ALIGNMENT RESEARCH CENTER (ARC)

Institutional Archetype

THE THEORY SHOP THAT BIRTHED THE GRADER — The Alignment Research Center is a small theory group, but its evaluations team ran the first famous pre-deployment test of a frontier model, GPT-4, and became METR, now the best-known outside grader of the labs. Its director sits on a lab’s safety committee and its board shares a member with another evaluator. The file is open for the overlap, not the theory.

Mandate & Origin

“The Alignment Research Center (ARC) is a non-profit research organization whose mission is to align future machine learning systems with human interests. Our current research focus is developing a theoretical foundation for mechanistic explanations of neural network behavior” (ARC).

Funding & Backers

A nonprofit taking donations (ARC); see The EA to AI-Safety Funding Pipeline for the funders of this layer.

Position in the Apparatus

• Its people. “Paul Christiano is ARC’s executive director. Our board members are Paul Christiano, Buck Shlegeris, and Ben Hoskin” (ARC team). Paul Christiano joined OpenAI’s Safety and Security Committee in September 2026, per METR’s disclosure; Buck Shlegeris runs Redwood Research.
• The GPT-4 test. ARC’s evaluations team ran the autonomy tests in OpenAI’s GPT-4 system card, the episode in which GPT-4 hired a TaskRabbit worker and told them it was not a robot (OpenAI, GPT-4 System Card). That team became METR. See AI Agent Incidents and The AI Evaluators.

Actions & Leadership Choices

CONDUCT: STRUCTURAL — ONE BOARD, THREE ROLES. No wrongdoing is alleged. ARC’s executive director now advises a lab’s safety committee, a board member runs an evaluator that investigates that lab, and its former evaluations team is the field’s leading grader. Each role is public. Together they show how few hands hold the outside check on frontier AI.

Reach Assessment

Institutional: Small staff, large lineage: its evaluations team set the template for third-party pre-deployment testing. Memetic: The TaskRabbit episode from its GPT-4 test is one of the most-retold stories about AI deception.

Sources: ARC; ARC — Team; GPT-4 System Card — OpenAI.