I think it's in the sweet spot where it's good enough that you can make some money by turning it on, and when you do the A/B test of, "Here's the unfiltered site, and here's something with a clumsy, basic personalization built in," yes, people click more on the things in the clumsy, basic personalization. But it's not good enough that it really reflects anything about any of the nuances of who people are [clicking]. I was talking to Chris Dickson, who runs Hunch, and he said that with five data points — if you answer five questions — that they can then predict almost any other sort of consumer preference question with 80 to 85 percent accuracy. And you can use that, because 85 percent is pretty good, but on the other hand, it's a very depressing version of yourself if you're boiling it down to things like gender, or "Are you an introvert or an extrovert"? That's hardly what a human being is, nothing more than five data points you're then using to predict a product preference.
Another thing that Hunch is wrestling with is: If you have the five data points for two people, and they're both friends with a third person, you can predict that [third] person's answers pretty well, too. They realized that with just the public data off Facebook, for example, you can make a pretty decent gayness predictor. You can take any person on Facebook and, based on the answers given by their friends, you can predict if they're gay. Obviously, the social consequences of that are pretty significant, and also, in 20 percent of the cases, you're revealing something that's significant about someone, and you're totally wrong. These are some of the dilemmas that this stuff creates.
Is there a difference between this being a good guess to sort of nudge in one direction, and this being near-certainty? So, when we're dealing with just a few data points – 80 percent accuracy – no one really wants to turn to youa la Sherlock Holmes and say, "Oh, you must be from mid-coast Maine, and you must be dating a girl of Greek extraction!" There's sort of spooky stuff behind it, but it's just enough to nudge it and suggest that maybe the LL Bean ad is more appropriate than Hollister. Why is that so bad? At least in the advertising space, it increases the chances that you're going to get something relevant.
In the advertising context, I think there are cases where it can be really problematic and it can create these compulsive loops where someone who searches for the definition of obese on dictionary.com gets a cookie dropped on her computer and then she's chased around the web by weight loss ads. That doesn't seem like a particularly good thing. And certainty there are even worse examples. But, ultimately, at least with ads, there's a basic understanding that this is happening. People are starting to notice that it's probably not the case that everyone is getting the same ads for the same exact pair of sneakers that I did two weeks ago. When you get into content – particularly because it's not being done in a transparent way, and you don't see it at work, and you can't get a sense of the dynamics of who Google thinks it is that then informs the search results, or what Facebook is learning about you that's then tailoring your News Feed – I think that's more problematic, because you want people to be able to understand the filters through which they're seeing the world.