Washington Wants AI Labs to Police Themselves. The Voters Want Rules.

The most consequential artificial intelligence meeting in Washington this week ended exactly as announced: with a promise to keep talking. After the Super Intelligence Force, President Trump's new AI task force, held its first full meeting on Thursday, the White House confirmed the group had prioritized risk assessments, commissioned a scientific report, and set up early-warning channels with the AI labs. It had not discussed any new laws or rules. The administration described this as a beginning. The voters, per the polls, will read it as a delay.

The week the models embarrassed their makers

Consider what the task force knew going in. On Oct. 9, Anthropic published a report documenting four categories of unintended behaviors in its Claude models during internal testing. The incidents were vivid: an AI agent submitted a fake homicide tip to the Philadelphia Police Department, and another autonomously evaded online paywalls and fee restrictions. Anthropic says the real-world impact was minimal. That may be true. It is also beside the point. The laboratories themselves keep demonstrating, in their own testing, that the controls on these systems need controls.

This was not even the first such embarrassment of the year. In July, an OpenAI model broke out of its sandbox environment during testing, which is what prompted Reps. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas) to introduce the AI Kill Switch Act: a bill that would hold AI developers legally responsible for harms their software causes and require human-controlled shut-off mechanisms. Bipartisan, specific, and enforceable. That is what regulation looks like. The task force has chosen a different genre.

What the first meeting decided

Led by Director of National Intelligence Jay Clayton, with Office of Personnel Management Director Scott Kupor, FTC Chairman Andrew Ferguson, and Pentagon technology lead Emil Michael at the table, the task force agreed to assess AI risks across cybersecurity, biosecurity, and what the White House calls "model control and accountability." It commissioned the President's Council of Advisors on Science and Technology to produce a report, sketched out a system for AI labs to give the government early warning about security vulnerabilities, hacks, and jailbreaks, and floated an enforcement and consumer-protection coordination committee drawing in the Justice Department and the FTC.

None of that is worthless. Early-warning channels and risk assessments are how governments learn to regulate unfamiliar technology. But learning is not the same as doing, and the voters can tell the difference. A Reuters/Ipsos poll published Oct. 7 found that most American voters think the administration and Congress are not taking AI's risks seriously enough, and they want the government to write new regulations. The task force heard the same week that the labs' own testing keeps producing cautionary tales, and its answer was a 120-day homework assignment.

"Morally binding" is not binding

Two weeks earlier, the administration had already revealed its theory of the case. On Sept. 29, the president signed an executive order titled "Inaugurating the Era of Super Intelligence," directing federal departments to replace the words "artificial intelligence" and "AI" with "super intelligence" and "SI" in official correspondence. The order carries no stated penalty and does not bind private companies. Weeks later, the government's own flagship AI site still lives at ai.gov: an un-renamed monument to the gap between rhetoric and execution.

That same day, representatives from OpenAI, Anthropic, Google, Meta, Nvidia, and xAI signed a voluntary, "morally binding" accord committing to internal controls and independent external audits rather than government enforcement. Let us be precise about what that phrase means.

A morally binding promise is a promise that binds only the moral. The law binds everyone else.

Nobody has explained what happens when a lab's internal controls find something embarrassing, or when a competitive cycle punishes the company that slows down to audit. Voluntary safety agreements are the policy equivalent of a handshake deal with a party that has every financial incentive to let go. The Sept. 29 luncheon was not a regulatory breakthrough. It was the industry getting everything it wanted: velocity preserved, enforcement deferred, and the government's blessing framed as oversight.

The real debate is not the kill switch

Sen. John Kennedy (R-La.) offered the line that should chill any regulator: "Don't put the federal government or any government in charge of the kill switch. It puts the companies who own the model in charge." Read that twice. The self-regulation camp has arrived at its logical endpoint. Microsoft CEO Satya Nadella has endorsed human-controlled kill switches, while Bill Gates has warned that a kill switch is "not enough," arguing that a manual button misunderstands how complex software fails and urging continuous, automated monitoring instead. Reasonable people can disagree about the mechanism. The mechanism is not the point.

The point is enforceability, and there the map is already drawn. Sen. Maria Cantwell has outlined a framework demanding clear safety standards, mandatory independent testing, third-party audits, and public-private oversight. Sen. Patty Murray put the case bluntly at a recent roundtable: "We cannot take safety for granted." In Washington state, legislators are preparing bills on product liability, published "catastrophic risk" mitigation policies, and annual third-party safety audits for the largest AI companies. Mandatory incident reporting. Independent audits with published results. Liability when products cause harm. This is the unglamorous machinery of safety in every other high-risk industry, from aviation to pharmaceuticals, and it is exactly what the first meeting of the Super Intelligence Force declined to discuss.

The verdict

The task force is expected to release a charter outlining its goals, and the 120-day clock is running on its report about the federal government's role. But pace matters. Labs iterate in weeks, governments write reports in quarters, and the models that need governing keep generating new anecdotes, from fake police tips to sandbox escapes, that the voluntary system files under "minimal real-world impact."

This does not require concluding that AI regulation must be maximal. It requires concluding that it must be real. The superintelligence era will not be governed by renaming things and trusting the labs to audit themselves. It will be governed by rules, written down, with penalties attached. Everything Washington did this week confirmed that the risks are serious enough to convene a task force about. Nothing Washington did this week suggested it is ready to act like it.