Gladstone AI's government-commissioned report warns of catastrophic AI risk
The four-person consultancy, paid $250,000 by the State Department, urged Congress to ban training runs above a compute threshold and restrict publishing open-weight models.
- Safety & alignment
- Government & policy
- Notable
Gladstone AI, a four-person AI consultancy, published “Defense in Depth: An Action Plan to Increase the Safety and Security of Advanced AI,” a report it had delivered to the US State Department the previous month under a $250,000 federal contract first commissioned in October 2022. Time magazine, which had obtained the document, reported its findings on the day of publication.
The report’s central recommendation was that Congress make it illegal to train AI models above a specified level of computing power, set — the authors suggested, without naming a figure — just above the compute used to train systems such as GPT-4 and Gemini. It proposed a new federal agency with authority to require government permission before frontier labs trained or deployed models above a lower threshold, and urged officials to “urgently” consider outlawing the publication of open-weight model parameters, with criminal penalties for violations. It also called for tighter export controls on AI chips, including monitoring technology embedded in the hardware itself, and increased federal funding for alignment research.
The report drew on more than 200 interviews conducted over roughly 18 months with researchers and executives at frontier labs, cybersecurity specialists and national security officials in several countries, and it framed its concern in explicitly military terms, describing advanced AI as creating “weapons of mass destruction-like” risks.
Reaction from policy specialists was sceptical of the proposals’ political viability. Greg Allen of the Center for Strategic and International Studies called an outright ban on training runs above a threshold “extremely unlikely to be adopted,” noting that existing US policy relied on monitoring and export controls rather than prohibition. The report’s authors acknowledged their recommendations, particularly the restriction on open-weight publication, would “upend” the AI industry and be unpopular within it. The State Department declined to comment on the findings. None of the report’s central proposals — the compute-threshold ban, the new licensing agency, or criminal restrictions on open weights — was enacted in the following two years, though it added to a body of national-security-framed argument for regulation that continued alongside industry-led self-governance efforts.