“Pickles,” chef Michael Laiskonis tells me as I poke at the plate he’s put in front of me. The menu said beet salad, but this isn’t that weary marriage of root vegetables with goat cheese, pistachio, and arugula. The former Le Bernardin chef reveals that it’s roasted beets and white beans with a roasted prune dressing, topped with pickles, an outlandish combination he created with IBM’s supercomputer, Watson. I try it, and the flavors taste fresh, unexpected, and satisfying.
Many innovations promise to optimize the experience of the chef in the kitchen. Too often, these “improvements” remove something essential to the process, the human element. However, not all technology is a cheat code. Some new technologies act more like a muse.
At SXSW this past year, IBM premiered its cognitive cooking collaboration with the Institute of Culinary Education. IBM’s supercomputer Watson used its big data capabilities to generate lists of complementary ingredients, which ICE chefs like Michael Laiskonis and director of culinary development James Briscione then developed into full-fledged recipes. At SXSW, they handed out food-truck goodies like chocolate burritos and also treated guests to a sit-down dinner with dishes like the surprising beet salad.
Today IBM has announced that Watson can now generate recipes itself, and in partnership with Bon Appétit, the company is releasing a beta version of an app called Chef Watson With Bon Appétit.*
The computer giant isn’t trying to replace people with robots that chop, sauté, and plate but rather help stoke the fire of chefs’ minds and lead the rest of us to better recipes. Instead of trying a handful of combinations at one time, the supercomputer can taste—virtually—thousands of ingredients to deliver an instant list of ingredients and rough set of instructions informed by Bon Appétit’s 9,000 recipes. For example, when I challenged an earlier version of the technology to suggest ingredients for an Indian cookie, it came up with a list of 21 options that fit the bill, including standards like cardamom and nuts as well as new goodies like gravy and peanut oil.
This is a step forward for Chef Watson: An earlier version generated ingredients, without any recipe framework. When it suggested an Italian roast with duck as the protein, Laiskonis and Briscione came up with remarkably different recipes. Laiskonis made a sous vide duck breast with apple-caramel puree; a tomato-ginger duck stock; and a gremolata of celery, olives, and herbs. Meanwhile Briscione made a fennel-cherry duck sausage, which he braised in a tomato-porcini sauce and served with an olive-cherry gastrique and poached-apple confit.
Even with a machine-generated recipe in hand, though, executing it is different from reading it. James Beard Award–winning chef Paul Qui told me the worst thing his staff can say is, “I followed the recipe, chef.” Raw ingredients like kale, yellowtail, peppers, and spices change from day to day and season to season. Watson can’t take that variability into consideration. Then there are color combinations, flavor balance, plating, seasoning, size, temperature, and texture (for example, knowing that a prune puree would work with the crunch of pickles on the beet salad). When real-life chefs look at the computer’s basic flavor combinations, they easily see omissions, like spice, heat, or acidity. “Watson’s not thinking for you,” said Adam Rapoport, the editor in chief of Bon Appétit. “It’s helping you think.” The app provides the general idea of a recipe, but a sophisticated cook will most likely want to modify it.
So how does Watson arrive at a recipe? The app sorts through all possible ingredient combinations and generates a list that takes into account surprise, pairing, and pleasantness.
Now that Watson has learned what makes a cookie a cookie, for example, it can compare its list of ingredients to the thousands of cookie recipes in its database. The more its list deviates from traditional recipes, the more surprising it is. Flavor pairing works similarly. IBM researchers catalogued more than 1,000 chemical compounds, and Watson uses that data to analyze shared compounds among ingredients. “In Western cuisine, the more flavor compounds the ingredients share, the more likely they are to taste good together,” said head researcher Florian Pinel. The opposite is true in Eastern cuisine, where opposing compounds make for intense contrast.
OK, but pleasantness seems more subjective, right? After extensive experiments using hedonic psychophysics (the psychology of pleasure and pain, for all us laypeople), IBM researchers developed a universal pleasantness score for those 1,000 flavor compounds. Watson uses that data to extrapolate a universal pleasantness score for any other compound. It can then add the scores of each compound to determine which combination is the most pleasing.
Going back to my Indian cookie experiment, coconut, ginger, and peanut oil all share one flavor compound, which has a high pleasantness score. Meanwhile cheese, coconut, and ginger share a compound that is slightly less pleasant, but that combination rated higher in terms of surprise. Watson uses all of this information to create a unique, surprising recipe.
It seems fitting that the cooking app turns on the idea of pleasure, since food connotes comfort and contentment. We are hardwired to connect food with the ultimate emotional communication, love. You bring food to your sick neighbor, celebrate your anniversary at a special restaurant, and make your mom’s spaghetti recipe when you’re lonely. In the book Like Water for Chocolate, the main character’s emotions transfer entirely to her cooking: When she can’t marry the man she loves, she bakes a cake while crying, and those who eat it vomit, cry, and pine for true love.
With Watson’s culinary endeavors and the new app, IBM and Bon Appétit aren’t trying to make anyone feel longing or joy but rather provide thought-provoking recipes. There’s no formula to harness imagination and convey a message, because creativity is subjective. A McDonald’s hamburger tastes delicious, but Qui says it ultimately fails because it doesn’t have “soul.” How do you quantify soul? It’s the difference between a cook and a chef, a tradesman and an artist, the recipes Watson generates and the recipes a chef creates. If technology can help a chef make something even more inventive, it’s not only interesting to use it but increasingly important.
For example, what if Watson’s future cousin could create flavor profiles that made indigenous but bitter plants taste better, thereby increasing a restaurant’s sustainability? Allowed chefs to offer personalized flavors of Soylent rather than the current chalky mess? Robyn Metcalfe, the director of the University of Texas’ Food Lab, imagines a day when Watson could modify a specialized diet with streaming data from, say, your Jawbone UP to compose a recipe that is ideal for you at that exact moment.
IBM’s cognitive cooking project works because it bridges technology with human imagination. A dish entirely fabricated by a machine will never be as satisfying as one from a human chef, because we want more than just food when we eat. We want a connection with another human being, whether that’s sitting at the family dinner table or savoring the latest special from a James Beard Award–winning chef. When chefs use a tool like Watson to invent new dishes, the possibilities are endless: Technology can inspire even greater creativity when you pair the depth of big data and machine learning with the breadth of human inspiration.
Update, June 30, 2014: This piece originally stated that the Chef Watson With Bon Appétit app beta was capped at 200 users. IBM has since stated that there is no longer a hard cap. (Return.)
This article is part of Future Tense, a collaboration among Arizona State University, the New America Foundation, and Slate. Future Tense explores the ways emerging technologies affect society, policy, and culture. To read more, visit the Future Tense blog and the Future Tense home page. You can also follow us on Twitter.