Demis Hassabis, on who funds AI safety: “Funding would need to be substantial and likely mostly come from industry”

Demis Hassabis has spent the year arguing that the most powerful AI systems should be vetted before they reach the public, by an independent body modeled on the watchdog that polices Wall Street’s brokers. The harder question is not whether such a referee should exist. It is who pays for it — and in a single sentence, the chief of Google DeepMind answered that too.

“Funding would need to be substantial and likely mostly come from industry, in order to attract world-class technical talent and provide the necessary compute resources for large-scale testing,”

Hassabis wrote that line in the framework essay laying out his proposal, quoted by Malay Mail. Read quickly, it sounds like responsibility: guardrails cost money, and the companies building the technology should foot the bill rather than leave the public to fund the cleanup. Read as a mechanism, it is the load-bearing line of the whole plan — the one that decides whose interests the referee ends up protecting.

The tension sits inside the sentence. Hassabis wants the body to be independent. He also wants it funded, “mostly,” by the very industry it would judge. Those two goals pull in opposite directions, and the history of the model he chose shows which one tends to win. The Financial Industry Regulatory Authority, his template, is paid for by the brokerages it oversees. Senator Elizabeth Warren has argued for years that FINRA acts “more in the interest of the brokerage firms that pay its bills than the individual investors it is meant to help protect.”

The objection is not fringe. Nader Henein, a vice-president analyst at Gartner, calls self-regulation simply “not viable,” because most vendors lack the capacity for it and the arrangement “necessarily” breeds conflicts of interest. Independent analyst Carmi Levy went further, describing the plan as “a self-serving roadmap for an industry bent on racing” ahead regardless of the harm. Even sympathetic voices want a firewall: Yoshua Bengio says he understands the pragmatism but insists on “a clear and precise roadmap to transition from a voluntary to a mandatory model.”

None of this makes Hassabis a bad-faith actor. He runs the lab behind AlphaFold, the system that predicted the shapes of proteins and reset how biology gets done — a record that gives his warnings about frontier capabilities real weight. That is exactly what makes the sentence worth dwelling on. The most credible technical voice in the room is also the one proposing the referee, drafting its rules, and arranging its budget from the same companies that would sit the exam. Voluntary submission, a short review window before release, industry money: each concession is reasonable on its own, and together they describe a body the labs could comfortably live with.

A safety regime is only as independent as its budget. The sentence that will decide whether this one protects the public or the labs is not the one about testing frontier models. It is the quiet one about who signs the checks.

Tags: AI safety, AI regulation, Google DeepMind, Demis Hassabis, FINRA