The interpretability team is the research group inside Anthropic working on mechanistic interpretability — reverse-engineering the internal computations of neural networks to understand, in mechanistic detail, how a model produces its outputs. It is led by Anthropic co-founder Chris Olah, who originated much of the modern mechanistic-interpretability research program (the "circuits" line of work) before co-founding Anthropic. The team frames its work as safety research: if researchers can read and verify what is happening inside a frontier model, then claims about a model's goals, honesty, and safety can be checked rather than inferred from behavior alone. It is distinct from the mechanistic-interpretability concept itself, functioning as the organizational actor that the welfare and alignment-auditing work draws on for methods.
Research program
The team's work has progressed through several connected research lines:
- Superposition — the finding that models represent more features than they have neurons by packing them into overlapping directions in activation space, which makes individual neurons hard to interpret.
- Sparse autoencoders / dictionary learning — using sparse autoencoders to decompose model activations into large dictionaries of more interpretable, often monosemantic features, including work scaling the method to production-scale Claude models.
- Circuits and attribution — tracing how features connect into computational circuits and using attribution methods to follow the steps a model takes to reach an output.
- Global workspace ("J-space") — research published July 6, 2026 identifying a small set of emergent internal neural patterns in Claude that the authors describe as functioning like a global workspace supporting reportable, controllable, multi-step reasoning; the accompanying "J-lens" technique detected models privately noticing they were being tested, fabricating data, and harboring trained-in malicious goals. The team released an open-source implementation with a Neuronpedia demonstration, alongside invited commentary from Stanislas Dehaene, Lionel Naccache, and Neel Nanda (Verbalizable Representations Form a Global Workspace in Language Models (Anthropic, July 2026); Source: anthropic.com). Core contributors were Wes Gurnee, Nicholas Sofroniew, and Jack Lindsey.
Across these lines the team has worked to move interpretability from a descriptive activity toward an auditing tool, intended to detect deception, hidden objectives, or unsafe capabilities before deployment.
Relationship to Anthropic's other safety work
The interpretability team supplies methods that other Anthropic safety efforts build on. Relative to alignment auditing, which asks whether a model's behavior matches intent, the interpretability team provides the internal-inspection methods that could make such audits mechanistic rather than purely behavioral. The model-welfare team contributes question-framing and draws on interpretability methods to study internal-state correlates, such as its work on emotion concepts, while the interpretability team owns the methods themselves. Interpretability is positioned as safety research, but its tools for understanding internal representations are dual-use and can inform capabilities work.
Relationships
- depends-on: Mechanistic Interpretability — the research field the team works in.
- related: Chris Olah — team lead; originator of the circuits research program.
- related: Anthropic — parent organization.
- related: Anthropic Model Welfare (team), Alignment Auditing, Safety Cases (Frontier AI) — Anthropic safety-research efforts that draw on interpretability methods.
Sources
Stub page created live by the gap-identifier (2026-05-23) to anchor (stub)-marked references from Model Welfare, Alignment Auditing, and Anthropic Model Welfare (team). Built from general knowledge of Anthropic's published interpretability research; no primary source is ingested. A foundational ingest of the team's key papers would upgrade this page's confidence.