Search Anthropic

Tools that explain why models behave as they do, so issues in model behavior are found and addressed as they arise.
A global workspace in language models
New interpretability research reveals an emergent mental workspace in Claude that holds internal thoughts that don’t appear in the model’s output.
Anthropic co-founder Chris Olah's remarks on Pope Leo XIV's encyclical "Magnifica humanitas"
The full text of Chris Olah's remarks on the Pope's encyclical on AI.
Natural Language Autoencoders: Turning Claude’s thoughts into text
AI models like Claude talk in words but think in numbers. In this study, we train Claude to translate its thoughts into human-readable text.
Emotion concepts and their function in a large language model
All modern language models sometimes act like they have emotions. What’s behind these behaviors? Our interpretability team investigates.
A “diff” tool for AI: Finding behavioral differences in new models
"Model diffing" is a powerful way to understand how models change during fine-tuning. Our new research paper extends this method to its most challenging and general use case: comparing models with entirely different architectures.
The assistant axis: situating and stabilizing the character of large language models
In a new paper, we look at several open-weights language models, map out how their neural activity defines a “persona space,” and situate the “Assistant” persona within that space.
Next-generation Constitutional Classifiers: More efficient protection against universal jailbreaks
Constitutional Classifiers++ improve on our previous approach, yielding a system that is even more robust, has a much lower refusal rate, and is dramatically cheaper to run.
Signs of introspection in large language models
Our new research provides evidence for some degree of introspective awareness in our current Claude models, as well as a degree of control over their own internal states.
Persona vectors: Monitoring and controlling character traits in language models
In a new paper, we identify patterns of activity within an AI model’s neural network that control its character traits, which we call persona vectors.
Open-sourcing circuit tracing tools
We recently introduced a new method to trace the thoughts of a large language model. Today, we’re open-sourcing the method so that anyone can build on our research.
Tracing the thoughts of a large language model
Today, we're sharing two new papers that represent progress on the development of an “AI microscope,” and the application of it to see new “AI biology.”
Auditing language models for hidden objectives
A collaboration between Anthropic's Alignment Science and Interpretability teams
Insights on crosscoder model diffing
Preliminary interpretability work on crosscoder model diffing, a method for understanding how two models differ.
Evaluating feature steering: A case study in mitigating social biases
Our preliminary (mixed) findings on how feature steering might play a role in creating safer model outputs.
Using dictionary learning features as classifiers
An interpretability update on using dictionary learning features as classifiers for detecting content or behaviors in model activations.
Circuits Updates – September 2024
Preliminary interpretability notes on successor heads and on oversampling topics in sparse autoencoder training for safety-relevant features.
Circuits Updates – August 2024
Preliminary interpretability notes on quantified autointerpretability evaluations for sparse autoencoders and self-explaining features.
Circuits Updates – July 2024
Five open challenges for mechanistic interpretability, plus notes on linear representations, rare features, and measuring feature sensitivity.
Circuits Updates – June 2024
Comparing alternative sparse autoencoder architectures on the sparsity and reconstruction tradeoff, plus notes on external interpretability work.
The engineering challenges of scaling interpretability