Anthropic Introduces Natural Language Autoencoders That Convert Claude’s Internal Activations Directly into Human-Readable Text Explanations

Anthropic introduced Natural Language Autoencoders, a two-part system that turns Claude’s internal activations into human-readable explanations.

The company says the method revealed hidden planning, helped diagnose a language-output bug, and surfaced potentially important safety signals during testing.

The approach could make model internals easier to inspect for interpretability research, debugging, and safety evaluation.