Trump may be forced to reveal secret rules feds use for AI safety testing

Those protections are part of a CISA provision that lapsed once, and Congress must soon decide if the provision should be extended. There’s discord between the House of Representatives and the Senate on the deadline to vote on the extension, but the earliest that could happen is September 30 and the latest is December 11.

“In either event, Congress will soon need to decide whether to extend CISA’s liability protections that, by the administration’s own account, sustain the GOLD EAGLE program that the administration has announced but has yet to explain,” Protect Democracy’s complaint said. “Indeed, a senior administration official publicly acknowledged at GOLD EAGLE’s launch that, absent that reauthorization, ‘this effort is fundamentally challenged.’”

Deana El-Mallawany, director of Impact Programs & Counsel at Protect Democracy, told Ars that “neither Protect Democracy nor Congress has enough information about the GOLD EAGLE program to make a fully informed decision on extending CISA. Our concern is that the White House is putting Congress in the position of voting on a CISA extension without knowing how the GOLD EAGLE program is operating, who’s involved, or what statutory authority is being invoked. That’s why we’re seeking these records.”

Protect Democracy alleged that without these records, the Trump administration could escape accountability for AI safety review.

“Congress cannot meaningfully oversee programs whose terms and asserted authority it has not been shown” and “courts cannot review action they cannot identify,” the complaint said.

El-Mallawany told Ars that if Protect Democracy gets access to the framework, all documents will be posted publicly. Without more transparency, the list of concerns that Protect Democracy has about Trump’s potential ability to leverage the voluntary framework over the AI industry is long, El-Mallawany said, adding:

The risks are abundant, including the risk that the White House could use the secret process to coerce AI companies over decisions like whether to deploy AI for lethal autonomous weapons and mass surveillance (as we saw with the clash between the Pentagon and Anthropic), to extort certain investors and employees for political and financial support or punish others for perceived political opposition, or to force AI companies to incorporate politicized viewpoints into their models—not to mention the risk that the review framework is simply ineffective, creating a false sense of security.