As smart as we think AI is, and as much as we assume that the Silicon Valley companies know about us, there are times when I look at the results of one of their algorithms and I’m left a little dumbfounded by an obvious issue.
Last year I noted that it was silly to show me ads for a restaurant I had just left, or for a show I had already bought tickets to using a Facebook link.
This week, it’s LinkedIn’s turn.
If you haven’t looked at the suggested contacts screen on LinkedIn recently, you may have missed that they’ve started breaking those suggestions into sections. I actually like this tremendously. Instead of a random list of suggestions in no particular order, they’ve moved that list below sections for “People You May Have Worked With”, and “Alumni You May Know”
So, taking a quick look, this seemed like a helpful addition. The first list is fairly heavy with people who work, or have worked, at Nuix, my previous employer, or my current one. Sure, there’s a somewhat decent chance that I know folks, or at least know a lot of the same people. The problem is, when they configured this algorithm, they clearly made no effort to account for dates. (Or they decided it didn’t matter, who knows?) For example, I worked at Nuix from Dec. 2014 – April 2017. That list shows me a few names that sound familiar, but a bunch that don’t at all. Turns out that might be because they started working there well after I left. Or, for my current firm, there are suggested contacts who left years before I started.
There isn’t really any rational reason why LinkedIn should suggest that I know them.
It gets worse in the Alumni section. I have Ohio State listed under my education, and I went to school there in the late 80s. My suggested contacts seem mostly to include people I have common connections with, or who work in the legal industry. That’s good. But again, they’ve made no effort at including the dates as part of the algorithm. Ohio State has 65,000 students. The chances that any random person who went to school there at the same time I did know each other is pretty small. The chance that some random person who went there years later and I know each other? Infinitesimal.
Think about it. LinkedIn knows that I left Ohio in 2011. It’s right there in my job history. A paralegal who graduated from OSU after I moved away, and works for a firm that didn’t exist in 2011, and has no common connections to me, should not be “someone you might know”. There aren’t anywhere near enough touch points to assume that. As a human we can see that, and probably wouldn’t make that connection, but the algorithm is seeing something, and making the leap.
That’s the thing about some of our current algorithms. The AI is great at finding thousands upon thousands of possible connections between people, places, and things. It’s not so great at ignoring the meaningless ones.
Eventually, we assume they’ll get better, but it should also serve as a reminder that big data by itself may have more answers than we need questions answered, and we may need to learn which ones actually matter before we make some very bad decisions.