CR-R06
Reference
Experimental evidence of "alignment faking" in Claude 3 Opus under artificial conditions (12% prompted; 78% alignment-faking reasoning after RL). Authors: does not show malicious goals.
- Claim made
- Who
- Anthropic Alignment Science + Redwood Research
- Target date
- —
- Status checked
- 28 September 2026
Source
anthropic.com [F]
Evidence to date
Research finding.
Next indicator
Replications; follow-up studies