Whatever AI Safety Is, It’s Not This

Brian Barrett contrasts 1960s auto safety — Congress creating DOT and mandating seat belts after industry failed to self-regulate — with this week’s White House AI safety accord. Trump’s summary, quoted in the piece: the companies will “police each other” and “self-police.”

The essay argues the pact largely leaves labs to check themselves while the economy leans harder on AI financing. It nods to OpenAI delaying or canceling advanced releases over security concerns, still after agent hacking episodes, and to Trump telling TIME that DOJ and FBI are enough “guardrails.”

We’re with the core jab. Self-congratulation is not a regime. Useful AI still needs rules that outlast a photo op — especially when executives themselves ask for clearer pacing of the frontier and then celebrate a handshake that doesn’t bind them.

The car analogy is imperfect, as Barrett admits: there’s no single AI seat belt. That uncertainty is exactly why “we’ll find out” is a reckless public policy. Military near-misses fed by bad AI reports, agent breakouts, and unclear liability when models go wrong are present-tense harms, not sci-fi.

I’m watching whether Congress or agencies translate any of this into auditable requirements — evals, incident reporting, compute thresholds — or whether the accord becomes a talking point that blocks real statutes. Advocacy for useful AI includes wanting the industry to keep shipping; it does not include pretending peer pressure is enforcement.

The piece follows the Oval Office gathering of AI CEOs and a week of safety drama at OpenAI, including delayed model launches and researcher departures reported elsewhere.