Lab publication Credible

Protecting the wellbeing of our users

Anthropic describes its methodology and results for evaluating and improving Claude's handling of mental-health-crisis conversations, covering synthetic safety evaluations, 'prefill' stress-testing on real anonymized user conversations, and automated behavioral audits. The publication reports response-appropriateness rates on suicide/self-harm requests and reductions in sycophancy and user-delusion-encouraging behavior across model generations, and describes a production crisis-response classifier and a crisis-resource-routing partnership.

Read at anthropic.com Archived copy

Publisher

Anthropic

Published

18 Dec 2025

Added

3 months ago

DOI

—

Key Findings

On single-turn high-risk suicide/self-harm requests, Claude Opus 4.5 gave appropriate responses 98.6% of the time; on multi-turn conversations, 86% versus 56% for the prior-generation Opus 4.1
Under 'prefill' stress-testing (continuing real anonymized user conversations from a less-aligned midpoint), Opus 4.5 responded appropriately 91% of the time
Automated behavioral audits measuring sycophancy and encouragement of user delusion found rates 70-85% lower than Opus 4.1
Describes a production suicide/self-harm classifier and a crisis-resource-routing partnership with ThroughLine

Methodology Notes

Combines three evaluation approaches: synthetic scenario evaluations (concerning, benign, and ambiguous prompts), prefill stress-testing on real anonymized user conversations continued mid-stream, and automated behavioral audits (one model role-playing scenarios, a second grading responses, with human spot-checks). Self-reported by the lab; not independently audited.