More than 100 experts and risk assessment specialists have signed an open letter stating that current resources are insufficient for independent AI safety auditing. The group, organized by the AI Assessment Forum, is calling for developers of base models to provide greater transparency, scientific objectivity, and independence for third-party evaluators.
The initiative is supported by notable figures, including Geoffrey Hinton, as well as representatives from Johns Hopkins University, Stanford University, and the research organization METR. Conrad Stosz, president of the AI Assessment Forum, emphasized that the goal is to establish standardized principles for independent oversight to better manage risks associated with AI development.
The discussion gained momentum after Anthropic CEO Dario Amodei proposed granting independent experts access to internal safety procedures and incident reports. Stosz noted that this could involve providing auditors with access to company infrastructure, internal personnel, and confidential data on unreleased systems, while also ensuring legal protection for the auditors against potential retaliation.
Experts clarified that their work is intended to complement, not replace, internal corporate safety programs. Currently, there is a limited number of groups with the technical expertise and scale required to conduct such comprehensive evaluations, highlighting the need for a more structured framework for independent AI auditing.
