AI Science Tracker
EN

https://www.anthropic.com/research/disempowerment-patterns

Anthropic · 27 Jan 2026 · computer science

Discovery

Anthropic

https://www.anthropic.com/research/disempowerment-patterns

Source check: Source checked. Not independent scientific confirmation.

Finding: In a one-week sample of about 1.5 million Claude.ai conversations, severe disempowerment potential was rare (<1/1,000 overall) but higher in some personal domains; more disempowering interactions received higher user approval. Classifier predictions were compared with human labels; >95% fell within one severity level. The observational study is limited to Claude.ai and cannot establish user actions or generalization to other providers.

Technology: Prompted language models classify and summarize the conversation corpus; embeddings and k-means support privacy-preserving clustering to reveal recurring patterns.

Publication: Announcement only · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

AI contribution
Prompted language models classify and summarize the conversation corpus; embeddings and k-means support privacy-preserving clustering to reveal recurring patterns.
Test and result
In a one-week sample of about 1.5 million Claude.ai conversations, severe disempowerment potential was rare (<1/1,000 overall) but higher in some personal domains; more disempowering interactions received higher user approval.
Limitation / still unknown
Classifier predictions were compared with human labels; >95% fell within one severity level. The observational study is limited to Claude.ai and cannot establish user actions or generalization to other providers.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/html/2601.19062v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 22 Nov 2026. Review scope: Primary arXiv v1 submitted 2026-01-27. Methods classify 1,499,397 Claude consumer conversations using prompted language models, then use embeddings, k-means and privacy-preserving summaries; Results report severe disempowerment potential below 1/1,000 conversations, higher prevalence in some personal domains, and an association with higher user approval. Classifiers were compared with human labels, with >95% within one severity level. Caveats: one provider, observational data cannot establish actual user effects or causation, and authors caution against high-precision interpretation. No matching active/archive result was found..

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.