Mitigating the Harm of Recommender Systems

Because recsys’ are based on user-driven behavior and content, recommender systems are vulnerable to bad actors with malevolent intent. This week’s discussion topic provides 3 articles that relate to bad-actor-driven content, its impact, and later we discuss ways we might mitigate it.

The Harm

The harm caused by recsys is rather prevalent in today’s society. This NYTimes article in the discussion prompt is about how YouTube’s recsys can lead one to videos produced by the far-flung fringes of the political spectrum – exposing users to ever-more-sensational programming. As an example of the damage, the author provided a link to a study titled Selective Exposure to Misinformation: Evidence from the consumption of fake news during the 2016 U.S. presidential campaign that concluded that misinformation favored Trump more than Clinton in the 2016 election.

In the same study they provide another table that provides a clear connection between popular websites that utilize recommender systems and the connection to misinformation:

Pre/Post Fake News Consumption

Pre/Post Fake News Consumption

According to this study, the top 10 apparent promulgators of misinformation (“fake news” articles) are big-time websites we all know and use – Google, YouTube, and gmail among them. The possible harm caused here is difficult to quantify but it’s clear that harm was the outcome.

Mitigating the Harm

There is no easy solution to this problem. Perhaps data could be gathered to identify when traffic is being directed to sites known for spreading misinformation and label the results accordingly. Human oversight could be effective but it’s an unlikely option for big tech. For each of these suggestions I can think of many reasons they’d be huge problems (e.g. criteria for labeling a “harmful site” could also be hacked, human intervention would be expensive/unsustainable).

Making truth as exciting and profitable as misinformation is already an impossible task without recsys recommendations but with them it makes this whole thing an anxiety-inducing problem to think about.=