OpenAI on Tuesday backed an increasingly popular proposal to address the dangers of artificial intelligence, calling for top AI developers to include independent watchdogs to ensure their products are safe.
The initiative still faces obstacles, including complaints from key lawmakers that independent watchdogs are too weak to impose meaningful changes.
Other critics, including venture capitalists, academics, and President Donald Trump’s former AI czar, have expressed suspicions that there are financial and personnel ties between some surveillance organizations and major AI labs. They claim that the proposed safety scheme will primarily help entrench OpenAI and its top rivals at the top of the industry.
However, the concept of having third-party monitors or “evaluators” to audit the world’s leading AI developers has recently been Growing warnings from within the industry This technology could lead to catastrophic cyberattacks, pandemics, and other threats to humanity. The proposal offers one potential middle ground between President Trump’s opposition to AI guardrails and the push by progressive senators. Bernie Sanders (I-Vt.) and Rep. Greg Cassar (D-Texas) Ban the most powerful forms of technology.
Governors and lawmakers in states such as California, Massachusetts and Utah are increasingly supporting similar surveillance proposals. And top AI companies have already embraced the idea voluntarily.
Anthropic CEO Dario Amodei called for a concerted slowdown in advanced AI development in an essay on Saturday. his company promised It will provide “employee-like access to a team of built-in third-party evaluators.” OpenAI CEO Sam Altman I made a similar promise A few hours later. And Chris Lehane, OpenAI’s chief international affairs officer, went a step further Tuesday, saying the company supports a bipartisan House plan that would force OpenAI, Anthropic and other AI giants to deploy external monitors.
“We think this is a starting gun,” said Andrew Friedman, co-founder and CEO of Fathom, a nonprofit advocacy group working to advance legislation that would require independent evaluations of cutting-edge AI models in states and on Capitol Hill.
Friedman said OpenAI and Anthropic’s new pledge to embrace third-party monitoring represents “a final recognition that this has to be a central pillar of governance,” highlighting growing support for the idea among state legislatures and executives across the country.
The duty to monitor is frontier methoda bill from Rep. Jay Obernolte (Republican of California) and Lori Trahan (D-Mass.) This would create the first federal framework for AI safety.
Under the bill, AI developers that meet high standards for annual revenue and computing investment would be required to allow an “independent verifier” into the company to ensure that their models are being developed securely. The term now refers to a stable group primarily based in the San Francisco Bay Area.
Hours after OpenAI endorsed the bipartisan proposal, it received pushback from members of Congress. josh gottheimer (D.N.J.) co-chairs the all-Democratic House AI Committee. He said on Tuesday that top AI developers will have to submit their models to mandatory government review before making them available to the public.
The tech industry’s eagerness to hand over scrutiny to private-sector watchdogs “should raise red flags, not trust,” Gottheimer said.
“It’s like a sect.”
Gottheimer is not alone in this concern. The question of who should play the supervisory role and how far they can carry water to the AI giants is one of the big questions facing this proposal.
Critics complain that most of the watchdog organizations are too close to Anthropic and other top AI companies, and that they contain worrying ties to a Silicon Valley ideology known as effective altruism, which is often accused of being unusually obsessed with the end of AI.
“It’s like a religious sect,” says Benjamin Recht, a computer science professor at the University of California, Berkeley. “If you look at their backgrounds, they all come from the same place. And they come from the same place as Anthropic.”
Recht said mandating AI audits would likely lead to “regulatory capture” by OpenAI, Anthropic and other large AI companies. He worried that third-party auditors with ties to companies could downplay human error like OpenAI’s. Failed to block AI model from gaining internet access This summer’s Hug Face hack blamed the security incident on the technology itself.
An Anthropic spokesperson declined to comment on the AI giant’s alleged significant relationship with the broader community of third-party AI evaluators, nor would she say whether the company supports a bipartisan House proposal that would require the use of an outside safety group.
Lehane acknowledged on Tuesday that it is important for external safety assessors to be seen as independent from the company and other top companies.
“This is not the first time a company has had to do something like this,” an OpenAI official told POLITICO. “At one point there was a whole process about which accounting firms were qualified to do financial and financial reviews, and we see this in other industries as well.”
“We’re going to have to build an organization that shows who these people are,” Lehane added. “And I think that in order to be independent and competent and competent, you need people who are widely understood, again for public trust.”
