Industry players, policy experts, and businesses are increasingly trying to establish reliable ways to assess advanced AI systems. Inside the White House, the approach so far has leaned toward voluntary, industry-led solutions, with little expectation of rapid federal regulation in the near term.
Current situation
AI CEOs are wrestling with how to test advanced models and how to coordinate on safety standards, including whether to engage with Beijing. The White House favors a framework in which the industry finds ways to police itself, which has set off a competition to decide who will be designated as third-party evaluators of AI systems and controls while companies await legislation or further White House action beyond voluntary measures.
There is already a varied ecosystem of safety and benchmarking groups, but some White House officials and AI executives consider those organizations too closely linked to top AI companies. Other possibilities include businesses and startups that already conduct evaluations, and a range of less conventional proposals has also surfaced.
Who should evaluate the models?
A White House official said companies have the most expertise on AI and understand what the White House’s AI voluntary framework requires, adding, “These people are not 12-year-olds.” The official also said frontier companies need to reach a consensus because they have not agreed on much so far, and that if firms deem the situation dire they have the right, reason, and ability to throttle their models.
Conflicts of interest and scrutiny of third-party evaluators
Some third-party evaluators have drawn criticism for close ties to the effective altruism movement and to companies they might be policing — METR and individuals associated with it were singled out after METR investigated the OpenAI–Hugging Face incident.
One lead outside investigator in the Hugging Face matter is married to Paul Christiano, a noted AI safety and technology figure who recently joined the board of OpenAI’s non-profit foundation. An Anthropic employee reportedly left that startup to work at METR as well. METR has stated it does not take funding from frontier labs; a spokesperson said Christiano joined the OpenAI board after the investigation concluded.
AI companies and safety officials note the research community is small and tightly connected, and that employees from frontier labs are well placed to evaluate model capabilities — yet that close-knit nature raises questions about independence.
How current evaluation systems developed
Some aspects of existing AI evaluation grew from efforts to showcase model capabilities, such as performance on coding benchmarks and math tests. A wide array of firms, including defense and tech contractor Booz Allen, evaluate AI systems for clients.
Eric Syphard, Booz Allen’s head of AI, said many evaluations miss critical elements — for example, measuring how often models create software vulnerabilities or how variable their responses are to different prompts. “They all lean in to that performance-only view of the world,” Syphard told Axios.
Alternative proposals and initiatives
Elon Musk suggested this week that U.S. and Chinese labs could review each other’s models, a prospect that many regard as unlikely given intense competition among leading players. Other, less conventional suggestions surfaced as well, including a former Trump adviser’s call for programmer John Carmack to step in.
OpenAI published a public incident reporting playbook this week after identifying six new incidents. Like the White House framework, OpenAI’s incident reporting concept is voluntary; OpenAI also proposes third-party auditors for more complex cases.
Criticisms and what to watch next
Henry Papadatos, executive director of SaferAI, said public transparency is crucial: current company actions are positive but rest on goodwill alone, which he believes is insufficient.
Looking ahead, Treasury Secretary Scott Bessent — who will lead talks with Chinese Vice Premier He Lifeng this weekend — said there is an opening for safety discussions with Beijing. Such international talks could complement industry self-regulation, but competition between parties and the voluntary nature of current proposals mean a definitive solution is not yet in sight.
