Do readers engage less with AI-generated social media posts?

On Medium, the paper compares engagement across posts its detector predicts as human-written (predicted-HWTs) and posts it predicts as AI-generated (predicted-AIGTs). The gap runs in one direction. Predicted-HWTs average 127.59 likes against 69.15 for predicted-AIGTs, and 7.38 comments against 4.16. The excerpt adds that predicted-AIGTs "exhibit a higher frequency of low 'Likes' counts" (Figure A3a), and that "across all follower count groups, AIGTs receive significantly fewer Likes and Comments compared to HWTs" (Table 5). The engagement comparison is reported for Medium only. The excerpt gives no engagement figures for Quora or Reddit.

The paper's reading is that "users in Medium are generally more willing to engage with human-written content." It then qualifies that reading: "the relatively small gap between the two suggests that AI-generated content appeals to users." The excerpt offers no mechanism for the gap beyond this interpretation. It does not say whether readers can perceive the difference, whether authors who use AI write about different subjects or differ in other ways, or whether the gap changed over the 2022 to 2024 window the tracking covers.

This result sits against the argument in Does polished AI output trick audiences into trusting it?, which holds that polished generated output borrows the authority audiences give to expert presentation. On Medium, the posts labeled as AI drew less engagement, which points the other way. Engagement is not the same as perceived authority, though, and the excerpt does not test that argument. The finding also supplies a behavioral counterpart to the belief-based perception gap in How much of the internet is AI-generated now?, where a user study measured what people believe about AI content. Here the measure is what readers did on the platform. The labels themselves come from the detector whose platform-level rates are reported in Is AI-generated content rising faster on some platforms?.

The excerpt does not establish whether the gap reflects reader preference, differences in what or how authors write, or post timing. The labels are also predictions. If the detector misclassifies some posts, the two groups blur together, which would tend to shrink the apparent gap. The excerpt reports no error rate on these posts, so the gap cannot be corrected for that. The supportable claim is narrow: in this Medium sample, posts labeled as AI drew fewer likes and comments in every follower group. That says something about one platform's audience in this period. It does not show that readers can tell AI text from human text, and it does not show that AI content is generally unpopular. The "appeals to users" reading belongs to the authors, and it rests on a gap they themselves describe as small.