Social media, we know, tends to strengthen our beliefsto make us more hostile towards different opinions and to close us in information bubbles, making civil confrontation with those who think differently increasingly difficult. This dynamic, however, may not be inevitable. A study published in Science in 2025 by Stanford University shows that it is possible reduce polarization intervening from the outside, without modifying the platform’s algorithm.
The researchers developed abrowser extension based on artificial intelligence that rearrange the posts present in the X feed. By showing the least polarizing content first and leaving the most polarizing content at the end, users become more open towards those who belong to the opposite political spectrum. Reversing the order of posts, however, increases hostility.
For now the tool works only on X and only in the browser version. The extension code, however, has been published in full, so as to allow other research groups to develop similar tools for other platforms. The research group also declared that it is working on a version that can be used directly within the apps, even if the implementation is technically more complex.
Let’s see how this tool works, how the experiment was designed and what effects it produced.
How it works: AI scores posts and reorders them
The research group built abrowser extensionthat is, a small program that you add to your browser – like Chrome or Firefox – to add new features. In this case, the extension change the order of posts in the X feed.
Every time a user opens the social network from the browser, the extension analyzes a group of approximately 35 posts. A first language model, i.e. an AI such as ChatGPT, Claude or Gemini, identifies those that contain character statements political. These are then sent to a second model, which evaluates the level of polarization. To do this it takes into consideration various factors, including support for violence politics, the aggressive language towards the opposing side and the distorted representation of political events. He then assigns each post a score ranging from 0, not at all polarizing, to 8, extremely polarizing.
At this point the extension does not modify or delete any content, but simply reorders the posts. If the goal is to reduce polarization, show those with the lowest scores first. If you want the opposite effect, bring those with the highest ratings to the top of the feed.
Tested on 1,200 US participants before the presidential election
To test and verify the effects of this extension, Stanford University has recruited further 1200 participantsall Americans and users of X. About two-thirds identified themselves as Democrats, while the remaining third identified as Republicans. The experiment took place in the weeks immediately preceding presidential elections Americans of 2024one of the moments of greatest tension in the political debate.
Participants were randomly assigned to two separate, parallel experiments. In the first one the feed was reordered by reduce exposure to content polarizingin the second for increase it. Each experiment also included a control group, whose feed remained unchanged, so that the effects produced by the extension could be compared.
The effects of the experiment, which lasted a week, were measured through a series of questionnaires. The researchers evaluated both emotions try while scrolling the X feed, both the attitudes towards the opposing partya measurement known as “affective polarization“.
Those who see less polarizing content are more open to the opposing party
From this experiment it became clear that the contents we are exposed to influence the way we perceive those who have different political ideas from ours. Exposure to a greater number of polarizing content made the participants more hostile towards the opposing party, while a feed composed of more moderate content has increased theopening towards those who belonged to the opposing political party.
These effects were measured through the “emotional thermometer”a scale from 0 to 100 used to rate the degree of liking or hostility towards the other party. Both effects exceeded the 2 degrees. This effect may seem very small, but it is a change comparable to approximately three years of the average evolution of affective polarization in the United States. The shift was seen among both Democratic and Republican voters.
Participants exposed to more polarizing content also reported feeling more get angry and more sad during the experiment, precisely because of the contents to which they were exposed. These emotions, however, were temporary and do not seem to explain the effects observed on affective polarization.
This study shows that it is possible to make social media less polarizing by intervening on the way content is presented to users, without necessarily removing it or limiting its diffusion. The tool developed by the researchers currently works only on browsers, via Chrome and Edge, and only on X. However, the extension code has been made public together with a detailed technical guide, so as to allow other research groups and developers to replicate the experiment and create similar tools for other platforms.
