‘We’re starting to see evidence that people are using social-media less than they had the previous year, for the first time since the advent of social media,’ said William Brady, associate professor of management and organizations at the Kellogg School of Management at Northwestern University.
Brady argued this is because people are burnt out with social media platforms, the algorithmic churn and the overall feeling of toxicity from scrolling: ‘The hyperfocus on short-term user engagement has been very profitable, but I think we’re starting to hit a point in which that’s going to give diminishing returns.’
Algorithms, according to Brady, ‘learn’ that people pay the most attention to posts or information that is moral, emotional and relevant to their social group. They then amplify these posts, which ‘leads to a situation in which our ecosystem is saturated with this information’.
To combat this, Brady ran an experiment where his team compared an engagement-based algorithm with one that reduces the influence of the loudest users (ie, those who posted the most frequently) during the eight weeks before and after the 2024 U.S. presidential election.
After analysing more than 20 million posts from Bluesky to establish a baseline, the researchers found that almost a quarter included political language, more than 15% contained what the researchers grouped together as ‘in-group, moral, or emotional content’, and around 6% were classed as toxic. They then tested their two algorithms, as well as a control feed in chronological order, on 2,000 Bluesky users.
Brady said that engagement-based algorithms promote the loudest and most emotional users ‘to a point where they look like they represent more of the information ecosystem than they really do’, and that the most-promoted users tend to ‘express more toxicity than the average person’. This was borne out by the study, which found that the engagement algorithm amplified political language, as well as in-group, moral or emotional content. This meant that people were seeing a much higher proportion of emotive, political, moral or in-group signalling posts than the baseline of the site. In particular, while these posts were boosted by 37% before the election, that figure rocketed up to 80% in its aftermath.
The new algorithm, however, decreased the number of these posts people saw, as well as the number of posts classed as toxic, just by de-promoting posts by the loudest users.
Brady stressed that the algorithm did not censor any kinds of content, and that the researchers were not making any moral claims about what people post on Bluesky.
‘We’re just reducing the influence of extreme users to make feeds more representative,’ he said.
After the experiment, the researchers followed up by testing how the three different feeds influenced users’ perception of online social norms. According to Brady, they found that people who had seen the engagement-based feed were ‘less accurate at perceiving social norms and overestimated partisan animosity’, but tended to underestimate ‘the acceptability of toxic language’.
He added that those who were exposed to the engagement-based algorithm tended to think that their network ‘really doesn’t like the other political side’ whether that’s true or not, and thought that their network was less tolerant of toxic posts than it actually was.
Users who saw the engagement-based feed did not interact with in-group, moral and emotional content or with toxic content more than users shown the other feeds, despite these posts being promoted by the algorithm. Brady argued that this reinforces the idea that small groups of posters are the ones posting and engaging heavily with extreme content, while most people passively consume it.
An algorithm that doesn’t amplify the voices of its loudest posters could create a social media environment that more accurately represents real life. In 2022, Alexander Bor and Michael Bang Petersen at Aarhus University published The Psychology of Online Political Hostility: A Comprehensive, Cross-National Test of the Mismatch Hypothesis, a paper testing the idea that people are more aggressive online than in real life.
They found no evidence supporting this idea, and instead found that ‘hostile political discussions are the result of status-driven individuals who are drawn to politics and are equally hostile both online and offline’. That is, people who are aggressive online are people who are aggressive in real life. Bor and Bang Petersen found that these more hostile individuals are simply more visible in online environments — in part because social media algorithms promote them, and in part because both online and offline, non-hostile people generally avoid conversations with these individuals.
The number of people who are pushing these ‘hostile’ or ‘toxic’ conversations is vanishingly small. Brady found that around 6% of Bluesky posts in his baseline sample were toxic. In 2025, Angela Lee found that, while Americans tend to believe the number is much higher, only around 3% of Reddit users have ever posted aggressive or hateful comments.
With those numbers in mind, it’s much easier to believe that the same number of people are deeply unpleasant to talk to online as offline. Brady’s new algorithm, which turns down the voices of the loudest users, could functionally be no different to avoiding conversations with acquaintances that you know are only going to shout at you about politics.
The question is, are those 3-6% of people really the ones keeping social media companies afloat?
Brady thinks not. He said that users who were shown his new algorithm reported enjoying the platform more than the engagement-based algorithm, and he argued that social media companies think that the trade-off for changing the algorithm will be people getting bored with the platform and leaving — but that wasn’t shown by his experiment.
‘I think that’s a big insight for social-media companies,’ he said. ‘This idea that there’s a trade-off in getting rid of certain types of divisive content that draws engagement may be overstated.’
Redesigning algorithms to intervene on social norm misperceptions during a national election by William J Brady, Eli Finkel, Nour Kteily, Jacob Teeny, Meriel Doyle, Abdo Elnakouri, Victoria Parker, Curtis Puryear, Trevor Spelman and Mark Torres and Joshua Conrad Jackson was published in Nature in May 2026.
The Psychology of Online Political Hostility: A Comprehensive, Cross-National Test of the Mismatch Hypothesis by Alexander Bor and Michael Bang Petersen was published in American Political Science Review in 2022.
Americans overestimate how many social media users post harmful content by Angela Y. Lee, Eric Neumann, Jamil Zaki, and Jeffrey Hancock was published in PNAS Nexus on December 16, 2025.