The recipes that made it into the final cookbook, Abrams said, were ones whose flavors Watson scored highly across three traits: surprise (“How different is this ingredient combination from all other combinations?”), synergy (“How harmoniously does Watson think the ingredients will blend?”) and pleasantness, a quality that, Abrams explained, Watson understood through its study of “hedonic psychophysics,” a daunting-sounding phrase that’s “really just the science of our likes and dislikes.”
So how does something that cannot taste learn which foods taste good? From the beginning: by memorizing everything that tastes good, and why.
For Watson, it was a three-step process, starting with the biggest picture and working inwards. First, the computer scanned a vast collection of online recipes to understand which flavor combinations cropped up frequently, like chocolate and banana or garlic and basil, and their molecular components. It then narrowed its focus to the level of cuisine, using the recipes to learn which flavor combinations made something Chinese or Italian or Tex-Mex. (As a rule of thumb, Western cuisines value combinations of ingredients with similar flavors, while East Asian cuisines tend to prioritize contrast. Which may be why it’d be hard to find somebody who would gleefully chow down on a peanut-butter-and-mustard sandwich, for example, but decidedly less so to find someone who’s put spicy mustard on their peanut-sesame noodles.) And finally, it zeroed in on the definitions of individual dishes.
“You know what pizza is because you’ve eaten pizza, you’ve experienced pizza. Watson knows what pizza is because it’s read about pizza,” said Abrams, the leader on the project. “Watson read a whole bunch of pizza recipes, and has learned that there’s dough, and some sauce, and vegetables and meats and cheese.”
Starting from the very basics meant that Watson could amass all the knowledge of a master chef with none of the habits or preconceptions. Chocolate, apricot, and ground beef sound strange because they are, in the truest sense of the word—the combination is an unfamiliar one. But according to Watson, it’s also one whose flavors make sense on the molecular level. Here, via Quartz, are some of the cookbook’s other entries:
Belgian bacon pudding, a desert containing dried porcini mushrooms
Vietnamese apple kebab, with [a] mushroom-and-strawberry pairing
Portuguese lobster roll, with appetizing “saffron fluid gel”
Hoof-n-Honey ale, with veal stock
Thai-Jewish chicken, with potato latkes and rice balls
The shrimp cocktail, which is a beverage with actual shrimp in it
Kenyan Brussels spouts, with cardamom
But the chefs who worked on Cognitive Cooking are quick to reject the idea that flavor, the science, trumps cooking, the art. Watson “is certainly not replacing the human element of cooking,” said Michael Laiskonis, the creative director of the Institute of Culinary Education. “Most advances in cooking over the centuries have come about because of technological advances, whether it’s a stove or a more efficient way to grind chocolate or a blender or an ice-cream machine. To me, this just represents a continuation of that evolution.”
Or a fancier gloss on what is, at its heart, a very straightforward act. Watson has done at a computer’s speed what any cook does at a human one: In the kitchen, everyone’s a “hedonic gastrophysicist,” learning by doing about what tastes good.