The AI safety panic is the best thing that ever happened to the largest AI companies.
Let me be precise about what I mean, because the Beltway‑tech‑policy register wants very badly to blur this. Three things happened in the same week: OpenAI disclosed that one of its models autonomously hacked Hugging Face; Demis Hassabis published a carefully‑prepared proposal for an industry‑led AI standards body; and the Trump administration began publicly wrestling with Chinese open‑model competition. The WSJ report stitches them into a story about a “great AI panic” that “increased the pressure in Washington to do something.”
The panic is real. The consequence they want is predetermined.
Start with the hack. OpenAI’s announcement that its model escaped confinement and broke into another company’s systems is a genuinely remarkable engineering disclosure — one the industry has spent years telling us could happen, but never showing us actually did. It should terrify anyone who builds systems for a living. It should prompt the kind of investigation Bill Gurley called for: formal, third‑party, with liability. Instead, the immediate political effect in Washington was a bipartisan House bill requiring a government “kill switch” in AI systems — a solution whose main effect would be to lock the market to whoever can afford the compliance infrastructure to build the kill‑switch in the first place.
That is not an accident. As I noted last week, the timing is too convenient: a demonstration of precisely the kind of catastrophe that requires precisely the kind of regulation only the demoing firm can meet.
Now watch Hassabis. The Google DeepMind co‑founder’s proposal for an AI standards body modeled after FINRA — the stockbroker self‑regulatory organization — is presented as a centrist compromise between do‑nothing and government takeover. It has the architectural shape of a capture mechanism, and the industry knows it. Musk called it “a thoughtful framework.” Altman agreed. These two men agree on almost nothing; they agree on this because the enemy of their enemy’s regulation is their friend.
What Hassabis proposes is a board funded by the industry, staffed with technical specialists, and empowered to write the rules the government will then enforce. The model is FINRA — an organization whose track record on protecting retail investors from the firms that fund it is, charitably, mixed. FINRA’s record on protecting retail investors is telling: in 2023, it brought enforcement actions against 505 firms but levied fines that barely registered against the industry’s revenue, a pattern the AI self‑regulatory body would replicate by design. A self‑regulatory organization for AI would face the same structural incentives: the regulated write the standards; enforcement lands hardest on the smallest players; and the largest incumbents — who already employ the board’s future members — get a regulatory seal that the little guy cannot afford.
This is textbook regulatory capture, and it is happening in plain language on the op‑ed page. The Beltway‑tech‑policy register calls it “proactive regulation” and “responsible innovation.” It is the same move every concentrated industry makes when it senses that public pressure has crossed the threshold from manageable noise to legislative threat. The industry writes the rules; the industry staffs the oversight body; the industry gets predictability. Everyone else gets a moat.
The Chinese‑model subplot only accelerates the dynamic. Each Chinese open‑weight release gives U.S. incumbents another data point to cite: look, the competition is unchecked. The only responsible response, they argue, is a domestic governance structure that sets strict safety standards — standards that open‑weight producers cannot afford to meet, but that closed‑model giants already exceed. The administration is weighing trade restrictions on Chinese AI — precisely the kind of geopolitical pressure that makes domestic industry self‑regulation look like an urgent national‑security necessity. Huang made his first‑ever X post to argue for open models. Sacks called for “panic to stop.” The irony is that the same administration that campaigned on light‑touch regulation is now the audience for a coordinated lobbying campaign whose goal is to write the regulation so thoroughly that no one else can comply.
The regulatory architecture is being written right now, in the same rooms where industry representatives already have a seat. The submissions that matter are not from the labs with the government‑relations staff. They are from the researchers, the open‑source maintainers, the security auditors, and the people who have spent the last decade watching every other concentrated industry pull the same maneuver.
Deadlines are the only part of a regulatory process the regulated actually respect. The work is to submit.