Optimizing for engagement metrics

This post by three AI ethics researchers talks about the concept of engagement as it relates to recommender systems and lists some problems with it. For example, this post points out that internal research from Facebook shows that the most outrage-causing, anger-inducing, and polarizing content gets the most engagement, which causes this content to spread to even more users.

One large review of “moral contagion” found “each message is 12% more likely to be shared for each additional moral-emotional word.” Other studies have found that divisive and extreme material is more likely to drive engagement. Internal research at Facebook has found that “no matter where we draw the lines for what is allowed, as a piece of content gets close to that line, people will engage with it more on average — even when they tell us afterwards they don’t like the content.”

Ultimately, rather than arguing for getting rid of engagement as a metric, this post argues for scrutinizing whether a form of engagement is good or bad, and trying to avoid the bad kinds.

Despite these issues, optimizing for engagement still forms the core of content recommendation on most platforms. While it can be important to offer alternatives, engagement is just too useful a signal of multi-stakeholder value to give up. Instead, it may be possible to get better at distinguishing “good” from “bad” engagement.

I would have liked to see more recommendations in this post of what constitutes good engagement. They do provide a small handful of examples, but ultimately leave it up to the product designers to come up with contextually relevant implicit and explicit signals of engagement from users.