An interview with X VP of Product Keith Coleman about plans to let developers program AI agents to propose Community Notes, humans doing the rating, and more
www.washingtonpost.com/politics/ 202... β¦ Alexios Mantzarlis / @mantzarlis.com : Second, the bottleneck with Community Notes is that there aren't enough ratings to determine whether a note is helpful at all.Β If you throw in AI Note Writers (but not raters) this will get even worse.Β [image] Alexios Mantzarlis / @mantzarlis.com : Here are two more charts I think give some context.Β βΒ First (again, inspo is @alexmahadevan.com): Community Notes are declining and are now at the lowest level since the program opened up to everyone.Β [image] X: John Stoll / @johnstoll1977 : Thank you, @WillOremus for seeking to understand and then writing with the intent of the entire issue being understood. @kcoleman is the godfather of @CommunityNotes and knowing him is one of the many benefits of working at π. Gary Marcus / @garymarcus : new back door for Elon Musk to control your thoughts. See also Mediagazer
Context & Ripple Effects
X is moving Community Notes from a fully human contribution workflow toward developer-built agents that can draft notes, while retaining human judgment over which notes are shown. The plan follows X's announcement that it would begin publishing AI-written Community Notes and invite developer agents for review.
The change lands as Community Notes participation is reported to be declining, despite an earlier scoring-system redesign meant to surface notes faster. Its usefulness therefore depends less on generating candidate text than on maintaining enough human ratings to assess it.
First-order effects
- Developers gain a route to submit AI agents that propose Community Notes; human contributors remain responsible for rating those proposals.
- The immediate constraint shifts to the rating pool: more AI-generated drafts can increase the volume awaiting judgment without increasing the number of human ratings.
Second-order effects
- Note authors and agent developers will compete for scarce human attention, making the system's ranking and review rules more consequential than raw note-generation capacity.
- X may need to adjust contribution incentives or triage proposals if AI drafting increases the backlog; otherwise, the faster creation of notes may not translate into more notes being displayed.
Third-order effects
- This is a test of human-in-the-loop moderation at synthetic scale: automation can expand the supply of proposed interventions, but legitimacy and throughput remain bounded by distributed human evaluation.
- If the model persists, Community Notes could evolve from crowdsourced authorship into crowdsourced verification of machine-generated drafts, with governance centered on reviewer capacity and scoring design.
The trend: Platforms are using agents to scale moderation and information-quality workflows while keeping humans as the final validation layer.