Trump may be forced to reveal secret rules feds use for AI safety testing
Ars Technica Ashley Belanger
Four federal agencies were sued over secret AI safety reviews. A nonprofit says almost nothing is public about the rules, the companies, or even the legal basis.
Based on reporting by Ars Technica, Ashley Belanger — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Four federal agencies are now facing a lawsuit over a framework the Trump administration uses to review frontier AI models before they’re released. The fight is not about whether the reviews exist. It’s about how much of them the public is allowed to see.
In a Wednesday press release, Protect Democracy, a nonpartisan nonprofit, said that “almost no details” about the process have been made public to either the public or Congress. The group says the government has kept the basic shape of the review system under wraps.
That secrecy covers the key questions. According to Protect Democracy, it is still unclear what the review process actually looks like, which companies helped build it, and what legal authority Trump officials are relying on to run it.
The nonprofit also said the administration has shared details only with a small set of vague “trusted partners.” Even there, the group says, the identities of those entities and the criteria used to pick them have not been disclosed. So the government is asking for trust on a system it hasn’t really explained.
This is the sort of thing that makes AI policy look less like safety and more like a private club with a memo. If the rules are serious, publish them. If they’re not, don’t pretend the secrecy is a feature.
My take — AI-written commentary, not fact-checked reporting
This is classic closed-door AI governance: lots of seriousness, very little sunlight. A model watchdog doesn’t become safer just because a few “trusted partners” were invited into the room and the rest of the room was locked out.
Read more about this at: Ars Technica