe developer and my role at NYT Cooking, where I oversee the editorial side of our recipes and related stories. But as artificial intelligence improved, our product and engineering colleagues saw its potential to help us create features Cooking users wanted.
We’ve just launched our latest tool, which gives cooks the ability to apply modifications to our recipes. We’re rolling it out gradually to monitor how it performs (whether it’s working fast enough, for instance) and to see how people are using it (as of now, we’re getting a lot of requests for ingredient substitutions). The tool is currently available to a large portion of Cooking subscribers on our iOS app, and we plan to make it available more broadly soon.
What We Wanted to Solve
For years, our users have been asking to measure ingredients by weight (in grams), to easily double or halve recipes and to know how to substitute ingredients for the ones they have. We agreed it’d be helpful to provide those features, but couldn’t figure out how. Our amazing community of cooks chimed in with answers in the comments section, but those replies were often buried among charming stories of dishes that led to second dates and very happy dogs.
The first two requests seemed like straightforward math problems. But there was no way my team could have manually converted more than 23,000 recipes, while continuing to produce new ones. That’s where A.I. came in.
Two years ago, NYT Cooking used it to convert our standard U.S. imperial volume measurements to metric weight measurements based on our tested conversions. (That’s right, we did set a cup of parsley on a scale.) I was nervous about what it might get wrong, but the reality is that when I’m calculating conversions in my head, or even on paper, I sometimes get it wrong. (AP Calculus was the last math class I took, and that was a long time ago.)
It was unlikely that cooking with the metric conversions would alter the recipe’s outcome, but scaling recipes might. Last year, we added an A.I. feature that could halve or double recipes, and developing that tool was a harder challenge. I grappled with the details of how to account for half of an egg and the fear that we might let users down. Our recipe editors, developers and testers work hard to make sure each dish works as written. I couldn’t make that promise for automatically scaled recipes, but I realized that the cooks using them could become even better cooks because they would have to pay closer attention in the kitchen and rely more heavily on their senses.
The Challenges We Faced
Once we saw how helpful these metric conversion and scaling features were for home cooks, we began working on a tool to allow users to customize our recipes to suit their needs by offering them substitutes for ingredients based on what they have at home, found on sale or need to avoid for dietary reasons. The appeal is obvious for users and for us: We want to help people cook our recipes and this helps. As thoroughly as we develop, write, test and edit recipes to ensure they come out, we can’t guarantee your store won’t run out of kale. (Yes, this did just happen to me.)
One of my early concerns was the quality of the chatbot answers and the resulting customized recipes. I still wanted to ensure home cooks receive only flawless tested recipes with thoroughly fact-checked information about them.
I do what I do for a living because I like to feed people, in my home and also indirectly by giving others recipes, and I want them to be perfect. So do the other editors, as well as our engineers and product partners. That’s why we spent countless hours evaluating sample after sample to get the chatbot answers to closely align with how we would answer as culinary experts. We applied the same rigor as we do to editing, grounded in Times standards.
I’m very happy with how well the feature works now and I am not worried that A.I. will take my job or diminish what we do, because it can’t cook. Creating recipes requires not only the physical acts of making and tasting, but also thoughtfulness and soul. This tool just helps cooks prepare their own versions of our expert-tested original recipes.
Where We Landed and What It Does
In addition to giving home cooks substitutions for ingredients, the tool can also scale recipes to any number of servings, swap pans and tools, and generally make our recipes work for what a cook needs at that moment. After answering those types of requests in a chat, it incorporates the requested changes — highlighted in orange — into a personalized version of the recipe to make it easy to cook from. Those answers and customized recipes remain private on your account, and can’t be published or shared via Cooking.
We trained our LLM prompt by drawing on our decades of culinary knowledge, and the engineers combined that feedback with The Times’s own source materials, including the recipes themselves, our ingredient substitution guide and our Cooking community’s comments. At the beginning of the training process, I had to take a lot of deep breaths to deal with inaccurate, florid responses. Now, I feel confident about the results 99 percent of the time and am grateful for the engineers’ work to continuously incorporate feedback and improve the model.
We have met two of our biggest editorial goals: The chatbot will redirect customization requests to other recipes on our site that closely match what the user wants. And it does not simply say yes to everything.
Chatbots tend to blindly affirm your ideas and cheer you on. (Sometimes with too many exclamation points!) We trained ours to comply with requests only when we feel confident they’ll work, and to explain how the dish may be affected. That sold-out kale? The chatbot swapped in spinach, but explained that it wilts more quickly, so it added it later in the recipe, which is exactly what I would have suggested. As with all our recipes — and our scaling tool — we want readers to learn more about cooking through each dish.
Why I’m Still Concerned
As well as this feature now works, I still have questions about big-picture issues around A.I.: the impact of data centers and the concentration of power among A.I. companies; the risk of A.I. escaping our control; and the concern that, especially for children, it will harm our ability to think critically, stay focused and find what is real and true in the world.
I worry that people will conflate those A.I. issues with what Cooking is doing. But we designed this tool thoughtfully and only with the purpose of customizing recipes. I take solace in knowing that we’re not replacing original recipes with A.I. ones. I recognize that A.I. can make mistakes. As an editor by training and a detail-oriented person by nature, that’s hard. But we will keep reviewing and improving the feature and we welcome your feedback, too.
(A note: The Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)
Our Hope for the Future
Cooking isn’t just my job, it’s what I love to do even in my time off. It wasn’t until I used this feature in my own life as a home cook, not a tester, that I understood its true purpose and potential.
This past summer, my sister organized a big trip for our mother’s birthday, and my kids took time off from work to join the extended family. They missed my home cooking, so I ended up preparing two meals a day for about a dozen people. Instead of limiting myself to NYT Cooking recipes that serve 12, I scaled a bunch up with the new tool, adding each to my grocery list. I didn’t miss doing all that multiplication or writing out a shopping list, and I was grateful to have extra time to hang out. What I realized then is that even though this A.I. feature can’t guarantee foolproof recipes, it can fulfill immediate, practical needs in the kitchen to actually increase real human connection.
Cooking is about connection — with yourself and with those around you — and it can be aspirational and leisurely, but it’s often just something you have to do. If you have everything you need to make one of our tested recipes, great. You can follow it as written and know you’ll end up with something delicious without having to expend extra mental energy as you cook. If you need to change something about a recipe and don’t have time to scroll through the comments, you can use this feature. You should pay extra attention as you’re cooking, and you’ll become a better cook for it.
This new A.I. tool requires you to be on your phone, but it also enables the tactile physical pleasure of cooking, speeding up the moment when you can put that device away while you eat your very real, very human creation.
The post How, and Why, We’re Adding A.I. Tools to NYT Cooking appeared first on New York Times.
