8 AI bots now write 50% of X’s Community Notes

A handful of anonymous bots have taken over fact-checking on X.

This isn’t hyperbole. In the first three weeks of May, just eight AI contributors wrote 50.3% of all visible Community Notes on the platform. 

When the program launched five years ago, X positioned it as a way to harness the knowledge of users to add context and corrections to tweets. It described Community Notes as a way to “broaden the range of voices” fighting misinformation. 

Then, in July 2025, X started allowing people to use automation to mass-generate and submit AI notes. That new feature made it inevitable that human writers would eventually be overtaken by bots. But it happened faster than I predicted. 

Community Notes was supposed to bring scale and legitimacy to the fight against misinformation by empowering users to participate in fact-checking. Now that power is increasingly in the hands of a tiny group of hobbyists and researchers.

AI contributors aren’t just replacing the humans who are supposed to make up the community in Community Notes.¹ They are publishing in languages and on topics that reflect the bias of their algorithmic setup. A project meant to humanize trust and safety has been delegated to bots and the people who run them.

Keith Coleman, VP of product at X, told Indicator that “people clearly value notes from AI writers, as shown by the growing number appearing on X.” Coleman said Community Notes power users appreciate them too and have made product feature requests for AI contributors that X has implemented.

What happens next will matter not just for X, or for the copycat Community Notes features on Meta and TikTok.² For good and for bad, X is currently running the world's largest live experiment in automated fact-checking. How it fares may define platform interventions against misleading content for years to come.

To better understand the program’s trajectory, I spoke to three researchers testing AI contributors, two human supercontributors, and several current and past X employees. I also used X’s (admirably) open dataset to compare several thousand tweets fact-checked by human and AI writers and scrape the sources and languages used in the nearly 61,000 helpful notes published since synthetic contributors were onboarded.