In July 2023, a research partnership between Meta and 17 external academics published its first results in the journals Science and Nature, based on experiments run on Facebook and Instagram during the 2020 US presidential election. The studies, led on the academic side by Andrew Guess and colleagues, were unusual because the researchers were allowed to change what consenting users actually saw, rather than relying on surveys or observational data alone.
The interventions targeted the recommendation algorithm directly. In one experiment, consenting users’ feeds were switched from Meta’s algorithmic ranking to a simple reverse-chronological order. In another, reshared content was removed from Facebook feeds. In a third, the amount of like-minded political content was reduced. The headline finding across these experiments was that the changes had little measurable effect on the political attitudes studied - affective polarization, ideological extremity, and belief in false claims did not move significantly over the three-month study. Switching to a chronological feed reduced the time users spent on the platforms, and removing reshares reduced users’ news knowledge. Meta highlighted that the studies found “little evidence” that key features of its platforms alone cause harmful polarization.
The results were debated immediately. Some outside researchers cautioned that a three-month window during a single election, on consenting participants, could not rule out larger long-term or society-wide effects, and that Meta’s funding and involvement warranted scrutiny.
Why business readers should care: this is the most rigorous public evidence to date on what feed algorithms do to users, and its nuanced, contested findings are a useful corrective to confident claims in either direction about the power of recommendation systems.