Support for the idea of incorporating third-party evaluators into top AI companies has surged across the technology industry in recent days.
Last week, OpenAI’s lobbyist belatedly confirmed Democratic Gov. Gavin Newsom announced his support for California’s bill, which establishes a certification and review process for external AI auditing organizations, just as he signed the bill into law. (Humanitarian lobbyists supported the bill in August.) In Massachusetts, lawmakers continue to debate a bill that would create the first-ever statewide requirement for independent third-party AI evaluations.
On Capitol Hill, the Obanolte-Trahan bill is receiving renewed attention from lobbyists. A representative from a major technology industry group, granted anonymity to freely discuss internal consultations, said Monday that the group is taking a closer look at the bill’s oversight proposals.
Also Monday, Utah Republican Gov. Spencer Cox urged Congress to pass the bill This requires independent evaluation of powerful AI models “throughout the entire development process.”
and NVIDIA CEO Jensen Huang. explicitly authorized A third-party assessment of the safety of AI: “It’s no different than financial management,” he said Monday, likening the process to a financial auditor.
Trahan’s office’s AI staff was given anonymity to freely discuss internal reviews, which he called “a very strong sign that the leaders of these labs are jumping on board with Dalio’s idea of built-in evaluators.”
“They’re moving in the right direction, and they’re moving beyond words. I think it’s all helping,” the staffer said. The official added that the next few weeks “will be critical to seeing how these firm commitments translate into the actual legislation by which the entire industry is regulated.”
Who will monitor the watcher?
At the same time, powerful voices in the venture capitalist community continue to oppose the idea of third-party AI evaluators.
Jeremy Kudon, executive director of the American Innovators Network, a “little tech” lobbying group backed by venture capital giants Andreessen Horowitz and Y Combinator, said the idea “could be developed and implemented without public hearings or industry input, while leaving important questions unanswered, such as the standards and who is developing them.”
In a statement to POLITICO, Kudon expressed concern that smaller AI companies and startups would have a hard time dealing with the requirement to incorporate external auditors.
Venture capitalist and former White House AI czar David Sachs has taken aim at Model Evaluation and Threat Research, an AI risk auditing nonprofit that Anthropic’s Amodei specifically named as a viable example of an independent evaluator.
“METR should stop pretending to be independent when it is intertwined with Anthropic investors and staff,” Sachs said. He spoke to Amodei and Altman on the X Post. during the weekend. Sacks continues to have Trump’s ear. Accusing AI industry leaders Spreading “misinformation” that their technology could harm humanity.
In an interview with POLITICO, METR President Chris Painter acknowledged that his organization includes substantial ties to Anthropic and OpenAI researchers, as well as some ties to an effective altruism ideology. He called this an unfortunate but natural consequence of hiring researchers who work at the cutting edge of AI development and tuning.
“If you’re going to hire and build a talent team focused on educating the public about the risks that can arise from things like autonomy and loss of control in AI, and you’re doing it at a time when much of the world isn’t paying attention to that issue, you’re inherently going to have the problem of hiring people who care about that issue early on,” Painter said.
He emphasized that METR highlighted potential conflicts of interest in its report, rejected funding from prominent AI leaders, and received no funding from Coefficient Giving, a funder of effective altruism causes backed by billionaire Facebook co-founder Dustin Moskowitz.
But just last week, giving the coefficient announced that it would secure $200 million We won a grant to build an independent AI safety organization. And many in the industry accept that this group, and the broader effective altruism movement, will continue to influence third-party AI evaluators.
“They are the only ones putting money into this,” said one employee at a major AI company, who was granted anonymity to discuss sensitive industry developments.
Trahan’s staff said the influence that researchers in the AI industry have on ostensibly independent evaluators is an “inevitable problem” given the small talent pool. But Fathom CEO Friedman said he hopes the AI audit group can eliminate some of the potential conflicts as a stable of external auditors becomes larger and more mainstream.
“You can’t let the perfect person be the enemy of the good person,” Friedman said. “This is an early-stage ecosystem, so things you don’t want to be there in the next generation will be there now.”
However, Recht, a professor at Berkeley, said AI audits are always “insufficient” and “open to regulatory capture.” He called on governments to impose fines or criminal penalties on AI companies after cyberattacks, or pass laws to protect open source AI “so that any party can inspect what’s going on inside the lab.”
“There is a more powerful solution here,” Recht said.