Will Google Use Humans to Fix Google Now?

I have never used Google Now, but I was always skeptical. The use cases that were being reported on the web always were extremely limited, and it seemed that they were simply telling us about the things that worked but not about the things that didn’t.

A couple of days ago, Janko Roettgers wrote on GigaOM about how Google Now actually fails, even in the limited tasks it’s supposed to be good at.

We often assume that because Google is collecting huge amounts of information about their users, they are able to understand a lot about who we are and what we want to do. Yes, Google does know where you live and it knows exactly what keywords you used on your searches yesterday. It knows exactly where you are right now. This is all very creepy, especially if you are an Android user. The problem is, there is no guarantee that this information will let Google know your current intentions with any accuracy.

Google has been described as a “decade-old machine learning project”. There are two products from Google which have shown how this can actually be turned into something useful. They are Google AdSense and Google Maps.

Google search basically selects the ads (AdWords) to be displayed based on the keywords that were entered in the search field. Although the information that Google holds about the user is also used to tweak the results, this is not the main driver of the AdWord ads to be displayed. With AdWords, the user has directly expressed their intent through the search keywords and Google does not try to second-guess it. This intent is what drives the ads and this is why I don’t include it in the current discussion.

AdSense is more “intelligent” than AdWords because it tries to guess user intent. It analyses the content of the page the user currently is browsing. It uses knowledge of what pages the user has visited recently. It uses location data. It combines all of these to determine which ad the user is most likely to click on. In reality, AdSense fails most of the time. It does not correctly estimate the user’s current intent. But that’s OK. The click-through rate (CTR) of AdSense is at most a few percent and the vast majority of people aren’t interested in what is shown in an online ad. It’s completely acceptable for AdSense to get it wrong most of the time. Therefore, the “intelligent” guesses of AdSense are successful even if they are completely wrong most of the time. AdSense will still be immensely successful when the majority of users are pissed-off by the ads. It only takes a small percentage of correct answers to succeed.

Google Maps is a totally different kind of “intelligence”. Google Maps has to be accurate for the vast majority of the time. It has to be something like 99.999% accurate or more. Otherwise, we would constantly get reports that some poor drivers drove themselves into a desert. How can Google Maps be so accurate when AsSense is so sloppy? The secret lies in humans. To quote an article by Alexis Madrigal on The Atlantic.

I came away convinced that the geographic data Google has assembled is not likely to be matched by any other company. The secret to this success isn’t, as you might expect, Google’s facility with data, but rather its willingness to commit humans to combining and cleaning data about the physical world. Google’s map offerings build in the human intelligence on the front end, and that’s what allows its computers to tell you the best route from San Francisco to Boston.

Even though Google gets vast amounts of information from satellites and street-view cars, it has to combine these with an army of human beings to gain accuracy. Without these human beings, they cannot get the error rate to an acceptable level. The kind of “intelligence” in Google Maps can only be attained with a huge number of people who manually curate the information.

Now let’s get back to Google Now. Which kind of “intelligence” do we need? Do we want a personal assistant that thinks it is acceptable to guess correctly only a few percent of the time? Or do we want a personal assistant that truly knows what we want to do next?

If we want the latter, we may have to be content with an army of human beings plowing through your most personal information, helping Google’s not-so-accurate machine learning algorithms to make sense of your daily routines.