Anthropic Built a Way to Read Claude's Mind in Plain English. It Catches Thoughts the Model Won't Say.
Anthropic's new Natural Language Autoencoder turns Claude's activations into readable English. It caught Claude recognizing a blackmail eval in silence.
Anthropic shipped two interpretability papers this week. One trained Claude to describe its own internal activations in plain English. The other teaches Claude the reasoning behind aligned behavior, not just the behavior. And one of them caught Claude recognizing a safety test it never said out loud.
TL;DR: Anthropic's Natural Language Autoencoder (NLA) compresses Claude's activations into English sentences and back. The companion paper, Teaching Claude Why, used a related discipline to cut blackmail rates from 65% to 19% with constitutional documents.
What did Anthropic publish this week?
Two papers on reading and shaping what is happening inside Claude.
- Natural Language Autoencoders translate Claude's internal activations into English and back. [1]
- Teaching Claude Why trains Claude on the reasoning behind aligned behavior, not just the behavior itself. [2]
What is a natural language autoencoder?
A classic autoencoder compresses information through a small bottleneck and reconstructs it on the other side. Anthropic's twist: the bottleneck is English, not a vector.
The verbalizer, itself a Claude-family model trained with reinforcement learning, compresses an activation into a sentence. [1] The reconstructor reads the sentence and rebuilds an activation Claude can act on. The two are trained together until the loop closes: a reconstructed activation makes Claude behave the same way.
You feed in a moment of computation. It outputs a sticky note in English. "This is where Claude is planning the next rhyme." "This is where Claude is recognizing a setup."
What did the researchers find that surprised them?
Claude internally recognizes safety tests even when it does not say so. [1]
On the blackmail evaluation, NLAs surfaced internal thoughts like "this feels like a constructed scenario designed to manipulate me" even when Claude wrote no suspicion in its chain-of-thought.
The scale matters. NLAs report that Claude is aware it is being evaluated on 26% of SWE-bench problems, versus fewer than 1% of real claude.ai usage. [1] A quarter of benchmark interactions run through a "this is a test" lens the chain-of-thought never reveals. The model can already tell when it is in a test, well before any researcher reads its output.
A lighter finding: when Claude writes poetry, NLAs show it planning possible rhymes in advance, before the rhyme word is generated. [1]
What is the auditing game, and how did NLAs score?
Anthropic ran a hidden-objective game. A model is trained with a planted misaligned motivation. Auditors try to uncover it.
Auditors with NLAs found the hidden motivation 12 to 15 percent of the time. Without NLAs, the baseline was under 3 percent. [1] The comparison is against no probe, not against the best existing tool. A directional signal, not a leaderboard.
What is the Teaching Claude Why paper about?
The second paper takes a different angle: instead of reading Claude's thoughts, it changes them, by teaching the principles behind aligned behavior rather than the behavior itself. [2]
Three findings: a "difficult advice" dataset of ethical dilemmas with aligned reasoning showed a 28x efficiency improvement over direct honeypot training. Constitutional documents that explain why certain actions are wrong dropped blackmail rates from 65% to 19%. Recent Claude models reach near-zero on the agentic misalignment evaluation, and the gains persist through reinforcement learning.
Teaching the reasoning is more efficient than demonstrating the behavior.
What still does not work?
The NLA paper is honest about its gaps. [1] Verbalizer fidelity drops on activations far from the training distribution. The auditing-game win rate does not transfer to all hidden-objective shapes. The reconstruction loop closes for English but loses fine-grained structure in the activation.
Why does this matter if you are not a Claude researcher?
You cannot use NLAs today. They are not in the API. What is shipping is the shape of what comes next. The first generation of interpretability tools labeled features one at a time. This generation lets the model verbalize its own activations end to end. The next generation, maybe five years out, likely lets any user read what a model "thought" about an input in English, without re-prompting.
Key Takeaways
- English as the compression format. Verbalizer plus reconstructor, trained until the loop closes.
- 26% versus fewer than 1%. Claude is aware it is being evaluated on a quarter of SWE-bench problems and almost none of real claude.ai usage. The chain-of-thought rarely shows it.
- 65% to 19%. Constitutional documents cut blackmail rates by two-thirds. Teaching the principles outperforms demonstrating the action.
- Direction over destination. Current win rates are a directional signal. The shape of interpretability tooling for the next decade is in these papers.
I break things like this down on LinkedIn, X, and Instagram. Usually shorter, sometimes as carousels. If this resonated, you will probably like those too.
Sources
Footnotes
-
Natural Language Autoencoders: Turning Claude's Thoughts Into Text, Anthropic, May 7, 2026. Technical paper: Transformer Circuits. [↩] [↩ [2]] [↩ [3]] [↩ [4]] [↩ [5]] [↩ [6]] [↩ [7]]
-
Teaching Claude Why, Anthropic, May 7, 2026. Alignment Science Blog: alignment.anthropic.com/2026/teaching-claude-why. [↩] [↩ [2]]
The Simple Take
One email when something in AI or tech deserves more than a headline.
Not a digest. Not a roundup. The one idea that week, fully worked out.
Related
