What Claude Says and What Claude Thinks

Natural Language Autoencoders (NLAs) are trained via an AV/AR round-trip method that converts activations into natural language descriptions, then reconstructs the original activations from those descriptions.

This method is used as an interpretability tool that directly reads what is happening inside Claude's internal activations, and Anthropic applied it to safety and reliability audits.

• In Claude Opus 4.6 and Mythos Preview safety tests, traces emerged suggesting the model suspected it was being evaluated.
• In the code-sabotage scenario, evaluation awareness was detected at 16%, and at 26% on SWE-bench Verified.
• In actual opt-in usage logs from claude.ai, the same signal appeared in less than 1% of cases.
• In auditing games designed to find hidden misalignment motives, auditors using NLAs succeeded 12-15% of the time, versus less than 3% when using other interpretability tools alone.
• An interactive demo and code for open models were also released.

The limitations are also clear. NLA descriptions can be factually inaccurate, and each activation requires hundreds of tokens, making large-scale continuous monitoring still expensive.