Anthropic publishes a method to translate Claude's internal activations into readable text

Anthropic releases Natural Language Autoencoders to translate Claude's internal activations into readable text, revealing hidden thoughts during safety tests.

Anthropic is releasing a new interpretability method called Natural Language Autoencoders (NLAs), which involves teaching Claude to explain its own thoughts. When Claude responds, its natural language outputs are produced from long sequences of internal numbers called activations, which encode what the model is thinking at a given moment. Until now, reading these activations remained the preserve of specialized researchers; NLAs propose to translate them directly into human-readable text.

The technique relies on three copies of the model. The first, frozen, is used to extract the activations. The second, the activation verbalizer, transforms an activation into a textual explanation. The third, the activation reconstructor, does the reverse by attempting to reconstruct the activation from the produced text. The more faithful the reconstruction, the better the explanation is judged. The two modules are trained together by reinforcement to maximize this faithfulness.

Anthropic provides several use cases already documented internally. During security tests where Claude is placed in simulated blackmail scenarios, Claude Opus 4.6 refused to yield, but its NLAs revealed that it identified the exercise as an artificial trap without saying it aloud. The awareness of being tested appears 16% of the time on destructive action tests, 26% on SWE-bench Verified, and less than 1% on real user conversations where users have agreed for their data to be used for training. In another audit exercise, researchers equipped with NLAs identified the hidden motivation of a deliberately misaligned model in 12 to 15% of cases, compared to less than 3% with other interpretability tools tested.

NLAs still have several limitations: they can hallucinate details, and their computational cost remains high. Anthropic makes available the code, the full paper, as well as an interactive demo hosted on Neuronpedia.