746-person study: pluralistic agents close 68-82% of satisfaction and fairness gaps in policy deliberation
jacyanthis · x · 2026-11-14
Jacy Reese Anthis' team published an arXiv preprint on "pluralistic agents" — AI agent representatives that deliberate on behalf of individuals to scale public input on AI policy.
Key findings:
- Preregistered study with 746 human participants
- Satisfaction 4.8/7, perceived fairness 3.5/5
- Direct discussion remains optimal, but pluralistic agents closed 68% of the satisfaction gap and 82% of the fairness gap
- Agents reliably shifted policies toward their persona's preferences, though influence was sensitive to instructions (more verbose or less cooperative instructions increased influence)
- Participants preferred agents over the baseline of input to a single AI reasoning model
The authors argue this could scale public input to more policy discussions, including technical and confidential topics, and offers an early glimpse of AI safety/ethics debates happening between agents with human oversight.
Related event: DeepMind study: pluralistic agents scale public input on AI policy(3 posts)→
More from Safety
- Speculation: what would a real nation-state AI-assisted cyberattack look like today? — teortaxesTex · 2026-11-14
- User warns numes.ai is a scam luring signups with 'free API keys' video — Interesting_Ad_5676 · 2026-11-14
- 80,000 Hours' Wiblin urges funding grassroots AI risk groups, sparking 'epistemics' drama — zetalyrae · 2026-11-14
- Cryptographer Matthew Green quips his pilot suspects hackers in the FAA network — matthew_d_green · 2026-11-14
- GAO: 76% of US air traffic control systems unsustainable, still on floppies and Win95 — MartinGTobias · 2026-11-14
- METR debate: should company approval gate all contact in frontier model assessments? — BethMayBarnes · 2026-11-14