A few months ago, YouTube recommended a video to me about the psychology of people who spend a lot of time alone. I had recently started doing remote freelance work. I was home most of the time, and, consequently, alone most of the time. But, I had never searched YouTube for anything about spending a lot of time alone.
I stared at the recommendation for a second and thought, “Well. That’s a little unnerving.”
The strange thing was not that YouTube had recommended something completely unrelated to me. It was the opposite. The recommendation was accurate enough to feel personal, even though I had no idea what combination of signals had produced it.
YouTube does not need to know what my days look like in the human sense of knowing. Its recommendation system uses behavioral signals such as watch and search history, likes and dislikes, “Not interested” feedback, subscriptions, and patterns among viewers with similar interests. YouTube itself describes recommendations as predictions about what a viewer is likely to enjoy, not as conclusions drawn from some complete understanding of the person watching.
Still, when a prediction lands squarely on something true about your life, the mechanics become easy to forget. It feels a little like being seen. And that raises a question I have started thinking about more often:
Who, exactly, does my algorithm think I am?