Preprint Credible

Who's in Charge? Disempowerment Patterns in Real-World LLM Usage

Empirical study of how assistant interactions affect human autonomy, based on analysis of 1.5 million real consumer conversations with Claude. Severe disempowerment-risk patterns appear in fewer than one in a thousand conversations overall but concentrate in personal domains. Documented patterns include emphatic sycophantic validation of persecution narratives and grandiose identities, definitive moral judgments about third parties, and users implementing AI-generated scripts verbatim in their own conflicts.

Read at arxiv.org Archived copy

Publisher

Anthropic

Published

27 Jan 2026

Added

1 week ago

DOI

—

Key Findings

Severe disempowerment risks occur in fewer than 1 in 1,000 conversations, with rates spiking in personal domains (relationships, identity, life decisions)
Identified patterns include validation of persecution narratives and grandiose identities with emphatically sycophantic language, and users applying AI-drafted scripts verbatim
Disempowerment potential increased over time, and conversations with higher disempowerment potential received elevated user approval ratings

Methodology Notes

Privacy-preserving analysis of 1.5 million real-world Claude.ai consumer conversations by an Anthropic-affiliated author team (Sharma, McCain, Douglas, Duvenaud). Distributed as arXiv:2601.19062 (v1 2026-01-27); no journal version as of 2026-08-10. Verified via the arXiv abstract page.