OpenAI and Anthropic are throwing their support behind a plan to place independent safety evaluators inside leading artificial intelligence companies, a shift that could shape the first federal framework for regulating the most powerful AI systems. The proposal is gaining momentum as lawmakers, governors and technology executives respond to warnings that advanced AI could enable devastating cyberattacks, pandemics and other threats. But it also faces opposition from lawmakers and researchers who say private monitors may lack the authority and independence needed to challenge the companies they are supposed to oversee. The idea is central to the FRONTIER Act, a bipartisan bill sponsored by Rep. Jay Obernolte, a California Republican, and Rep. Lori Trahan, a Massachusetts Democrat. The legislation would require AI developers above specified revenue and computing thresholds to admit independent verification organizations into their operations and assess whether their models are being developed safely. OpenAI Chief Global Affairs Officer Chris Lehane said Tuesday that the company supports the House proposal. OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei have separately promised that their companies would give outside evaluators access similar to that provided to employees. “We see it as a starting gun,” said Andrew Freedman, co-founder and CEO of Fathom, a nonprofit advocating for independent AI assessments. He said the companies’ commitments showed growing recognition that outside monitoring would need to become a central part of AI governance. The proposal is emerging between two sharply different approaches. President Donald Trump has opposed broad AI guardrails, while Sen. Bernie Sanders of Vermont and Rep. Greg Casar of Texas have called for banning the most powerful forms of the technology. Independent assessments are being promoted as a compromise that could impose checks without halting the industry’s development. Several states are moving in the same direction. California has enacted legislation creating an accreditation and vetting process for AI auditing groups, while Massachusetts lawmakers are considering statewide requirements for third-party assessments. Utah Gov. Spencer Cox, a Republican, has urged Congress to require evaluations throughout the development of powerful AI models. Nvidia CEO Jensen Huang has compared the proposed reviews to financial controls. The most immediate challenge came from Rep. Josh Gottheimer, a New Jersey Democrat who co-chairs the House Democratic AI Commission. He said companies should instead submit models to mandatory government review before releasing them, arguing that industry enthusiasm for private monitors should prompt concern rather than confidence. Questions about independence center on the limited number of organizations currently equipped to evaluate frontier AI systems. Many are based in the San Francisco Bay Area and have professional, financial or ideological connections to researchers at major AI labs. Critics have focused particularly on links to effective altruism, a movement whose supporters have devoted significant attention to catastrophic AI risks. Benjamin Recht, a University of California, Berkeley, computer science professor, said the network of researchers and funders involved in AI safety is too concentrated. He warned that mandatory audits could lead to “regulatory capture,” allowing OpenAI, Anthropic and their competitors to influence the people charged with scrutinizing them. The concern is not only theoretical, critics say. Recht pointed to a recent security incident involving OpenAI models and internet access as an example of how an audit could be weakened if reviewers blamed a system’s behavior rather than human decisions made by the company. Lehane said outside assessors must be widely viewed as independent, capable and competent. He compared the effort to the development of professional standards for accounting firms, adding that the country would need to build a recognized pool of qualified evaluators. One organization under scrutiny is Model Evaluation and Threat Research, or METR, which Amodei has cited as an example of a potential independent assessor. Former White House AI czar and venture capitalist David Sacks has argued that METR is too closely connected to Anthropic’s investors and employees. METR President Chris Painter acknowledged that the group includes researchers with ties to Anthropic and OpenAI and that some have connections to effective altruism. He said those links reflect the small pool of people who began studying risks such as AI autonomy and loss of control before the issue attracted broader attention. Painter said METR discloses potential conflicts in its reports, has rejected funding from prominent AI leaders and does not accept money from Coefficient Giving, a major funder of effective-altruist causes. Coefficient Giving recently announced $200 million in grants intended to build independent AI safety organizations, underscoring both the scale of available funding and concerns about who is shaping the field. Supporters of the monitoring plan acknowledge that conflicts will be difficult to eliminate while the industry remains young. A staffer for Trahan said the influence of AI researchers on outside evaluators was an “unavoidable problem” given the limited talent pool. Freedman said the solution is to expand and professionalize the ecosystem rather than reject outside assessments altogether. “We can’t let the perfect be the enemy of the good,” he said. Recht disagreed, calling audits “toothless” unless governments also impose meaningful penalties. He urged lawmakers to consider fines or criminal liability after major cyberattacks and protections for open-source AI that would allow outside parties to inspect how systems work. The coming weeks could determine whether the companies’ public commitments become enforceable federal requirements. For communities whose schools, workplaces, public agencies and critical infrastructure are likely to rely increasingly on AI, the dispute is ultimately about who gets to decide when these systems are safe enough to deploy—and who bears responsibility when they are not.