In front of me sit two tantalising cakes. One is a glossy, mirror-finished chocolate creation, the other a buttercream-frosted variety. Both look great, but there’s a fundamental difference between them. One is the brainchild of 2020’s Great British Bake Off winner Peter Sawkins. The other was dreamed up by artificial intelligence (AI).
They have been made as part of a challenge to see if AI can rival humans when it comes to baking. The brief was to come up with a cake recipe incorporating Maltesers.
Sawkins’ mousse cake is decadent, with a crunchy base and a caramel centre, laced with soy to emphasise the savoury hit you get from malt. The AI creation is less delicate, its body a hybrid of cake and cookie, its coating infused with Marmite.
The i newsletter latest news and analysis
After trying both, I’m relieved to find I marginally prefer the one created by my fellow human – but they both taste very good. Which raises the question: is baking another area where we could be left redundant one day?
Creativity if often cited as an area where machines won’t be able to replace us. But if AI can create recipes from scratch, could tomorrow’s Bake Off winners be algorithms, basing their creations on data and equations rather than experiment and experience?
It’s worth bearing in mind that the Google Cloud AI used for the challenge is itself the product of one person’s love of baking.
“At the beginning of the pandemic I found myself with a lot more time at home, so I started baking,” explains Sara Robinson, senior developer advocate at Google Cloud. She started to think about what all the recipes she was using had in common.
“A cake has a specific ratio of flour, fat, sugar and eggs, and other types of baked goods have other distinct ratios.”
Because “machine learning at its core is all about finding patterns”, Robinson wondered if she could combine her “newfound love of baking” with her research.
She built an AI model, then gathered a range of hundreds of recipes for a data set. The program examined which ratios of ingredients make different types of foods – basically learning the theory of baking. This knowledge meant it could invent its own recipe, whether for cakes, cookies, scones, traybakes or a combination.
That’s where the computerised creativity stopped. The AI still needed a human to write instructions on how to combine the ingredients and come up with an idea of incorporating Maltesers imaginatively to meet the brief. (It was also a human that baked the cake, following the AI recipe – though doubtless a robot could quickly be invented to do this job.)
AI successes and failures
Lee Se-dol, a master player of the Chinese strategy game Go, announced in 2019 that he was retiring, after losing in 2016 to the AI program AlphaGo, developed by Google’s DeepMind unit.
Inverness Caledonian Thistle FC used an AI camera system to record matches, but in a match last October it mistook the bald head of a linesman for the ball.
An algorithm is producing the best computer-generated poems yet after analysing half a million lines of verse, in a project led by the artist Es Devlin, it was reported this month.
In 2019 Hong Kong investor Samathur Li Kin-kan said he would sue the firm Tyndaris, claiming its supercomputer had cost him millions.
While baking is often described as being a science as well as an art, Sawkins finds that life experiences are key to his recipe ideas – it’s creativity that comes before technical considerations.
“I think about the stories, connections and memories – the nostalgia of the bake,” he says. “Then I think about the flavours and how I’ve been creative with them, then finally the textures.”
In this case, his creation started with the childhood memories that Maltesers inspire in him: his mother’s traybake and happy times “snaffling it” with his brother.
The next part is where AI could prove helpful, Sawkins believes. “If I’m making a cake, I’ll look up different recipes that make this same thing and they all have different ratios of ingredients.”
He averages these out, looking at the boundaries when it comes to adjusting them, and what qualities he wants in his final product. Aggregation and analysis of information is the perfect task for a machine sous chef.
Robinson says that the benefit of AI in all kinds of fields is in creating “space for creativity” for its human counterparts, rather than stealing our star role, and the same proved true with baking.
“AI can make a great helper in the kitchen. For me it provided a great base for recipe development, because it’s not something I had done before. To start with the ingredient amounts that the model gave me provided confidence that the recipe was going to turn out as I expected, then that gave me room for creativity… It was the human element that really made the recipe shine.”
“Do I think my role is safe from the AI machine?” Sawkins asks. “I think so.” But he adds: “It could be a winning combination rather than one or the other winning.”