IBM's Watson can't taste what it cooks

Years before he headed up the team to build IBM's Jeopardy computer, Watson, David Ferrucci co-authored a book. It's called Artificial Intelligence and Literary Creativity--Inside the MInd of Brutus, a Story-telling Machine. The idea was to program a computer to recognize various elements of story-telling--from coming up with a plot to creating characters--and then let it start authoring. Brutus effectively assembled words into coherent sentences and paragraphs, Ferrucci later told me. But it had no way of knowing if what it created was any good.

And that's the problem with teaching machines to be creative. To date, they haven't been blessed with good taste. Sad to report, that's even true when Watson ventures into the kitchen. But perhaps it can find a role as a wild and inventive sous-chef.

Anshul Sheopuri is on the IBM team training Watson for kitchen work. The idea, essentially, is that most humans are locked into cooking with combinations that we know. Tomatoes, for example, go with oregano, and with Parmesan cheese. Someone might try throwing in a pinch of marjoram with that, or an ill-advised turnip. But outside of celebrity chefs, most of us paint with a small and conservative pallet.

IBM's goal is to get Watson to experiment, suggesting combinations most people would never consider. For this, the IBM team has the computer reduce food to its molecular components, and then rearrange them based on each ingrediant's chemistry and what the computer has gleaned from lots of written material about food. It then comes up thousands of new recipes. Based on its readings, it gives each recipe a score for novelty, its quality to blend with other foods (pairing) and "pleasantness."

Of course, this is all academic. For the computer, pleasantness, along with everything else in the universe, is simply a string of ones and zeros that appears to correlate with other ones and zeros that make up its point of reference. It's up to human beings to pick the most promising looking recipes, cook and taste them.

The results include novel blends, such as the sweet and piquant Caymanian Plantain dessert (above). Have to say, though, I glanced at the recipes, and they didn't look as surprising as I would have imagined. I wouldn't be at all surprised to see the Swiss-Thai quiche on the menu of the Asian fusion restaurant down the road. But the caution could come from conservative humans. Maybe the people on the IBM team picked the recipes that they liked best, and that they figured fellow humans coming up to their food truck at South by Southwest were most likely to enjoy. Watson is hard-pressed, at least for now, to make such judgments. So my guess is that it didn't get a chance to fly its freak flag.

With time, this type of computer chef will be able to incorporate all sorts of other analysis in its recipes, incorporating data about health, disease, prices and supply chain. In fact, those areas are in Watson's comfort zone. Cooking? Well, it's still learning. The problem, from Watson's perspective, is that the boss has this thing called a tongue.