The longer people talk about the same thing online, the more negative they get. A study of more than 2 billion Reddit comments found that this slide happens not because users want to be nasty, but because negativity is the easiest way left to say something new.
The work comes from Hongkai Mao, Yuan Chang Leong, Yutong Jiang, Alex Koch, William J. Brady and Joshua Conrad Jackson, with affiliations at the University of Chicago Booth School of Business, Stanford, and Northwestern’s Kellogg School of Management. It was published in the Proceedings of the National Academy of Sciences on June 22, 2026.
The case for “differentiation”
The authors start from a simple idea. Social media rewards users who stand out, but every new comment in a thread makes it harder to say anything original. Negative statements help because there are more varied ways to be negative than positive, and because positive comments are the norm, so a negative one is more noticeable.
The researchers framed it with Tolstoy’s line about families: happy ones are alike, while each unhappy family is unhappy in its own way.
What 2.05 billion comments showed
The team analysed 2.05 billion comments from 2,150 topic-based subreddits posted between 2010 and 2022. In a sample of 115,961 threads, comments posted later in a thread were more negative than earlier ones.
The shift within threads was small but consistent. Among the first five comments, 25.21% were negative, compared with 25.96% among the final five, and negative comments also grew more intensely negative.
Whole communities drifted too. When each subreddit’s monthly tone was tracked over time, 62% became more negative and 38% became more positive. Every cohort of subreddits, grouped by year of creation, moved toward negativity.
The key test was uniqueness. Using a language model to measure how far each comment sat from the thread’s average meaning, the authors found later comments were more distinctive, and more negative comments were more distinctive as well. Once uniqueness was included in the model, the link between a comment’s position and its negativity was no longer statistically significant.
Bigger communities were also more negative, which fits the idea that crowds make it harder to stand out. Accounting for semantic diversity cut that size effect by 28%.
The drift was strongest where conversations began on a happy note. Threads and communities that started out negative actually became slightly more positive over time.
To rule out a general souring of public mood, the authors compared Reddit with 4.91 million news articles from the same period. The news showed no trend toward negativity.
An experiment with 4,000 commenters
Because the Reddit data cannot prove cause, the team ran a preregistered experiment with 4,000 Prolific participants. Over five “generations,” people read earlier comments on a news headline and added their own, and a language model rated each comment’s tone.
Half were offered a $5 prize for writing the most unique comment, and half for writing the comment most similar to earlier ones. Under the uniqueness incentive, comments grew more negative across generations, most steeply for the most positive headline, about atomic bombing survivors winning the 2024 Nobel Peace Prize. Under the conformity incentive, the pattern reversed and tone drifted upward.
The authors acknowledge that the observational data alone cannot settle the cause. Communities may form during bursts of enthusiasm that naturally fade, an explanation they consider less likely but cannot fully exclude.
Their framing places part of the problem in platform design: threads that keep thousands of people on a single topic push every new commenter to find a fresh angle, and the darker angles are easier to find.
Study Details:
- Title: Differentiation drives the erosion of positivity on social media
- Authors: Hongkai Mao, Yuan Chang Leong, Yutong Jiang, Alex Koch, William J. Brady, Joshua Conrad Jackson
- Journal: Proceedings of the National Academy of Sciences
- Publication Date: June 22, 2026
- DOI: 10.1073/pnas.2527316123
