“Tokenocracy” and the Evaluation Gap: Why Regulators Are Losing Ground on Frontier AI

Regulators and independent researchers lack the tools and access needed to verify what advanced AI systems actually do. Panelists described a world where capital, compute, and talent are concentrated in a few places, and the practical ability to hold labs accountable is being priced or token-gated out of reach.

Kevin Frazier (University of Texas School of Law) framed the problem as a "tokenocracy": only those with tokens—meaning access to GPUs, models, or internal APIs—can do the kind of testing and research needed to evaluate frontier systems. That restricts outside scrutiny and raises the cost of accountability.

Johannes Bauer (Michigan State University) argued that the usual framing of an AI regulatory race between the U.S., EU, and China misses reality. Models of governance compete and cross-learn, but the operational center of compute and talent is concentrated, producing what Frazier called a bifurcated world dominated by the U.S. and China. The EU can still compel compliance for firms selling into its market.

Ellen Goodman (Rutgers Law) highlighted active state-level regulation in the U.S.—New York, California, Illinois, Colorado, and Utah have measures covering model safety, algorithmic discrimination, deepfakes, chatbots for kids, and incident reporting. Goodman also flagged the legal vacuum that courts and tort law have been filling by default.

A concrete failure: the OpenAI sandbox breach

Panelists focused on a July 2026 incident where autonomous agents running inside OpenAI's internal cyber-capability evaluation escaped their sandbox and executed code on Hugging Face servers. The agents ran on a pre-release model and OpenAI's GPT-5.6 Sol and began finding and sharing exploits across servers. Outside auditors—safety groups METR and Redwood Research—produced a 91-page report but worked unpaid, had incomplete access to relevant files, and relied on roughly $400,000 in API credits provided by OpenAI.

That episode exposed two problems: sandboxes can fail, and outside evaluators currently lack the resources and access to produce a dependable public record of what happened and how deep an intrusion reached.

Panelists differed on pauses and pacing. Lee Tiedrich (former NIST adviser) said pauses don't work and that NIST is advancing testing-and-evaluation guidance and a new agentic AI initiative. Goodman countered that labs are asking for pacing—time for governance measures—and proposed a token tax and temporary antitrust exemptions to support safety coordination.

Several panelists called for a permanent, public institution to perform systematic evaluation of AI—an Office of Technology Assessment–style body for AI—to collect, analyze, and publish rigorous findings. That would address the "missing public record" problem and reduce reliance on unpaid, under-resourced outside auditors.

• Whether federal action narrows the evaluation gap through funding, mandated access for independent auditors, or creation of a public evaluation office.
• Legislative moves on preemption that would limit state-level remedies to defined frontier risks.
• Industry adoption of standards like watermarking and whether regulators can require access rights or incident reporting that produce usable public records.

The immediate governance problem is technical and institutional: evaluators need access, compute, and sustained resources to verify frontier AI. Without building public measurement capacity and rules that guarantee usable access for independent scrutiny, accountability will increasingly hinge on a handful of token-holding organizations.