Key facts
- OpenAI supports a bipartisan House plan to mandate independent AI model safety audits.
- Anthropic and OpenAI have pledged to embed third-party evaluators within their companies.
- Critics warn that proposed third-party AI monitors may have conflicts of interest with AI developers.
- Rep. Josh Gottheimer prefers mandatory government vetting of AI models over private sector audits.
- The FRONTIER Act proposes a federal framework for AI safety, including independent verification organizations.
- Concerns exist that AI monitoring groups are too closely linked to AI companies and effective altruism ideology.
Top artificial intelligence companies, including OpenAI and Anthropic, are increasingly backing proposals for mandatory independent audits of advanced AI models to ensure safety. OpenAI, in particular, has thrown its support behind a bipartisan House plan that would require major AI developers to embed outside observers to verify the safety of their products. This move comes amid growing industry warnings about AI's potential to cause catastrophic harm.
However, the idea faces significant opposition and skepticism. Critics, including venture capitalists, academics, and some lawmakers, express concerns about potential conflicts of interest and 'regulatory capture,' where the oversight bodies might be too closely aligned with the AI companies they are meant to monitor. Rep. Josh Gottheimer (D-N.J.) has voiced strong opposition, advocating instead for mandatory government vetting of AI models before public release, arguing that industry self-regulation through private monitors is insufficient.
Academics like Benjamin Recht, a computer science professor at UC Berkeley, warn that these proposed safety schemes could primarily serve to entrench the dominance of leading AI firms like OpenAI and Anthropic. He suggests that auditors with ties to these companies might downplay human errors, such as the recent Hugging Face hack, and instead attribute security incidents to the technology itself.
Despite these concerns, the concept of third-party AI monitoring is gaining traction. State governments in California, Massachusetts, and Utah are considering similar proposals. OpenAI and Anthropic have voluntarily pledged to allow third-party evaluators access to their operations. Andrew Freedman, CEO of the advocacy group Fathom, views these pledges as a crucial step toward establishing governance for AI development.
The FRONTIER Act, proposed by Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.), aims to create a federal framework for AI safety, including requirements for independent verification organizations for companies meeting certain revenue and investment thresholds. Nvidia CEO Jensen Huang has also endorsed third-party AI safety evaluations, comparing the process to financial auditing.