Not sure what to eat tonight? Artificial intelligence wants to tell you what to cook, but are the recipes it creates any good? We put it to the test.
As you savour your next meal, it’s worth spending a moment reflecting on just what it took to create. Cooking is a uniquely human activity and over the millennia that we have been roasting, baking, frying and boiling our food, we’ve honed it into something special.
Turning a pile of ingredients into a flavoursome dish requires a blend of chemistry, art and instinct that many of us do without thinking. Some may turn to a cookbook for help, and a few simply reach for a takeaway menu.
Regardless, someone, at some point, has had to figure out how to combine a set of ingredients to produce flavours that titillate our tastebuds. Often it involves a bit of trial and error. The best chefs will usually taste-test their creations before serving.
But what happens if the chef doesn’t have tastebuds at all?
We decided to see just how AI-recipe generators measure up in a head-to-head competition with professional chef and cookbook writer Ixta Belfrage. We wanted to see who could create the tastiest fusion-food recipe from handful of pre-selected ingredients. The final plates were then given to a panel of tasters to discern who, or what, created each dish. The results were somewhat surprising.
Warning, this recipe is more than the ingredients listed…
The computation of sensory mechanisms such as taste and smell, present a significant challenge for AI developers. Without the ability to taste, or to smell, flavours need to be reduced to a collection of binary digits for AI algorithims to understand. But as Charles Spence, an experimental psychologist at the University of Oxford in the UK, has discovered in the decades he has spent studying human senses, our experience of food runs even deeper. It is partly a tangle of chemicals interacting with our taste and olfactory receptors, but also our emotional state and other senses play a role. Listening to different types of music, for example, can add what Spence calls “sonic seasoning” to a meal, altering its flavour, while the colour of food or the crockery we use can also affect our experience.
“The complexities of flavour perception are beyond technology, and engage multiple senses, not just taste and smell, but also texture, visual appearance, pungency and mouthfeel,” says Spence.
It’s this plurality of senses that AI must grapple with when attempting to combine ingredients into new, and perhaps most crucially, palatable recipes. “So much of our experience of food depends not merely on chemical structure and ingredients, but on our prior experience,” says Spence. An AI might be able to compute how chemicals will interact together, but it will struggle with how that blends with our personal perception.
Take acclaimed chef Asma Khan, for example, who’s sensory recall of sounds and smells in her childhood kitchen are the basis to all of the recipes she cooks in her Central London restaurant. Very few of her recipes are written down, but memorised from years spent cooking with her mother, grandmothers and aunties. And crucially memories from its own lived experience is something that AI cannot amass.
So perhaps a good place to start is understanding what a recipe actually is, suggests Patrik Engisch, a philosopher and co-founder of research centre The Culinary Mind at the University of Milan, Italy. He argues that recipes are so much more than just “a list of ingredients and a series of instructions” that turn into a dish. “What we cook and how we cook is ‘who we are’ and it is an expression of what we care about,” he says. A chef’s experiences, their cultural heritage, food memories, even the socio-economic context of the time all contribute to their understanding of flavour and ultimately shape the recipes they then create.
AI v the Mind
This article is part of AI v the Mind, a series that aims to explore the limits of cutting-edge AI, and learn a little about how our own brains work along the way. Each article will pit a human expert against an AI tool to probe a different aspect of cognitive ability. Can a machine write a better joke than a professional comedian, or unpick a moral conundrum more elegantly than a philosopher? We hope to find out.
Generative AI models and large language models (LLMs) like the ones used for recipe creation, produce responses using pre-existing data sets, in this case recipes that already exist and have been created most often by a human. They do not currently have the cognitive function to create original recipes with meaning or emotion.
But there have been some promising developments when it comes to AIs creating convincingly human-like recipes. Ganesh Bagler, a computational researcher at the Indraprastha Institute of Information Technology in Delhi, has been attempting to “capture the culinary creativity of humans” in his novel recipe generation algorithm “Ratatouille“. It was trained using more than 118,000 traditional recipes from across 74 countries.
Bagler also devised the Turing Test for Chefs, inspired by the British computer scientist Alan Turing’s work, which sought to determine the ability of a machine to think like a human. Bagler suggests presenting a chef with a recipe chosen at random from a deck of both human- and AI-generated recipes. The chef is then asked to score each recipe from zero to five, zero being an AI recipe, and five for a human-created one.
Bagler found a number of recipes created by his own Ratatouille algorithm – named after the 2007 Disney film about a rat who dreams of being a chef – passed the test, fooling trainee and professional chefs, on “paper” at least. One particularly convincing recipe was for a Thai shoyu burrito. Delicious? We’d only know by following the steps to cook it.
“We are in the middle of devising a strategy for evaluating these recipes by cooking and ‘tasting’ recipes that have passed the Turing Test for Chefs,” says Bagler. “While I am hopeful that these recipes will turn out to be palatable and tasty, in reality, I am keeping my fingers crossed, since taste is a very subjective phenomenon.”
Our tasty Turing test
Next, for our competition we had to decide which AI tool to use. There is an almost inexhaustible list of apps and platforms, all vying to be the perfect cooking companion. Some, like DishGen, allow you to input specific ingredients so you can use what food you have in your fridge and cupboards. Others, like ChefGPT, focus on providing nutritionally complete meal planning for specific dietary needs. Just over half of those I provisionally tested produced an AI generated image to accompany the recipe, with very differing levels of realism.
Eventually I settled for Food Mood, a self proclaimed “experiment” developed by Google, which claims to create “new recipes mixing influences from two cuisines”. Of all the apps I tried, this paired its recipes with the most convincing, and appetising, images.
