Teresa J Blair has a handful of cookbooks to her name. Each has a snappy, catchy title: The Ultimate Crockpot Cookbook for Beginners: A Comprehensive Guide to Slow Cooking Success for Novice Chefs, Featuring Mouthwatering Recipes, Time-Saving Tips, and Essential Techniques is a great example. Teresa has written about canning and preserving, the Mediterranean diet and anti-inflammatory recipes. Well done, Teresa. But don’t expect this cookery writer to turn up to any book signings or appear on Saturday Kitchen. It’s not that Teresa is a diva; it’s that she’s not a real person. She’s an AI.
In the past couple of years, something disconcerting has been happening in the cookbook marketplace. Entire books are springing up, written by AI. In just over a week, between 2 and 14 December 2023, Teresa managed to publish four books – which means Nigel Slater will need to speed up his whisking significantly. She has a rival, as well: in January, the fictional Lillian D Stewart published five recipe books on pasta, gout food, lectin food, Mediterranean food and the dash diet. At first glance, Teresa’s books look plausible. Some people may be fooled by them. But can AI actually emulate human chefs?
In 2023, Joanne Lee Molinaro noticed that, among the infuriating rip-offs of her Korean Vegan Cookbook, there was one culprit that looked almost identical to her book. Written by one “Rachael Issy”, this Korean vegan cookbook had a cover so similar that it appeared to be actually trying to deceive readers. “I knew that AI could also be used to exploit the work of others,” Molinaro wrote on her website, “that its growing utility threatened human thought and labor with their obsolescence. But it never occurred to me that ChapGPT could cut so close to my bones. And man … it hurt.”
Molinaro’s many followers confirmed to her that this supposedly vegan cookbook boasted recipes featuring eggs and dairy, and even at one point suggested chicken. To head off the implication that she had written it, Molinaro threatened legal action and her followers bombarded the book’s (now deleted) Amazon page with negative reviews. “Make no mistake,” Molinaro wrote about the experience. “AI will thrive in a playground where productivity and fat margins are valued above creativity and equity.”
Molinaro’s example might be an anomaly at the moment. But there are people out there working on harnessing AI for food writing. One of them is Paolo Rosson, the author of The AI Cookbook. Ominously subtitled A Taste of the Future, Rosson’s book, released in 2022, was an experiment in which he helped GPT-3 to produce AI recipes. The book’s photos, made with DALL-E, are wholly convincing. Rosson corrected some of the most glaring mistakes in the recipes – one called for two kilos of butter, for example – but he let the model do as much of the work as possible.
I cook one of the book’s bizarre “fusion dishes”, flicking past “shepherd’s pie sushi” and “Thai green curry lasagna” to get to “Sri Lankan currywurst curry”. The eloquence of AI models has improved in the past two years, but it’s no surprise that the prose won’t win a Pulitzer: “Sri Lankan currywurst is a flavourful and hearty dish that makes a great main course for lunch or dinner. It pairs well with white rice and a side of green beans or another vegetable.”
I choose this dish because, like a lot of AI, it makes me feel slightly queasy. I suspect that when I serve it to my family, they will order a takeaway or head to the pub. But, miraculously, the dish works. “This was actually one of my favourite recipes from the book,” says Rosson, who was similarly surprised. “I remember that, at the very beginning, when I was still fine-tuning the creation process, the AI recommended the use of bratwurst ice-cream as a side dish, which was a bit much for me.”
Rosson has hit the nail on the head there: because it has never tasted the food, the AI has no idea whether what it suggests will taste good or not. Surely it can only reproduce recipes that already exist? “I wouldn’t trust an AI-generated tabbouleh recipe in the way that I would a Felicity Cloake ‘Perfect…’ recipe,” says bestselling food writer Rukmini Iyer. “An AI hasn’t cooked and tested six recipes and compared them before coming up with its own.”
Having bought two of Teresa J Blair’s masterpieces, I am keen to see the inside of a pure AI book. Unfortunately, there are zero photos, just a plain list of hundreds of recipes. The Ultimate Crockpot Cookbook for Beginners is an unconventional crockpot cookbook, I discover, in that 99% of the recipes are for mojitos – not a drink known for being made with the aid of a crockpot. I make one of the 100 recipes on offer, called Gimme My Mojito!
Although the formatting is a little odd – “Ingredients5”; “3fresh mint sprigs” – it’s perfectly lovely. I later discover that every single word is copied from this mojito recipe. Because you can’t copyright a formula but can copyright the specific wording of a recipe, says Ryan Abbott, the author of The Reasonable Robot: Artificial Intelligence and the Law, this means that Teresa has probably committed copyright infringement.
I have an even better time with Teresa’s The Ultimate Anti-Inflammatory Cookbook for Beginners. Here I am reminded why proofreaders exist. Something in the AI processing for this book took objection to the word “and”, turning it into “&;” in every instance. It inadvertently leads to beautiful phrases such as “h&ful cori&der” and “using an immersion blender or even by “h&”. We know that AI struggles with hands, but this is ridiculous.
The Japanese hotpot I attempt – not obviously anti-inflammatory, like all the other recipes – is one of the most disgusting meals I have ever eaten. Even if we ignore the gibberish instructions – “Sprinkle a little of the chilli-ginger-garlic-sesame combine true between everylayer”; “Give it a two-min simmer once it reveryes a boil before turning it” – the whole thing is a soggy, limey mess. My children barely touch it and I can see betrayal in their eyes. Thanks, Teresa.
Though AI recipes can be dreadful, that won’t stop them making people a quick buck. As Matthew Kupfer pointed out in a fantastic Twitter thread on another AI crockpot book, these cookbooks can climb the Amazon rankings because the reviews themselves are written by AI, a trend becoming increasingly prevalent online. (Teresa has only ever been reviewed twice – the review is exactly the same, despite being written for two different books, and was clearly written by AI.)
You can find gibberish AI recipes on YouTube as well. One channel, SuperRecipess, has 1.19 million subscribers despite being driven by AI and despite its videos being called things like “I never bought ice-cream again, I only make it like this now” and “This recipe surprised everyone for how easy and delicious breakfast can be!” While the footage itself seems real, the commentary – “And if you’re still watching this video, you’re loving our recipe!” – is almost certainly generated by AI, and the recipes are often extraordinarily disgusting.
The law is playing catch-up with the rate at which AI is accelerating. But if an AI cookbook is trained on recipes that it has hoovered up from the internet and elsewhere, this might not be illegal in and of itself. “If you were writing a cookbook, you might buy other cookbooks and read them and draw inspiration from them,” Abbott says. “Training on copyrighted material has never constituted copyright infringement for a person.” Everything rests on how the word “training” is defined.
Though AI cookbooks might not be hugely impressive at the moment, cook and food writer Ravneet Gill points to a dystopian future in which publishers might want 10 books on air fryer recipes generated quickly. Rather than paying an author between £30,000 and £100,000 to do so, they might simply use AI and pay a popular food writer a £10,000 endorsement fee. “It’s very dangerous,” she says. “But unfortunately it’s not going away.”
As in other creative spheres of life, however, will we be happy to be dictated to by robots? Even Rosson, who wrote an entire book with AI, is sceptical. AI might be able to reproduce his grandmother’s cake recipe perfectly, but would the pleasure of eating it be as great? “If I knew that a recipe was created by AI – even if I knew that it was meant to recreate my grandmother’s cake – it would mean less to me,” he says. “Because underneath, I know that it hasn’t been made by a real person.”