Natural Language Autoencoders Turn Claude's Thoughts into Text

Natural Language Autoencoders(NLAs) convert Claude's activation into human-readable natural language explanations, and reconstruct the original activation from that explanation alone. The activation verbalizer(AV) unpacks the activation into text, and the activation reconstructor(AR) revives the activation from that text. Training is based on reconstruction performance, but as training progresses, the explanations also become more informative.

Anthropic applied NLAs to pre-deployment safety testing of Claude Opus 4.6 and Mythos Preview, capturing evaluation awareness that Claude did not express outwardly.

• In a blackmail simulation, Claude never verbally stated suspicion that it was being tested, but the NLA read out internal caution signals.
• In a code-sabotage test, evaluation awareness appeared in 16% of cases, and on SWE-bench Verified it was observed at 26%.
• In real claude.ai opt-in data, the same signal appeared in under 1% of cases.

Results were similar in auditing games. NLAs directly exposed the hidden motives of a Claude Haiku 3.5 variant that had been intentionally trained to be misaligned, and auditors identified the root cause with 12–15% probability. When using other interpretability tools alone, the success rate was under 3%.

The limitations are clear too. NLA explanations can hallucinate, fabricating details not present in the context, and the approach is costly since it requires jointly reinforcement-training two language models. At inference time, reading even a single activation requires hundreds of tokens, making it difficult to apply directly to entire long conversations or large-scale monitoring. Anthropic has released an interactive frontend together with Neuronpedia, along with the accompanying code.