Lawsuit Demands Disclosure of Secret AI Testing Rules in the US

The nonpartisan nonprofit Protect Democracy has filed suit against four US federal agencies, seeking publication of the framework of rules the Trump administration uses to conduct safety testing of cutting-edge AI models before their release. According to the plaintiffs, “almost no details” have been disclosed to either the public or Congress: it is unknown how the procedure is structured, which companies are involved in developing the criteria, and on what legal basis officials are conducting these reviews in the first place.

The organization calls this a dangerous lack of transparency. Decisions about which models are cleared for release may turn out to be the “most important policy question” of this White House. Meanwhile, as noted, OpenAI has already reached a private agreement with the government to limit distribution of its newest models to a circle of “vetted partners,” whose names and selection criteria have not been disclosed.

Suspicions of Corruption

The plaintiffs allow that the mechanism could be used in bad faith: officials could fast-track approval for firms favored by the administration and punish those out of favor. As an example, they cite a recent court ruling that found it unlawful to blacklist the company Anthropic for being “excessively progressive.” In Protect Democracy's view, the public cannot blindly trust the authorities in so sensitive an area, all the more so amid rising cyberthreats.

It is separately emphasized that the framework of rules itself, unlike the “classified benchmarking process,” bears no formal secrecy designation. Even so, the White House refuses to share details. An administration spokesperson told reporters that just because a document isn't classified doesn't mean it will be “broadcast to all and sundry.”

GOLD EAGLE and a Question for Congress

In July the White House launched the GOLD EAGLE platform, where industry partners help agencies identify cyber vulnerabilities, and on August 3 it announced the completion of a voluntary model-testing system. Here too, the plaintiffs note, neither the participants, nor the terms, nor the legal basis of the program have been named.

This matters because Congress will soon have to decide the fate of the 2015 Cybersecurity Information Sharing Act (CISA) — apparently the only legal footing for GOLD EAGLE. One senior official acknowledged at the program's launch that without its renewal, “this initiative is fundamentally vulnerable.” As the organization's representative, Diana El-Mallawany, put it, lawmakers are effectively being asked to vote without having been told how the program works or who takes part in it.

Protect Democracy sent identical FOIA requests to the four agencies, but not one has yet produced documents. The suit demands that all unclassified information be provided by September 30. As a counterpoint to the federal approach, the plaintiffs point to California's SB 813 bill, which, as state Senator Josh Becker emphasizes, was drafted entirely in public.

Source: Ars Technica