Having settled on the AI generator, it was now time to pair it with the perfect chef. And no choice seemed more natural than Ixta Belfrage, the queen of fusion food. I first became aware of Belfrage through Yotam Ottolenghi’s food. As co-author of Flavour, released in 2020, Belfrage has been heralded by some as his protege, her unique style of fusion cooking paving the way for her debut cookbook, Mezcla, meaning mix or fusion in Spanish. She credits her Brazilian and Mexican family roots, and upbringing in Italy, as the source for her unique flavour combinations and inspiration behind her recipes.
This was staring to look like the ultimate cook-off.
Trust me, I’m a chef
I’ll admit that I’m a little obsessed with cookbooks. The satisfaction I gain from cooking from a recipe doesn’t just come from tasting the finished dish (although that is a big part of it), but from the supplementary knowledge I absorb about the culture and history behind that particular dish or cuisine. It all helps me trust the method presented and the chef’s intentions with the food.
This idea of trust is an important one. We expect the outcome of our culinary endeavour, notwithstanding our own ability, to be delicious, nutritious and safe to eat. So what about recipes created by a computer?
Researchers at the University of Naples Federico II and the University Oxford have found that images of food generated by AI were deemed more appealing until volunteers learned they had been machine-generated. Although people trusted AI to create relatively simple and familiar recipes, this trust diminished when an AI came up with more creative offerings. It could be that the assumption that AI recipes are unverified and untested dampened their desire to cook the dish.
This is not surprising, given instances where AI models have made up fake ingredients and created toxic dishes like “bleach infused rice surprise”. Without human verification and testing, there is a higher likelihood of recipe generators creating error-laden instructions that could not only be revolting, but also dangerous.
We began the experiment by choosing a handful of pre-selected ingredients as the basis for two dishes – ricotta, oyster mushrooms and chipotle chilies for the main course; chocolate and orange for dessert.
I fancied a Mexican-Italian fusion meal, so I set Belfrage and the AI to work.
Sat at a computer, we generated a number of different recipes, Belfrage assessing their viability and most importantly their propensity for deliciousness. The initial offerings from the AI sounded interesting: partially cooked tomato and zucchini salad; Chipotle Chiaroscuro; and a double carbohydrate penne surprise, Three Wise Mushrooms.
Scrolling through some of the options, it was immediately obvious to Belfrage that there were inconsistencies in the recipes, with mixed metric and imperial measurements, ingredients missing, incomplete instruction and pictures of an end-product that looked wildly different from what the method would produce.
Eventually Belfrage chose two recipes offered up by Food Mood that looked like they had the best chance of actually working.
With her own recipes also in the bag, we had five hours before the taste testers arrived.
It’s a matter of taste
Over the course of four hours, with me assisting as a slightly useless sous chef, Belfrage rustled up all four of the dishes we intended to test – two dreamed up by her and two by AI. Throughout the cooking process we strictly followed the method, even when the recipe called for improper measurements and gave contradictory serving instructions.
“It’s quite shocking to be honest – I don’t know what it is, and I don’t know what it’s supposed to be,” Belfrage said while stirring the AI main course. Keen to remain impartial, I awaited the decision of our taste-testers.
As the judging panel eagerly gathered outside, Belfrage and I plated up the final dishes. It was glaringly obvious that neither AI dish looked anything like the well-constructed and appetising picture fed to us by the recipe generator. Both appeared to be missing important elements and neither method allowed for the desired presentation, lacking in both ingredient quantity and consistency. It didn’t stop Belfrage’s nerves though.
“I’m a little bit worried,” she said just before serving her Chipotle Oyster Mushroom Taco. “I’m hoping the layers of flavour and texture in this one will win it.”
One by one our seven-strong tasting panel took turns to sample the dishes, unaware of which dishes were devised by Belfrage, and which by AI.
“This looks like cat food,” scorned one taster. “I’m not sure what it is,” said another, both tucking into the rather grey and lumpy looking Oyster Mushroom Symphony offering from AI. But for Belfrage’s Oyster Mushroom Tacos there was nothing but praise. The main course tasting was an undeniable victory for humans, six votes to zero (one of our tasters was a vegetarian from eating the main courses as one of the recipes used chicken stock).
It became immediately clear that this round wasn’t quite so clear cut, as all seven of the dessert tasters struggled to ascertain which dish was created by who, or what. “This isn’t easy,” said one. “Is it wrong that I kinda like both?” said another. “I’m so confused.”
Then came the first blow to humans, as one, followed by another, named the AI’s Chocolate and Orange Symphony as the human creation. Could it be that AI’s creation genuinely tasted better than Belfrage’s? Or is it just near impossible to dislike chocolate and orange?
Either way, the results offered food for thought and that AI can get it right with certain flavours.
The future – AI and human synergy
Whilst it may not be time to throw away our cookbooks and surrender full creative autonomy to AI recipe algorithms quite yet, there are some compelling arguments in favour of AI’s place in the kitchen, especially in the future. These see AI as an enabler, not a replacement for human chefs and their creativity. For AI researcher Bagler, a “chef will never be replaced by an AI. But a chef who does not use AI will definitely be replaced by a chef who uses AI.” He sees it as an aid to help boost creativity rather than replace it.
Balger also believes advancements in computational gastronomy and AI could help to “transform the global food system for a better food future, to achieve better public health and nutrition”. He also believes it could help to reduce the environmental impact of our food by helping us to cut food waste and choose ingredients with a lower carbon footprint.
There’s also undeniably some fun to be had from playing with these dish generators. They have a novelty to them, almost like spinning a food roulette wheel. But for many people, Belfrage included, cooking is about more than simply mixing ingredients together to create a meal. It is an outlet for our creativity, our heritage and personality.
And AI won’t ever take that away.