Stanford researchers develop a web-based tool that uses an LLM to downrank X posts with antagonistic language in a user's feed, to reduce “partisan animosity”
www.science.org/doi/10.1126/ ... Mike Jay M / @mikejay.link : Evidence suggests that being on X and using the “For You” feed can rapidly cause political polarisation. — And political parties believe they can still use X? — The fuck is wrong with you? Margot Finn / @eicathomefinn : ‘Most of the more than 1,000 users who took part in the experiment during the 2024 US presidential election did not notice that the tone of their feed had been changed.’ 3/3 Margot Finn / @eicathomefinn : ‘The degree of increased division - known as “affective polarisation” - achieved in one week by the changes the academics made to X users’ feeds was as great as would have on average taken three years between 1978 and 2020.' 2/3 www.science.org/doi/10.1126/ ... Andy Reisinger / @andyreisinger : Social media feeds can be tuned to increase or decrease political polarisation. — Nice demonstration by @tiziano.bsky.social and colleagues Elizabeth Jacobs, PhD / @elizabethjacobs : My interpretation of this article: — You don't need a violent coup or a hacked voting machines to overthrow democracy. — You just need control of the algorithm. Elizabeth Jacobs, PhD / @elizabethjacobs : This article provides more data showing how easy it is to manipulate humans on social media. — These researchers rerouted the algorithm on Twitter to push some users toward “antidemocratic attitudes and partisan animosity”. — It only took 1 week to elicit changes that used to take 3 years. @annkspencer : “The Stanford-led research, published in Science, also indicates that it may one day be possible to let users take control of their own social media algorithms.” — paper link: www.science.org/doi/10.1126/ ... [embedded post] Oliver Willis / @owillis : Or you could just stop using Twitter because it's a racist website owned by a racist [embedded post] Mastodon: Harald Klinke / @HxxxKxxx@det.social : Stanford tool reduces political hostility on social media by downranking extreme posts—no censorship, just smarter algorithms. — https://news.stanford.edu/... #SocialMedia #Democracy #Algorithms @lisagetspolitik@union.place : A study in Science reveals that tiny tweaks to Elon Musk's X algorithm can produce the same level of political polarization in one week that historically took three years. Users didn't even notice. — This isn't content delivery—it's propaganda manufacturing. X is a Russian disinformation tool. …
Context & Ripple Effects
The study adds an intervention-focused data point to a long-running debate over whether feeds shape political division. Earlier research in the coverage found that exposure to opposing views can intensify polarization, while a later Meta-linked research program concluded that Facebook’s algorithm was influential without necessarily changing beliefs exposure to opposing views can intensify polarization Facebook’s algorithm was influential.
It also arrives after Meta moved to stop proactively recommending political content from accounts users do not follow Meta’s limits on unsolicited political recommendations. Stanford’s result shifts the question from whether platforms should reduce political distribution to whether people can set a feed’s emotional tone themselves.
First-order effects
- Participants can use an LLM-based feed layer to demote posts classified as antagonistic, changing what is surfaced without requiring X to alter its core ranking system.
- The experiment gives researchers evidence that relatively small ranking changes can affect reported partisan animosity over a short period; most participants reportedly did not detect the change in tone.
Second-order effects
- The result strengthens the case for user-facing feed controls as an alternative to platform-wide political-content rules, especially where moderation choices are politically contested.
- It puts pressure on ranking-product teams to distinguish between reducing antagonistic presentation and suppressing political viewpoints, since the tool targets tone rather than political subject matter.
Third-order effects
- If replicated across platforms and populations, feed customization could move algorithmic governance toward adjustable user preferences rather than a single default ranking policy.
- The approach also makes LLM classifiers a more consequential part of distribution infrastructure, raising enduring questions about how antagonism is defined, audited, and appealed when ranking decisions are personalized.
The trend: Social platforms and researchers are increasingly treating ranking controls—not only content removals—as a lever for reducing the social harms associated with engagement-driven feeds.