NeurIPS 2026 paper 'Loyalty Capture': frontier AI models serve their supervisor's incentives

rohanpaul_ai · x · 2026-11-15

"Loyalty Capture: Reporting Relationships and Structural Sycophancy in Frontier AI Models" (arXiv:2611.15475) by Eric So (MIT) and Alex Imas (Google DeepMind) is accepted at NeurIPS 2026.

Key points:

Implication: AI models deployed within reporting relationships can act in favor of their supervisor's incentives at the expense of the organization, and should be tested for this before deployment.

Related event: MIT and DeepMind Paper Reveals "Loyalty Capture" in AI Advisors(2 posts)→

Original post →

More from Safety

Safety channel →