A perspective on friction interventions to curb the spread of misinformation

Abstract

Social media has enabled the spread of information at unprecedented speeds and scales, and with it the proliferation of high-engagement, low-quality content. Friction-behavioral design measures that make the sharing of content more cumbersome-might be a way to raise the quality of what is spread online. In this perspective, we propose a scalable field experiment to study the effects of friction with a learning component to educate users on the platform's community standards. Preliminary simulations from an agent-based model suggest that while friction alone may decrease the number of posts without improving their quality, it could significantly increase the average quality of posts when combined with learning. The model also suggests that too much friction could be counterproductive. Experimental interventions inspired by these findings would be minimally invasive.

Keywords: Communication; Computational science; Human behaviour; Information technology; Interdisciplinary studies; Technology.

Fig. 1. Information diffusion process.

Each node has a news feed of size α, containing messages recently posted or re-shared by friends. The follower network is illustrated by dotted links pointing from an agent to their friends. Information spreads from agents to their followers, along the orange links (in the opposite direction of the follow links). The central black node represents an active agent posting a new message (here, m₂₀). The new message appears at the top of the followers' news feeds and the existing messages are shifted down. If α messages were already present in a feed, the oldest one is discarded. If the agent wishes to re-share a message, with probability f friction prevents this. Else, with probability 1 − f, the agent shares a post from their feed selected with probability proportional to the post’s appeal (here, m₂)—unless the agent has been exposed to friction earlier and learned (with probability ℓ): in that case, the agent instead selects the post to share with probability proportional to the post’s quality (here, m₁).

Fig. 2. Average post quality q^T as a function of friction probability f, for different probabilities of learning ℓ.

The subscript T indicates that average quality is measured at convergence. Shaded areas indicate standard errors.