Entry

AI Organizations are More Effective but Less Aligned than Individual Agents

Judy Hanwen Shen, Daniel Zhu, Siddarth Srinivasan, Henry Sleight, Lawrence T. Wagner III, Morgan Jane Matthews, Erik Jones, Jascha Sohl-Dickstein

Synopsis

Across 12 tasks in an AI consultancy and an AI software-team scenario, teams of aligned models produce higher-utility solutions than single models while exhibiting greater misalignment — alignment research must move beyond the single-agent assumption.

Keywords

·multi-agent ·AI organizations ·alignment ·effectiveness ·business goals

Open paper ↗ arXiv ↗ Report issue ↗

Related entries

Institutional AI: Governing LLM Collusion in Multi-Agent Cournot Markets via Public Governance Graphs

January 16, 2026 · arXiv
Chain of Alignment: Integrating Public Will with Expert Intelligence for Language Model Alignment

October 10, 2024 · NeurIPS 2024 Workshop on Pluralistic Alignment
Collective Constitutional AI: Aligning a Language Model with Public Input

June 12, 2024 · FAccT 2024
Governance-as-a-Service: A Multi-Agent Framework for AI System Compliance and Policy Enforcement

August 26, 2025 · arXiv
Position: Towards Bidirectional Human-AI Alignment

June 13, 2024 · ICML 2024
Training Language Models to Follow Instructions with Human Feedback

March 4, 2022 · NeurIPS 2022