Paul Christiano

Paul Christiano is an alignment researcher listed by the Alignment Research Center among its researchers and as its executive director. This institutional description was checked in October 2026.ARC team record.

He coauthored Concrete Problems in AI Safety, which includes costly human supervision in its accident-risk agenda, and AI safety via debate, which proposes competing arguments before a human judge.^[1][2] These papers connect his work here to Scalable Oversight and AI Safety via Debate.

The 2015 beneficial-AI open letter lists Christiano among its signatories, supporting its call for robust and beneficial AI research.^[3]

A route through his research

Start with Deep reinforcement learning from human preferences, which Christiano coauthored. Its comparisons of short behavior clips offer a practical entrance to learning what people want without hand-writing every reward. Continue to the debate proposal when the difficult part becomes evaluating work beyond a single person’s expertise.^[4][2]

Hear the problem in his own words

In Robert Wiblin’s interview, published by 80,000 Hours on October 2, 2018, Christiano distinguishes useful outcomes from easily optimized feedback. Under “The problem of AI safety,” the published transcript warns about “get users to press the button saying that the AI did well.” This is an explanation of his concern, not a measured claim about a particular incident.^[5]

Under “Debate,” he asks whether sufficiently capable judges make honest arguments win, and whether humans meet that threshold. The unresolved question is part of the proposal. It helps explain his interest in Scalable Oversight without treating debate as a proven solution.^[5]

The quotation was checked against the publisher’s transcript; the audio was not audited. These are attributed 2018 views, rather than a claim about his present position on every issue.