Almost every social media user has encountered posts that are irritating, contradict their beliefs or make them wonder why such content appeared in their feed in the first place. New research offers one possible explanation: recommendation algorithms can interpret negative reactions as signs of interest. Researchers examined the feeds of 715 US-based users of X and compared the values expressed by those users with the content selected for them by the platform’s recommendation system. The results suggest that the algorithm does not necessarily prioritise posts aligned with a person’s values. In some cases, it is more likely to amplify content that conflicts with them. The explanation lies partly in how engagement is measured. Likes, replies, clicks and other interactions do not necessarily carry equal weight. People frequently like posts they agree with, but when they take the additional step of writing a reply, they may be doing so because they disagree, are angry or want to challenge what they have seen. The researchers found that X’s algorithm gives particularly strong weight to replies. A user may therefore write a critical response to a post they strongly dislike, while the system primarily registers intense engagement. It then learns that similar material is effective at generating a reaction and may recommend more of it.
This can create a feedback loop: people see content that annoys them, reply to it, the algorithm interprets the interaction as valuable and then serves them more comparable posts. The researchers also identified differences between political groups within the US sample. For participants who identified as Democrats, the content amplified by the algorithm was more than four times more misaligned with their stated values than it was for Republicans. The authors argue that this does not necessarily demonstrate political bias in the algorithm itself. Their evidence suggests that Democrats in the sample were more likely to object to the content they replied to, creating a stronger feedback signal for the recommendation system. The broader problem is that an algorithm optimised for engagement does not necessarily understand why someone interacts with a post. From the system’s perspective, outrage, disagreement and approval can all appear as different forms of the same valuable commodity – attention. Researchers warn that this dynamic may contribute to political polarisation rather than simply exposing users to different viewpoints. One possible alternative would be to give people greater control over the values and priorities used to organise their feeds. The practical lesson is surprisingly simple: repeatedly arguing with posts you do not want to see may unintentionally teach the algorithm to show you more of them.
