The rise and rise of the chatbot cook

In the bygone days of May 2024, it did not seem as though AI had a future in the culinary sphere. In a now-immortalised screenshot posted on X, Google’s newly released “AI Overview” tool responded to the search query “cheese not sticking to pizza” by suggesting that the user add “about ⅛ cup of non-toxic glue” to the sauce to make it stick. That same month, it told another person to eat “one small rock per day”, and advised a third that petrol could be used to make “spicy spaghetti”.

Two years later, generative AI has been rebranded as the ultimate kitchen companion. Google now markets its AI assistant, Gemini, as a tool you can talk to while cooking, promising “expert advice and personalised help”. OpenAI, meanwhile, has advertised ChatGPT as “your new sous chef, sommelier and creative partner”.

For many home cooks, the technology has been a boon. People are using AI chatbots to do everything from planning weeknight meals around their budgets to generating shopping lists from what’s already in the fridge, making cooking easier and more accessible. But for professional recipe developers like Georgina Hayden, the change is not necessarily welcome. “If it ends up being really good at recipe writing, I’ll be gutted,” she tells me as I walk into her home in north London. “And also, I’m out of a job.”

When we meet in December 2025, Hayden is in the final stages of developing recipes for her next cookbook, Medesque. A follow-up to the bestselling Greekish, it draws on her Cypriot heritage and years of work researching and writing about food. She has agreed, somewhat reluctantly, to spend the day stress-testing her latest recipes – and developing new ones – with the help of ChatGPT.

Hayden starts by serving me a date, walnut and chocolate cookie, a recipe from her upcoming book inspired by ma’amoul biscuits, shortbread-style cookies eaten across the Middle East. The cookie is, objectively, perfect – chewy without being claggy, sweet, but not overly so. Hayden had made at least 10 other batches, but in each, something was off.

The previous evening, she had entered her recipe into ChatGPT, asking why the cookies tasted “a bit eggy”. The app suggested that she reduce the amount of egg white, and that she increase the ratio of brown sugar, because it “balances flavour and masks bitterness or egginess”.

The result was “impressive”, she says, “and annoying”. It’s also not all that surprising. When researchers have tested AI chatbots on recipe development, they tended to perform best on structured work like error-spotting and adjusting ingredient ratios. “I could have figured that out eventually with lots of testing and tweaking,” Hayden says, “but ChatGPT nailed it”.

There is already an emerging industry dedicated to the idea that cooking is more science than art. Ganesh Bagler is a pioneer of “computational gastronomy”, where his team at the Indraprastha Institute of Information Technology uses AI models to analyse the molecular compounds of food and suggest ingredient pairings based on shared chemical structures. His research culminated in the creation of Ratatouille, a tool trained on a database of over 118,000 recipes from 76 countries, designed to generate recipes on demand, based on criteria such as the type of ingredients you have available to you, or your desired cuisine.

The computational gastronomy world is one of double-blind studies, spreadsheets and algorithms. Bagler’s lab, which does not have an in-house kitchen, relies on partnerships with an Indian branch of Le Cordon Bleu culinary school to test AI-generated recipes for objective “parameters” like aftertaste and mouthfeel. Bagler admits he hasn’t tasted anything life-changing yet. (“Some have surprised me positively. Some have surprised me negatively,” he says. “We are still evaluating the recipes”.) But he does believe, wholeheartedly, that “cooking can be boiled down to the subtle details of actions and quantities, so it can be made into a quantifiable science”.

Hayden is not convinced. While generative AI might nail the “maths and science” that is so central to baking, she believes it lacks the “human element”, the creativity and intuition required to nail a recipe in the kitchen.

We decide to push ChatGPT further. “Please write me a chicken thigh recipe in the style of Georgina Hayden,” we prompt the model. The response is generated in seconds: “There’s something undeniably comforting about a tray of golden, crisp chicken thighs,” it begins. “This recipe is simple, the kind of thing Yiayia would throw together instinctively.”

At first, Hayden’s attention turns to the ingredient list. She can already tell the recipe lacks depth, pairing honey and cinnamon in a way that she thinks will taste too sweet. She would typically balance a dish like this with paprika or chilli to prevent a “strange finish on your palate”.

We agree to give ChatGPT the benefit of the doubt, and test the recipe as is. Then Hayden reads the introduction again. “Why is it talking about my yiayia?”

Large language models like ChatGPT and Gemini are trained on all of the work (or the data) that Hayden has produced throughout her two-decade career: her four cookbooks; the recipes she has written for The Observer, BBC, Guardian and Telegraph; her Instagram page, with its 3,400 posts, many of them deeply personal accounts about learning to cook from her grandmothers in Cyprus. That a chatbot can create a facsimile of her voice in seconds is unsettling, to say the least.

“It just angers me,” she says. “I’ve inherited recipes from the women in my family, and most of them aren’t really around any more.” She says that the beauty – the art – of recipe development is capturing “memory and history and heritage”, and translating that into a tangible narrative.

ChatGPT’s mention of her grandmother reminds her, she says, of the controversy around Makan, a cookbook by the Michelin-starred, British-Singaporean chef Elizabeth Haigh. The book was withdrawn in 2021 after the food writer Sharon Wee accused Haigh of copying from her own book, Growing up in a Nonya Kitchen. Not only were the recipes very similar, but her stories were too – passages about learning to cook from her mother were paraphrased and, in some cases, lifted wholesale from Wee’s book. “We know it’s wrong to steal someone’s recipes and memories, because look what happened there,” she says. When I point out that ChatGPT wasn’t lifting directly from any of her work, instead bizarrely repurposing her old material, she doubles down. “It is stealing, because that’s not me saying it,” she says. “It’s not for them.”

