Cooked Data: IBM's Watson Wants Access To Your Kitchen

Bottom line: probably not optimal for picky eaters...JL

Alexandra Kleeman reports in The New Yorker:

I.B.M. exposed its algorithms to the entire recipe archive of Bon Appétit, as well as to recent research in “hedonic psychophysics”—“the psychology of what people find pleasant.” The algorithms also took note of which ingredients tended to be combined, and inferred the roles they seemed to play in a dish. The result is a browser-based Web app that allows users to generate recipes by selecting a permutation of ingredients and a style of cuisine.

The interface for Chef Watson, I.B.M.’s artificial-intelligence cooking app, is simple and welcoming, a minimalist canvas of four empty text fields and four dove-gray circles. You type in the ingredients, or let Chef Watson choose them for you according to its own mysterious logic: tomato, garlic, onion, purple seedless grape. These four ingredients, Watson declares, have a “synergy” of a hundred per cent—they are an unimprovable combination, chemically speaking. But, as an embodied being who has tasted those ingredients, you might be skeptical about combining them—especially when you scroll down to the suggested recipes and discover, near the top of the list, something called Purple Seedless Grape Starch Dish.

The recipe also calls for “sixty-seven medium trimmed Easter-egg radishes,” black beans, cinnamon, curly parsley, marjoram, and Calvados. Cook, salt to taste, then top with Jack cheese, olive oil, and the grapes, “for squeezing over.” And there you have it: the computer-assisted future of cuisine, in the form of a pile of sweet-smelling, mud-colored radishes.

So far, artificial-intelligence researchers have mostly built machines capable of demonstrating their own prowess. At I.B.M., engineers have used natural-language processing and enormous computational power to beat the most proficient humans at our own games, like chess and “Jeopardy!” Having achieved these goals, Watson’s handlers now imagine a more intimate, domestic role for A.I. To create Chef Watson, I.B.M. exposed its algorithms to the entire recipe archive of Bon Appétit, as well as to recent research in “hedonic psychophysics”—“the psychology of what people find pleasant.” The algorithms also took note of which ingredients tended to be combined, and inferred the roles they seemed to play in a dish. The result is a browser-based Web app that allows users to generate recipes by selecting a permutation of ingredients and a style of cuisine. Watson can invent several dozen recipes that prominently feature prunes; it can satisfy a request for banana biscotti in a Creole or a Basque style; and it makes suggestions that no human would ever make, like adding milk chocolate to a clam linguine or mayonnaise to a Bloody Mary.

With Watson’s help, I cooked some eggplant fritters that made convenient use of every sad, wrinkling root in my refrigerator’s crisper. (Combining seemingly incongruous spare ingredients is the app’s most practical function.) I made a butternut-squash-and-shrimp sandwich—a tuna-and-pickle sandwich from the Bon Appétit archive, transformed by Mad Libs logic. I made a caper-and-fennel salad that was lovely, though I left out the suggested cocoa.

After a week of collaborating with Watson, I began to worry that I wasn’t giving it a fair trial. Perhaps, by using whatever I had on hand and selecting for novelty, I was making Watson seem kookier than necessary. I decided to impose the social pressure that a skittish cook like me needs: I scheduled a dinner party for myself, my husband, and four nonjudgmental friends.

On the morning of the party, I wandered the grocery store with phone in hand, trawling the app for familiar summertime dishes with just a dash of robotic weirdness. I settled on a cherry-tomato gazpacho, followed by a clam-and-salmon paella and a maple-syrup ice cream. Watson’s gazpacho recipe called for cabbage; or, according to a drop-down menu of psychophysically similar ingredients, I could substitute squash blossoms or watermelon radish. The app’s quantitative approach made cooking a simple, combinatorial thing, an equation with variables waiting to be filled in.

When it came time to prepare the paella, Watson was cagier. “Add enough fish stock to measure the remaining fish stock mixture,” it told me, an intriguing Zen koan but hardly a useful instruction. When I took the paella out of the oven, I found myself poking at a heap of tough undercooked rice, gooey overcooked rice, unopened clams, and desiccated salmon.

By dessert, I was ready to mutiny. I decided to discard Watson’s ice-cream recipes, all of which called for butter or garlic or curry powder, and go it alone. That Chef Watson unleashed an improvisational cook within me is evidence of how frustrating the program often is, and how productive that frustration can be. Sifting through dozens of recipes had taught me that ice cream was just a creamy base and flavoring, so I relied on my human intuition. I boiled blueberries with brown sugar to make a compote, which I stirred into a vanilla-flavored base and sweetened with maple syrup. For the first time all evening, my guests looked delighted. ♦