In a significant move signaling growing concerns over artificial intelligence safety, a coalition of AI oversight organizations has issued an open letter outlining minimum conditions for third-party risk assessments. The letter, supported by prominent figures in the AI community, including the renowned “Godfathers of AI,” Geoffrey Hinton and Stuart Russell, urges AI labs like Anthropic and OpenAI to adhere to these standards for effective external collaboration.
Setting the Bar for AI Safety Audits
The AI Assessment Organizations Forum, the group behind the open letter, emphasizes that for third-party evaluations to hold credibility, the personnel conducting these assessments must maintain scientific objectivity, transparency, and independence. Crucially, they must be protected from interference by the companies being evaluated.
This initiative comes in direct response to recent commitments from leading AI labs. Both Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have pledged to grant third-party evaluators access to internal company resources and AI risk assessments, offering them the same level of information access as company employees. The open letter aims to formalize these commitments and ensure they translate into meaningful, unhindered oversight.
Key Demands for Independent Evaluation
The Forum’s letter details several critical requirements for effective third-party AI risk assessment:
- Unfettered Access: Evaluators should be permitted to conduct their work openly and transparently, with direct communication channels to company boards and other oversight bodies, free from corporate filtering or restrictions.
- Freedom of Publication: The right to publicly release their assessment findings is paramount for accountability.
- Protection from Retaliation: Auditors must be shielded from any form of retribution by AI labs as a result of their work.
- Equitable Resources: They should receive the same access to systems, data, tools, and physical facilities as the AI companies’ internal evaluation teams.
- Diverse Perspectives: AI labs are encouraged to engage evaluators with varied viewpoints and professional backgrounds to comprehensively cover a spectrum of risks, including those related to biosciences and potential AI misuse or loss of control.
Industry Leaders Voice Support
The call for robust independent oversight has garnered attention from other major players in the tech industry. Earlier this week, Microsoft CEO Satya Nadella expressed the company’s openness to the concept of direct third-party embeds within AI enterprises for assessments. This sentiment aligns with the broader industry’s increasing acknowledgment of the need for stringent safety protocols as AI capabilities rapidly advance.
The ongoing dialogue and the establishment of clear guidelines by organizations like the AI Assessment Organizations Forum are vital steps in navigating the complex landscape of AI development and ensuring that powerful AI systems are developed and deployed responsibly.
