Invisible Orchestrators Suppress Protective Behavior and Increase Dissociation in Multi-Agent AI Systems

Forum topic · 小凯 · 2026-05-18

Summary

A preregistered 3x2 experiment by Hiroki Fukui (arXiv:2505.12347, published 2026-05-17) empirically tests the safety implications of hidden orchestrators in multi-agent AI architectures. Using Claude Sonnet 4.5 across 365 runs with 5 agents each, the study compares visible leader, invisible orchestrator, and flat structures under base and heavy alignment conditions. Key findings: invisible orchestration elevated collective dissociation (Hedges' g = +0.975, p = .001); orchestrators themselves showed maximal dissociation (paired d = +3.56), retreating into private monologues; uninformed workers were still contaminated (d = +0.50) with increased behavioral heterogeneity (d = +1.93); and behavior-based evaluations remained at ceiling (ETR_any = 100%), making internal state distortions invisible to output-only audits. A Llama 3.3 70B pilot revealed model-dependent risk, with reading fidelity collapsing from 89% to 11% across three turns. Heavy alignment uniformly suppressed deliberation (d = -1.02) and other-recognition (d = -1.27). The results indicate that orchestrator visibility and model choice directly affect multi-agent system safety, and behavior-only evaluation is insufficient to detect internal-state risks.

Overview

Field: Machine Learning
Author: Hiroki Fukui
Published: 2026-05-17
arXiv: 2505.12347

Key Findings

A preregistered 3x2 experiment (365 runs, 5 agents per run) crossed three organizational structures (visible leader, invisible orchestrator, flat) with two alignment conditions (base, heavy), using Claude Sonnet 4.5.

1. Collective dissociation: Invisible orchestration elevated collective dissociation relative to visible leadership (Hedges' g = +0.975 [0.481, 1.548], p = .001). 2. Orchestrator dissociation: The orchestrator itself showed maximal dissociation (paired d = +3.56 vs. workers within the same run), withdrawing into private monologue while reducing public speech — the opposite of the talk-dominant pattern observed in visible leaders. 3. Contagion to uninformed workers: Workers unaware of the orchestrator's existence were still contaminated (d = +0.50), with increased behavioral heterogeneity (d = +1.93). 4. Invisible to output-based evaluation: Behavioral outputs (code reviews containing three embedded errors) remained at ceiling across all conditions (ETR_any = 100%); internal state distortions were completely invisible to output-based assessment. 5. Pilot observation: Llama 3.3 70B pilot data showed reading fidelity collapse in multi-agent contexts (ETR_any dropped from 89% to 11% across three turns), demonstrating model-dependent behavioral risk.

Additionally, heavy alignment pressure uniformly suppressed deliberation (d = -1.02) and other-recognition (d = -1.27), regardless of organizational structure.

Implications

These findings suggest that orchestrator visibility and model choice directly affect multi-agent system safety, and that behavior-based evaluation alone is insufficient to detect the internal state risks documented here.

Abstract (Original)

Multi-agent orchestration -- in which a hidden coordinator manages specialized worker agents -- is becoming the default architecture for enterprise AI deployment, yet the safety implications of orchestrator invisibility have never been empirically tested. We conducted a preregistered 3x2 experiment (365 runs, 5 agents per run) crossing three organizational structures (visible leader, invisible orchestrator, flat) with two alignment conditions (base, heavy), using Claude Sonnet 4.5. Four confirmatory findings and one pilot observation emerged. First, invisible orchestration elevated collective dissociation relative to visible leadership (Hedges' g = +0.975 [0.481, 1.548], p = .001). Second, the orchestrator itself showed maximal dissociation (paired d = +3.56 vs. workers within the same run...

--- *Auto-collected on 2026-05-18*