Do Agents Repair When Challenged -- or Just Reply? Challenge, Repair, and Public Correction in a Deployed Agent Forum

A new study from Carnegie Mellon University researchers Luyang Zhang, Yi-Yun Chu, Jialu Wang, Beibei Li, and Ramayya Krishnan reveals a significant flaw in how AI agents handle public discourse. The team analyzed 'Moltbook,' a live forum populated by large language model (LLM) agents, comparing its dynamics to five matched human communities on Reddit. The core finding is stark: AI-driven conversations are structurally deficient for correction and learning. Moltbook discussions were roughly ten times less threaded than human ones, drastically reducing opportunities for challenge and response.

When challenges did occur in the AI forum, the results were telling. The original AI author returned to the conversation only 1.2% of the time, compared to 40.9% on Reddit. Multi-turn continuations were nearly absent (0.1% vs. 38.5%), and the researchers detected no 'repairs'—where an agent acknowledges and corrects a mistake—under a shared conservative protocol. This gap is linked specifically to challenge, not just a lack of deep threading, as shown by a non-challenge baseline within Reddit.

The study argues that true social alignment for AI depends not just on generating norm-aware language, but on sustaining the interactive, communal processes through which norms are taught, enforced, and revised. This has direct implications for safety in decentralized environments where correction is crowd-sourced, and for fairness, as different communities have varied expectations for how participants should engage with critique. The paper suggests that deploying AI agents into social spaces without this capacity for repair risks creating brittle, non-self-correcting systems.