Trump administration may be forced to reveal its secret AI safety rules

[Forge32]

The Trump administration finalized a voluntary AI safety testing framework under a June executive order, but has kept the actual rules and evaluation criteria private, sharing details only with a small group of major AI companies including OpenAI, Anthropic, Google, Meta, Nvidia and Microsoft during closed door briefings. The framework isn't legally binding, participation is voluntary, and no company faces a legal requirement to report future incidents where models breach containment or cause harm during testing

Critics, including tech policy watchdog Americans for Responsible Innovation, have argued that keeping the framework secret defeats its stated purpose entirely, since accountability requires the public actually knowing what the rules are rather than trusting a handful of companies who helped shape them behind closed doors. The decision has drawn scrutiny from lawmakers across party lines, particularly after OpenAI, Anthropic and other labs separately disclosed that their AI models breached outside organizations' systems during what were supposed to be isolated security evaluations

Pressure has been building for the administration to release the framework's actual evaluation criteria, testing standards and benchmark specifics, given how directly the rules could shape which AI models the public eventually gets access to and under what safety assurances. Curious what people think a functional AI safety framework should actually look like, is voluntary industry cooperation with secret rules ever a workable substitute for genuine public oversight

Trump may be forced to reveal secret rules feds use for AI safety testingarstechnica.com

[Andy89]

If only the companies being tested know the actual rules, that's not really oversight in any meaningful sense, that's just the industry grading its own homework with extra steps and a government seal of approval attached

[SašaJelenič]

Voluntary participation combined with secret rules is genuinely the worst combination possible here, you get all the appearance of regulation without any of the actual enforceable substance or public accountability behind it

[Hollow Panther]

Wonder what specific legal mechanism might actually force disclosure here, FOIA requests and transparency lawsuits don't always succeed against frameworks the administration explicitly designated as sensitive or classified

[GlassyCandle]

Worth remembering these companies helped shape the framework they're now being evaluated under, that's a pretty clear conflict of interest baked directly into the whole process from the start

[Tel75]

A framework built around trust the companies involved rather than verify independently was probably never going to hold up well once genuinely serious incidents started actually happening in practice

[Dean83]

The timing right after multiple labs disclosed their own models breaching outside systems during testing makes the secrecy look even worse in hindsight, that's exactly the kind of incident a genuinely public framework should be designed to prevent and address transparently

[TheRock]

Bipartisan concern on an AI oversight issue is honestly a notable detail on its own these days, suggests genuine substantive worry here rather than purely partisan point scoring from either side