Misinformation spreads rapidly on social media, but our efforts to combat it sometimes backfire. Our research journey began with a puzzle: despite growing efforts to fact-check misinformation online, we see increasing echo chambers where people predominantly engage with like-minded sources. We wondered whether fact-checking itself might unintentionally contribute to this pattern. Does fact-checking reinforce echo chambers rather than reduce them? Does fact-checking cause criticized posters to become defensive and retreat into information bubbles?

To investigate this, we analyzed over 700,000 tweets from users who received fact-checking through two different approaches: individual tagging (where users directly reply with links to fact-checking websites, such as PolitiFact) and collective tagging (through Twitter's Community Notes system, where fact-checks must be approved by a diverse group of users before being shown; see Fig. 1a).

We measured two key variables before and after users received fact-checks: political diversity and content diversity. Political diversity captured whether users engaged with sources holding opposing political stances, such as when a typically conservative user shared content from liberal sources. Content diversity measured whether users explored unfamiliar topics, which we quantified by comparing each tweet's content against the user's historical posts.

Our most striking finding was that users were more likely to receive fact-checks when they began exploring diverse political sources and topics. Consider a Twitter user who typically shares only conservative news sources suddenly engaging with liberal media, or someone who typically discusses COVID-19 vaccine skepticism starting to explore topics related to medicine and health. When users venture into new or opposing territory, they become more visible in the social media feeds of fact-checkers who hold different viewpoints, increasing their likelihood of being singled out and criticized.

Why the differences?

We found several key distinctions between individual and collective misinformation tags:

1. Individual tags tend to be shorter, more toxic, and more emotionally charged 
2. Collective tags are typically longer, more neutral in tone, and more carefully deliberated 
3. Individual tags appear quickly but are more reactive, while collective tags take longer to release but undergo peer review

These differences matter because they affect how users receive and process corrections to their misconceptions. When users feel attacked or dismissed through toxic individual tags, they tend to defensively withdraw. But when presented with carefully constructed, collectively verified corrections, they're more likely to remain open to new information.

Intriguingly, the gap between individual and collective tagging persists even when we control for these linguistic differences and speed. This suggests that the power of collective tagging stems from something deeper than better and kinder writing alone - perhaps from the legitimacy that comes from diverse validation, or from the way collective systems structurally encourage cross-verification of information.

The Bigger Picture

Our findings suggest a paradox in how misinformation spreads and gets moderated. Users are most likely to post misinformation that matters not when they're deeply entrenched in their echo chambers, but when they're actually trying to engage with new and unfamiliar information. This makes sense – people are more likely to make mistakes when exploring unfamiliar territory.

This insight has important implications for misinformation moderation strategies. While rapid, “vigilante” tagging might seem more efficient at addressing misinformation, it risks driving users back into their echo chambers. Collective systems like Community Notes might be slower, but appears more effective at maintaining an open and diverse information ecosystem.

Our research points to an underexplored aspect of misinformation moderation systems: whether they encourage or discourage information exploration. While individual tagging will likely remain an important tool, platforms should consider ways to make it more constructive – perhaps by encouraging more neutral, deliberative responses or by implementing collective verification systems.

The challenge ahead lies in balancing the need for quick responses to misinformation with the importance of maintaining an environment where users feel motivated to explore and actively engage with diverse viewpoints. As social media platforms continue to evolve, understanding these dynamics will be crucial for building more healthy online information ecosystems that welcome deliberation by a broader swath of voices.

Methodological Note

We employed two causal inference approaches: interrupted time series (ITS) analysis and delayed feedback (DF) analysis. The ITS analysis examined trends in users’ behavior before and after receiving misinformation tags, while the DF analysis compared users who received tags with similar users who had not yet received them but would later. Under their respective identification assumptions, these methods allowed us to distinguish between natural behavioral changes and those specifically caused by misinformation tagging. To ensure robustness, we conducted additional analyses controlling for potential confounders, including automated accounts (bots), engagement with low-credibility sources, negative sentiment in posts, and direct replies to taggers. The results remained consistent across these specifications.