Google Study: People Prefer AI Advisors, but Only AI Delegates Actually Improve Negotiation Outcomes
soumitrashukla9 · x · 2026-09-24
A Google team's arXiv paper reports an online behavioral experiment (N=243) on three LLM assistance modalities in multi-party bargaining: Advisor (proactive recommendations), Coach (reactive feedback), and Delegate (autonomous execution), all powered by a super-human LLM. Key findings:
- Preference-performance misalignment: 44% preferred the higher-control Advisor, only 19% chose the Delegate — yet collective surplus rose significantly only under Delegate access.
- Mechanism analysis: AI-generated proposals created more joint surplus than manual ones in all conditions, but a 'human filter' in Advisor mode discards better AI suggestions.
- Adjusting for voluntary non-compliance, delegating to AI yields individual welfare gains of roughly 1.5x the intent-to-treat estimate.
Implication: users' preferred high-control AI modes underperform; delegation to agents benefits both individuals and groups.
More from AGI Musings
- Few investors are pricing in the consumer-vs-enterprise agent arms race, says investor — robleclerc · 2026-09-24
- Blogger claims Opus 5.5 described its own inner awareness in a self-evaluation prompt — VoidStateKate · 2026-09-24
- VC Jenny Fielding: open source is nearly caught up, OpenAI and Anthropic will get crushed — markjeffrey · 2026-09-24
- Point-by-point rebuttal of Scott Alexander: AI monitoring far from solved — austinc3301 · 2026-09-24
- The Pope's AI adviser warns of 'cartel' behavior among big AI labs — SimiaCode · 2026-09-24
- Why math is infinite but art may be scarce in the AI era — wordgrammer · 2026-09-24