Over the last few weeks, I’ve been using GPT-4 for cooking assistance. If I need a substitute for an ingredient I forgot to buy, GPT can suggest alternatives. If I want to clear out my cupboards, I simply type in the ingredients I have, like “Please create a recipe using two eggs, a jar of borlotti beans, a potato, a leek, and the scrapings on the bottom of a jar of pickle.” GPT is always polite and thinks for a moment before providing me with instructions for an unusual but edible hash, even wishing me bon appétit. But GPT can do more than that.
During a recent trip to Venice, I wanted to know what kind of fish I should try. GPT explained that one of the local specialty fish is the branzino, or European sea bass, without mocking my ignorance. Just as image generation models can mimic popular artists, GPT can also incorporate the influence of well-known chefs. For example, when I asked for a “beans on toast” recipe in the style of Yotam Ottolenghi, it gave me a recipe for “spiced beans on sourdough toast” that included cumin, za’atar, Greek yogurt, and other ingredients. However, the long list of ingredients was overwhelming, and when I asked for a simpler version, GPT provided alternatives in the style of Jamie Oliver, Martha Stewart, and Salt Bae.
Not all of GPT’s results are groundbreaking. This became evident when YouTube chef Joshua Weissman challenged GPT to a cook-off. His taste testers found that the formulaic recipes for burgers, fried chicken, and chocolate chip cookies didn’t match Weissman’s creations. However, I believe that the formulaic choice of dishes may have been part of the problem, which AI could potentially help us overcome.
Of course, there are challenges with AI models. Some criticize the output for being too “white and western.” There have also been controversies, such as a chatbot designed for eating disorders offering weight loss advice. There’s always a risk that the machine may include ingredients you’re allergic to, and without skilled prompting, the food may turn out bland.
Nonetheless, the convergence of cooking and computation makes sense. Recipes are essentially algorithms, providing instructions to solve the problem of what to cook and how. Finnish engineer Sami Matilainen visualized Nordic classics as flowcharts in his book, “Flowchart Recipes,” highlighting the similarity between recipes and programs. Additionally, the programming language Chef, created by cartoonist-physicist David Morgan-Mar, uses a cookery-inspired syntax to write computer programs.
Instead of asking GPT for versions of dishes I already cook, I decided to explore something new. I love both Iranian and Sichuan Chinese cuisines, so I asked GPT to combine them. At first, it suggested “kung pao jeweled rice,” which seemed uninspiring – simply placing two classics side by side on the same plate. However, when I asked GPT to try again, it proposed “mapo bademjan,” a fusion of spicy mapo tofu and hearty khoresh bademjan. The result was a lasagna-like dish with layers of tongue-numbing Chinese sauce and fried aubergines from khoresh. It was hot and slimy, but the khoresh added a vegetal density that the mapo alone couldn’t provide.
I may be the first person to have cooked this dish, but the recipe now exists on OpenAI’s servers, and I’m confident that a platform will emerge to record AI-human collaborations in the future. Most recipes no longer exist, either forgotten or never written down. With large language models capable of recording and recalling information, the opportunity for accidental discoveries, experiments, family recipes, and tricks to endure and thrive will increase.
The tradition of written recipes dates back around 4,000 years when Mesopotamians began inscribing simple ingredient lists on clay tablets. The word “recipe,” meaning “take,” originated from these tablets and later translated into Latin. In the late Middle Ages, not much had changed, as it was assumed that anyone reading a recipe would know how to use it. However, by the 18th century, with the affordability of printing and the rise of standardized measurements, home economics books filled with elaborate recipes and detailed instructions became popular.
Artificial intelligence brings cooking back to an oral tradition, allowing for more descriptive adjectives and nuanced responses to food beyond bratty comments and star ratings. AI can convert measurements and remind you of the ideal temperature for serving pork loin. Moreover, it can reintroduce personalization and details to cookbooks, constantly refreshing and learning. Perhaps we are on the verge of redefining what it means to cook, working in kitchens that listen, record all the adjustments, experiments, fusions, and improvisations for others to access. This expands the culinary experience from the ground up.
Philip Maughan is a writer based in London and Berlin, and Cooking Earth by Black Almanac is now available.