The chicken thigh recipe ends by suggesting that we add “a wedge of lemon if you’re like my mum and want that extra zing”. Hayden recalls that she wrote about her mum’s love of tangy, lip-puckery salads in the introduction to a recipe in one of her previous cookbooks. “It’s obviously read that and taken it and then spat it back out at me,” she says. “It feels quite alarming.”

When the ChatGPT chicken comes out of the oven it is fragrant, the skin on the thighs starting to crisp, the potatoes beneath it soaking up the juices. I’m ready to dig in, but Hayden is having none of it. “No, it’s still very anaemic,” she says, shaking her head at the tray. “People never cook chicken thighs long enough.”

The thighs go back in the oven for another 15 minutes and emerge crisper, browner. “I just think it looks more delicious now,” she says. And she’s right, of course. It is a victory for the human cook.

Bagler, however, might not see it that way. To him, the problem of the pale chicken thighs is a technical one, solvable with more data and better algorithms. “There are reluctant people who don’t want to accept the promise of AI, which is going to change the way we cook,” he says. But in three to five years, as the technology improves, he thinks even the most fervent sceptics will be converted.

“While love and emotion have their own place, they are not the ingredients of cooking,” he tells me. In fact, Bagler would happily encode his own mother’s cooking into a machine, “so wonderful would it be, especially if it can be left back for eternity, so anyone can test it out”.

It is easy to hear this and feel, to use Hayden’s word, gutted. Could a perfectly optimised Ratatouille, or IBM’s Chef Watson recipe tool, or ChatGPT, or Gemini, replace the work of recipe developers entirely? Could it be the death knell for a food industry that has already been upended by social media?

Ruby Tandoh, a writer and chef whose book All Consuming traces how technology has transformed the food industry over the past 75 years, says we must resist “declinism”. “Every single time that the dominant medium of food culture changes, something about the recipes themselves evolves,” she says. “It was possible to be declinist when mass-market cookbooks first came in, in the last century. You could have been declinist about food television, and obviously people have been. There was a lot of panic about [recipe aggregator] sites such as Allrecipes in the early internet days, because people felt like it put at risk the work of actual recipe writers and professionals.” But the food industry has continued to adapt.

Tandoh has always viewed food culture as “inherently iterative”, and she doesn’t object to recipes being freely built upon and shared online. But she is concerned that a culture of openness is being exploited by a handful of big tech corporations, the work of thousands of recipe developers “turned into profit for a very few people at the top” like never before.

The technology companies argue that their tools will support, not replace, human labour, but Tandoh thinks it is inevitable that the market will be transformed. “These kinds of advances really do decrease the quality of, and flatten and homogenise, the middle,” she says. “But excellence will always flourish.” What will survive, she predicts, is the narrative recipe: “Anything that resists being pillageable or decipherable by these platforms.”

Perhaps that is why ChatGPT’s mention of Hayden’s grandmother is so uncomfortable. When so much of what recipe developers do – from deeply personal writing, to planning, to flavour combinations – can be automated, what does it mean to do something only a human can? AI is forcing all of us to consider the true value of our labour.

Andrea Borghini, a food philosopher at the University of Milan, has spent a lifetime studying the intangible value of recipes. When I put to him Bagler’s view that love is not an ingredient, that recipes can be automated away, he laughs heartily. Why does a forged Van Gogh cost less than the original, he asks, if the colours are identical? Because the original contains spontaneity, innovation, a specific unrepeatable moment, the artist’s touch.

“If you want to define it just based on the colours, go ahead,” he says. “But you can’t explain why one costs more than the other.” The same, he argues, applies to cooking. You can make sushi rice in a machine, but it will never taste like something produced by the legendary Japanese chef Jiro Ono, because Jiro’s sushi is inseparable from who he is, his decades of practice, his relationship with the customer sitting across the counter. “The way it’s done, who’s doing it, the emotional aspects – these are part of the definition of the procedure,” he tells me. “I’m sorry, you can’t exclude those.”

Borghini is no AI sceptic, however. A few months ago, his mother passed away, forcing his 83-year-old father to learn how to cook, something he had never tried because of what Borghini calls a classic “Italian division of labour”. After trying various websites without much luck, he turned, of all places, to Meta’s AI chatbot on WhatsApp. It is now his go-to kitchen assistant.

It is in moments like these, when AI is a companion rather than an author, that it is most useful, Borghini says. This, of course, is exactly what the technology companies are selling: a kitchen companion that makes your life easier, a sous chef. For thousands of amateur cooks, the technology’s benefits are clear. Even Michelin-starred chefs have told the New York Times that AI is helping them discover flavour combinations they would never have otherwise considered. But these companions have learned to cook, in part, by reading Hayden’s recipes, her columns, her stories about her grandmother – and, so far, she isn’t seeing any of the benefits herself.

We try one last experiment. Hayden takes a photograph of her incredibly well stocked fridge and asks ChatGPT to generate dinner for her children. The chatbot recognises all of the ingredients, and comes up with a step-by-step recipe for beef tacos with paprika, cumin, garlic and tomato paste, topped with cheese, tomatoes and yoghurt. Putting aside her anxieties, “as a busy working mum, these are brilliant”, she says. Although she might have added a bit more spice.

Illustrations by Lew Pearce