The Government Says Its AI Testing Is Voluntary, but It Doesn't Look that Way

In Brief

AI policy writer Dean Ball, a former White House staffer who helped write the administration's AI Action Plan, argues that an executive order billed as voluntary frontier-model testing has quietly become an involuntary approval regime, and that nobody knows what a lab must do to get a model cleared for release. 

His fix: stop regulating individual AI models and start regulating the labs themselves, with government-certified independent auditors checking each lab against its own published safety framework.

What Happened

Earlier this month, the Trump administration signed an executive order billed as a voluntary testing program for frontier AI models, the most advanced and capable class of AI systems. AI policy writer Dean Ball argues it has become something very different. In a new essay called "What Should Be Done," Ball makes the case that the order has created a de facto involuntary licensing regime, meaning the government is effectively requiring its own approval before a model can be released even though the program is officially called voluntary.

Ball is worth listening to here. He is a former White House AI policy staffer who helped write the administration's AI Action Plan, and he is set to join OpenAI soon, though he stresses the essay reflects his own views, not the company's. He says recent events prove the point: the administration revoked public access to Anthropic's latest frontier model, and now OpenAI's GPT-5.6 is being limited to a small set of users at the government's request. His central worry is that nobody, including the administration, knows what a lab would actually have to do to get a model approved for broad release.

Ball is careful to say the administration is not directionally wrong and the catastrophic risks are real. But he warns the current approach is economically dangerous. Labs recoup the cost of training a frontier model in the first few months it is broadly available, so every week it's delayed threatens the business model. And the hundreds of billions of dollars going into data centers assume a roughly global market, not access for 100 government-approved companies.

SmarterX founder and CEO Paul Roetzer examines Ball's essay on Episode 222 of The Artificial Intelligence Show.

The Key Numbers

~500 - Top-tier U.S. AI researchers, by Roetzer's generous estimate

3-6 months - Time Ball gives the government before its approach breaks the AI industry

100 - Government-approved companies with access to latest AI frontier model

2-3 months - When a safety standard written today could already be obsolete

Is It Really Voluntary Testing?

Ideas not Complaints. Ball opens by returning to a warning he made when the order was signed. "I argued that the executive order on cyber and AI, which claimed to establish a voluntary testing program for frontier AI models, was really establishing a de facto involuntary licensing pre-approval regime," he writes. "This analysis has proven correct." Roetzer's read is that the piece is valuable precisely because it moves past complaint. "I just want people to bring ideas to the table, not just complain about things," he says.

A lack of experience and expertise. Ball notes that almost no one now setting AI policy in the government has frontier-lab experience, and that a leader hired to run the government's AI standards center was fired within days. Roetzer thinks this situation is a big problem. There are maybe 500 top-tier researchers in the country, he says, possibly far fewer, and none of them are leaving a lab to do this. "You're going to go work for the government instead of in one of the labs?" Roetzer says. "No one is going there. So what do you have to do? The government has to borrow the talent."

An ecomomics problem. Ball warns that "no one is building $100 billion data centers to serve frontier models to whatever 100 companies the U.S. government will allow access," and that the current approach "will break the AI industry within three to six months." Roetzer sees no middle path. "I actually see this is a binary thing," he says. "You either allow them to sell these models and make money, or you don't, and if you don't, the industry crumbles."

Concentrating the models concentrates the power. Ball argues that bad AI futures are likelier if only a narrow set of people can access frontier AI because that set is already composed of the most economically and politically powerful groups. Roetzer puts a sharper edge on it.

"What happens when you give powerful people more power? Not great things for democracies."

—Paul Roetzer, founder and CEO of SmarterX, Episode 222 of The Artificial Intelligence Show

Ball's prescription is to stop regulating individual models, which change too fast to pin down, and instead regulate the labs as entities: federalize state laws that require labs to publish safety frameworks, then have independent, privately run auditors check each lab against its own plans and its internal governance of recursive self-improvement loops, where AI is used to improve the next version of AI. The government would not conduct reviews; it would certify the auditors, the way accountants are licensed. Roetzer keeps circling the same practical gap. "Who are you going to get to do these audits?" he says. "Who is going to audit labs instead of working in them?"

SmarterX Take

The most useful thing about Ball's essay is that it reframes the unit of regulation. For two years, the assumption was that you govern AI by setting thresholds on models, originally based on the raw compute used to train them. More efficient algorithms blew that up, letting labs hit the same capability with far less compute. Any model-based threshold the government writes today is obsolete within a year, which is exactly why Ball wants to regulate the lab, not the model.

The harder problem is the one Roetzer keeps surfacing: the people technically capable of auditing a frontier lab are the same people the labs are paying enormous sums to build the models. A certified-auditor system only works if independent verification becomes a real career. Ball points to the bipartisan Great American AI Act from U.S. Reps. Jay Obernolte, a republican, and Lori Trahan, a democrat, as the closest thing yet to this vision. The fact that a workable proposal now exists in Congress is worth tracking.

What to Watch

A market panic is the real near-term risk. Ball notes that U.S. reindustrialization efforts, from nuclear energy to natural gas to batteries, are predicated on expected demand from the AI industry. If the government cannot produce a clear approval process and that demand looks unlawful, the shock could cascade well past the frontier labs and into the broader economy.

The jobs debate may follow the exact same arc. Roetzer thinks AI safety regulation just hit the inflection point everyone predicted and nobody prepared for, and that the economy is next in line. He expects a jobs report or a major AI-driven layoff to arrive the same way the model restrictions did, suddenly and after years of warnings went unheeded.

Only 13% of Organizations Can Govern AI at Scale

While Washington debates how to govern the labs, most companies cannot yet govern their own AI use. According to the 2026 State of AI for Business Report, only 13% of organizations have all four governance foundations in place (an AI roadmap, an AI council, generative AI policies, and an AI ethics policy), and 32% have none of them at all.

The report is built on more than 2,100 responses across roles, functions, and industries, and the governance gap maps directly onto the regulatory uncertainty Ball describes. If the rules of the road are about to be redrawn, the organizations with a roadmap, a council, and clear policies will adapt fastest, and most are not there yet. Read the full report →