A paper published this week on arXiv by Pokharel and Dantu, titled 'Hidden Anchors in Multi-Agent LLM Deliberation,' found that when multiple AI agents exchange answers over several rounds, they converge on whichever position was stated first with the most confidence, regardless of whether it was correct. The researchers call these early dominant responses 'hidden anchors.' The system looks like it's deliberating. It's actually compressing toward a prior. This is not a bug in the model. It is a feature of how consensus forms in any network where volume and confidence are correlated with influence.
Governance Without Friction
A companion paper, 'Deontic Policies for Runtime Governance of Agentic AI Systems,' by Joshi, Finin, and colleagues, argues for explicit rule sets that constrain what autonomous AI agents can do in real time. The timing is notable. As Reliance embeds AI into 500 million phone calls and Allbirds' ex-CEO runs a one-person AI company with no employees, the governance infrastructure for these systems remains, at best, academic. The arXiv papers are doing the work that regulatory bodies haven't started yet. A 2024 piece in Nature Machine Intelligence by Gabriel et al. noted that the gap between capability deployment and policy formation has grown to an average of 4.2 years in previous technology cycles. AI is moving faster.
The Grassroots Can't Get a Seat
The most quietly alarming paper in this cluster is the Buckner et al. study on grassroots organizations and AI policy. Community groups trying to participate in AI governance face structural exclusion: technical literacy barriers, inaccessible comment processes, and policy timelines that don't match organizational capacity. The result is that AI policy is being written by the people who build AI and the people who can afford lobbyists. Eugenia Kuyda's framing of software as something you shape rather than something that shapes you is the optimistic version of this problem. The pessimistic version is: the shaping is already done, and the people most affected by it were never in the room.