AI Evaluation Forum Releases Open Letter Setting Minimum Standards for Third-Party Risk Assessment

An AI evaluation forum made up of several watchdog groups has released an open letter setting minimum standards for third-party risk assessment. It says AI labs like Anthropic and OpenAI must meet specific requirements to work more effectively with outside evaluators. The letter was signed by dozens of people in the AI field, including computer science pioneers Geoffrey Hinton and Stuart Russell.

The forum says on-site third-party evaluators must stay scientifically objective, transparent and independent, and be fully protected from interference by the companies they assess. Evaluators should be able to communicate directly with company boards and other bodies with special oversight powers, without corporate filtering or limits, and have the right to publicly release their findings.

Evaluators should also be protected from retaliation by AI labs for their work, and get the same access to systems, data, tools and physical spaces as internal evaluators. The letter also argues that every AI lab should bring in evaluators with different views and professional backgrounds to cover risks such as biosecurity and AI loss of control.

Microsoft CEO Satya Nadella said the company welcomes the idea of letting third-party evaluators work directly inside AI companies.