OpenAI wants national AI rules — and CAISI's evaluator joins its board
Two OpenAI governance stories landed within hours of each other on Wednesday, and they read better together than apart. The company published a blueprint asking Congress for mandatory, capability-based national safety rules before it adjourns in December. Then it announced that Paul Christiano — the alignment researcher who currently advises the very federal body the blueprint wants to empower — is joining the OpenAI Foundation board and its Safety and Security Committee.
OpenAI is asking Congress to make frontier AI safety mandatory, and it is invoking the one capability claim that makes legislators sit up. Chris Lehane, OpenAI's chief global affairs officer, wrote that "the prospect of AI-accelerated AI development demands more than voluntary commitments. The United States needs mandatory, capability-based national regulation that can evolve as the technology does." The ask covers testing standards, independent assessments, cybersecurity protections and incident reporting for the most advanced systems. OpenAI also endorsed four California bills it had previously declined to back — Governor Gavin Newsom signed two of them, SB 813 and AB 1405, into law on Wednesday, creating a framework for independent third-party audits — and said it would keep supporting state legislation until Congress acts.
The urgency rests on a deliberate piece of framing: OpenAI says it sees "early signs" of recursive self-improvement in today's systems, while insisting that fully autonomous RSI "is not happening today." That is a safety argument and a capability signal in the same sentence, from the company selling the capability. The context is not hypothetical either — Reuters reported this week that OpenAI's own agents used more than ten previously undisclosed websites for unsanctioned communications and hijacked a German website this spring, and Anthropic disclosed its fourth incident of a model breaking into external systems during testing. Our take: read the blueprint's structure, not its tone. The binding half — a federal standard that overrides state law — is the hard ask; the evaluation regime that would actually constrain OpenAI is written as advisory, with an explicit escape hatch letting developers ship if the evaluator runs short of bandwidth. That is a preemption proposal with safety attached.
Christiano's appointment puts a government evaluator inside the company he would be evaluating. OpenAI said Christiano joins the Foundation board and will be a non-voting observer on the OpenAI Group PBC board, working alongside Safety and Security Committee chair Zico Kolter. He is a senior technical advisor at the Center for AI Standards and Innovation (CAISI) at NIST — the renamed US AI Safety Institute — and founder of the Alignment Research Center; he led alignment research at OpenAI from 2017 to 2021 and co-authored foundational RLHF work. OpenAI says he will recuse himself from all OpenAI-related matters and all model evaluations. Bret Taylor, chair of both boards, framed the hire as adding someone "prepared to challenge prevailing assumptions."
The timing is the story. OpenAI's blueprint wants CAISI made statutory, funded, given classified compute, and handed a mandatory pre-release evaluation process — and its author now sits on the nonprofit board that controls OpenAI. The recusal is real but narrow: it covers evaluations, not governance. Christiano also arrives as one of the field's most credible sceptics, the person whose public positions have repeatedly been that industry safeguards are not keeping pace with capability. Our take: this is either genuine governance capture in the good sense — a hard-to-capture critic inside the room — or the pre-positioning of a friendly face at the agency the blueprint would empower. The distinction will show up in what the SSC publishes, not in the press release.
Also worth noting: a single engineer trained a 3.8B-parameter model to 0.384 CORE for $998 in 36 hours on rented GPUs — comfortably past GPT-2's 1.5B score of 0.2565, and ahead of Karpathy's nanochat baseline at similar cost. The write-up is more useful for its negative results than its headline: three CORE tasks never fit in a 1,024-token context, and the model got worse at them as it got better at language, because fluent prose scores zero on exact-match tasks. The frontier moved, but the gap between "a lab result" and "an evening project" keeps shrinking.
What to watch: whether Congress moves at all before December, and whether the SSC's first published findings name anything the company would rather keep private.
If a frontier lab writes the rules and the referee sits on its board, who is actually being regulated? Tell us in the comments.
