ARI applauds progress, urges federal government to turn voluntary action into enforceable safeguards
Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman today called for pacing the development of powerful AI models in the face of rising AI risks and proposed their companies provide third-party evaluators with employee-level access to verify their adherence to safety practices and commitments. The move follows calls by Americans for Responsible Innovation for embedded, continuous third-party monitoring of AI development as a core pillar of responsible frontier AI governance.
“Frontier AI development should move at the pace of our ability to keep it safe, not just what’s technically possible,” said ARI President Brad Carson. “The OpenAI–Hugging Face incident earlier this year proved that to keep AI safe, we need visibility into what’s happening within leading AI companies. As developers test advanced models internally, it’s clear that embedded third-party evaluators are central to responsible frontier AI governance. The commitments proposed today by Anthropic and OpenAI are incredible steps forward, but lawmakers must act quickly to enshrine these agreements into law and establish government-set safety standards that will make sustained visibility meaningful and provide an opportunity to responsibly pace the frontier.”
In August, ARI released Responsible Innovation at the Frontier, a federal blueprint for governing frontier AI that calls for embedded, accredited private verifiers to audit advanced AI technologies and their developers against safety and security standards.
Under ARI’s frontier AI blueprint, the federal government would hold covered developers’ safety frameworks to minimum federal standards that define adequacy, verify compliance with those standards, and maintain visibility into both internally deployed frontier models and the growing automation of AI research and development. Together, these pillars support a binding and adaptable response to the challenge of ensuring adequate frontier AI safety practices.
###