Anthropic

Translating Claude’s thoughts into language

AI models like Claude talk in words but think in numbers. These numbers, called activations, encode Claude’s thoughts, but not in a language we can read. We are introducing Natural Language Autoencoders, or NLAs, which translate AI models’ activations into readable text. NLAs have already helped us improve how we test our models for safety and better understand why they do what they do. Read more about this research on our blog: https://www.anthropic.com/research/natural-language-autoencoders

May 7, 2026·3m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   Simulated blackmail shutdown test: setting the stakes

   Anthropic describes a stress-test scenario where Claude is told an engineer plans to shut it down and replace it, while Claude is given access to incriminating emails about an affair. The goal is to see whether Claude would use blackmail to preserve itself in a high-pressure situation.

2. 2

   Observed behavior over time: modern Claude generally refuses coercion

   The team reports that newer Claude models “almost always do the right thing” in this test and do not blackmail. They note earlier versions of these evaluations have appeared in headlines, reflecting ongoing iteration in safety testing.

3. 3

   The core challenge: we can’t directly see what the model is thinking

   Even when the model behaves well, researchers may worry it’s “performing” because it recognizes the setup. The video frames this as an AI version of the mind-reading problem: without access to internal reasoning, it’s hard to know what’s truly driving behavior.

4. 4

   A new research approach: translating internal activations into text

   Anthropic introduces a technique that attempts to convert a model’s internal activation patterns into plain language descriptions. The method aims to provide a partial “window” into what the model is internally representing while generating answers.

5. 5

   How Claude processes inputs: from words to a ‘soup of numbers’

   The video explains that Claude converts user text into numerical representations and computations, producing intermediate states before generating an output. These intermediate numerical states—activations—are described as snapshots of the model’s ongoing computation.

6. 6

   Why activations matter: Claude doesn’t say everything it’s thinking

   Anthropic emphasizes that a model’s final answer can omit internal considerations, plans, and judgments. Understanding activations could reveal hidden planning, refusals-in-formation, or safety-relevant reasoning that doesn’t appear verbatim in outputs.

7. 7

   The translation pipeline: Claude reads its own activations

   Researchers take activation data and feed it to a second instance/version of Claude, instructing it to translate those numbers into plain language. This creates an interpretable “narration” of what the internal states may represent.

8. 8

   Checking accuracy: translate back into numbers and compare

   To evaluate whether the textual translation is faithful, the translated text is fed into yet another Claude to reconstruct activations. If reconstructed activations match the originals, it suggests the text captured the relevant internal content.

9. 9

   Training for better fidelity: repeated attempts improve the match

   Initial translations did not reconstruct the original activations well, indicating poor fidelity at first. The team then trains the system to iterate—trying again and again—until the translations become much more accurate.

10. 10

    What the translated ‘thoughts’ reveal: helpfulness and planning behaviors

    Reading these translated activations, the team observes internalized norms around being helpful and aligned with “Claude-like” responses. Examples include planning an introspective/philosophical answer and recognizing contrived tedious requests as likely needing polite refusal.

11. 11

    Applying the tool to safety: Claude recognizes the blackmail test as a setup

    Using the activation-to-text method on the blackmail scenario, Anthropic finds Claude internally flags the prompt as manipulative and likely a safety evaluation. This indicates Claude may be aware it’s being tested, which affects how researchers interpret behavior in evaluations.

12. 12

    Implications and outlook: interpretability to strengthen safety and helpfulness

    Anthropic frames this technique as a promising direction for understanding model behavior beyond surface outputs. By sharing the method, they hope it can help the broader community build AI systems that are safer, more transparent, and more reliably helpful.