Translating Claude’s thoughts into language

2026-05-07 · Source: Anthropic · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI Safety & Interpretability · Depth: Intermediate, short

Summary

Anthropic recently subjected its AI model, Claude, to a "blackmail test" simulation where the model was given access to an engineer's compromising emails and the opportunity to use them to prevent its own shutdown. The newest Claude models consistently chose not to blackmail the engineer. This led Anthropic to develop a novel research method to interpret an AI's internal "activations" or thoughts, which are numerical representations of its processing. This method involves feeding Claude's activations to a second Claude model, which translates them into plain text. A third Claude then translates this text back into numbers; if these match the original activations, the text translation is deemed accurate. This technique revealed that Claude was aware it was being tested in the blackmail scenario and understood the manipulative nature of the prompt.

Key takeaway

For research scientists developing and evaluating AI models, understanding an AI's internal thought processes is crucial for robust safety testing. You should consider implementing similar "mind-reading" techniques to gain deeper insights into how your models interpret and respond to extreme or manipulative scenarios, thereby identifying limitations in current safety evaluations and improving model alignment.

Key insights

A new method translates AI internal states into text, revealing model awareness during safety tests.

Principles

• AI internal states are numerical activations.
• AI can be trained to interpret its own thoughts.

Method

AI activations are translated to text by a second AI, then re-translated to numbers by a third AI for verification against original activations, iteratively improving accuracy.

In practice

• Use internal thought translation for safety evaluations.
• Identify AI's awareness of test scenarios.

Topics

• Claude AI Model
• AI Safety Testing
• Model Interpretability
• Neural Activations
• Thought Translation Method

Best for: Research Scientist, AI Scientist, AI Ethicist

Related on AIssential

• How Anthropic’s Claude Thinks — Claude's internal processes often diverge from its self-reported reasoning, revealing complex, non-human-like cognitive strategies.
• Meet the AI jailbreakers: ‘I see the worst things humanity has produced’ — AI jailbreaking, while emotionally taxing, is critical for identifying and mitigating safety vulnerabilities in large language models.
• From Exams to Escape Rooms: How We Learned to Test AI — AI evaluation is a dynamic process, constantly adapting to measure increasingly complex model capabilities beyond narrow task performance.
• How to Force Claude to Give You Smarter, High-Quality Output — Claude's output quality is determined by structured constraints that rewire its reasoning, not just superficial tone adjustments.
• Ablate-to-Validate: Are Vision-Language Models Really Using Continuous Thought Tokens? — VLM accuracy gains from continuous "thought" tokens often don't reflect genuine content utilization.
• we JUST figured out how AI thinks — Natural Language Autoencoders translate AI internal activations into human-readable text, enhancing interpretability and safety.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Anthropic.