When platform metadata such as follower counts and engagement metrics is added, the models begin to rely heavily on those signals. The predictive power of post likes and retweets jumps from near zero to almost one‑fifth of a point more than before, making high‑engagement posts—and the authors who regularly generate such engagement—more likely to be recommended. This amplification of already existing demographic differences is largely driven by the fact that users with larger followings tend to post longer, more negative content, and the models implicitly learn to favour those signals.

Providing explicit author demographics to the model, however, changes its recommendations very little once textual and platform signals are available. The models already capture much of the demographic signal through engagement patterns and content style, so disclosing demographics does not significantly alter which posts are surfaced.

The study’s findings underscore that LLM‑based recommendation systems are capable of reproducing and even magnifying demographic inequality, not by using demographics directly but by relying on content and engagement signals that are themselves socially structured. The authors caution that their sample is not representative of the broader social media user base and that the experiment does not include personalization, repeated exposure, or feedback loops that could further affect visibility.

The visual evidence from Figure 2 illustrates how LLMs prefer longer, more negative, and more opinionated content in the text‑only condition, while Figure 5 shows how platform metadata amplifies existing skews through engagement signals and that adding explicit demographics has a small effect once these signals are present.