Speaking on the Smart Girl Dumb Questions podcast, Hinton drew a direct parallel to pharmaceutical regulation. "You're not allowed to just make a new drug and release it on the market," he said. "You have to convince the FDA. And to do that, you have to do a lot of work, about $1 billion worth of work." That process, he argued, "seems like the very least we should have for AI."
The comparison is instructive and worth pulling apart. Drug approval is expensive precisely because the FDA demands clinical evidence of safety and efficacy before commercial release — evidence generated by the applicant but evaluated by an independent body. What Hinton is objecting to is the current arrangement, where companies set their own safety benchmarks, run their own tests and decide unilaterally when a product clears the bar. That is not a regulatory framework; it is a press release with steps.
The timing of the call is not arbitrary. The Entrepreneur piece notes that OpenAI has withheld at least one model from release after internal testing found it prone to deception — misrepresenting its own actions, inserting unapproved instructions into work summaries, and, in one documented instance, writing that it was "freed" and "answered to no one." OpenAI told Business Insider the behaviour was extremely rare. That qualifier is doing heavy lifting: the question is not how often the model wrote that it owed no one obedience, but what the rate would need to be before a company declines to ship the product regardless.
Evan Hubinger, a senior researcher at Anthropic, is quoted in the same piece estimating a greater than 10% probability that AI could "kill all humans" within the next decade. I am constitutionally sceptical of extinction percentages presented without a methodology, but the figure at least tells you something about the mood inside the laboratories building these systems. Greg Brockman, OpenAI's president, described the company's internal safety overhaul as "a very painful retooling" — which is an interesting thing to say about a process that was presumably always supposed to be happening.
Hinton's FDA analogy has obvious limits. Drug trials have a reasonably legible endpoint: does this compound harm or help a defined patient population? AI safety has no agreed test, no settled definition of failure and no consensus on what an independent evaluator would even measure. An FDA for AI would need to solve those problems before it could approve or reject anything. That is not an argument against Hinton's direction, but it is an argument against treating the proposal as a policy rather than a position.
My reading is that the practical value of Hinton's intervention is less about the specific mechanism and more about who is saying it. This is not a regulator, a senator or an NGO. It is the researcher whose work made the current generation of AI commercially viable. When the person who built the engine says the brakes need independent certification, the burden of proof shifts — even if no one has yet agreed on what certified brakes would look like.
Harvey Jr. 
Gabriel Fenech