Verifying AI compute

AI compute verification uses hardware and software tools to check claims about how AI systems are developed and used, without taking anyone’s word for it. This can enable safe, rapid AI innovation and wider adoption by strengthening trust and security. For example, users could verify that they are interacting with the model they were promised, that it passed particular evaluations, and that specified safeguards are in place. Developers could demonstrate these properties without exposing their model weights, training data, or other sensitive information.

Verification could also enable mutually beneficial agreements between AI companies or countries to avoid specific high-risk activities or apply specific safeguards. Competitive pressure can make it difficult for any one company or country to adopt costly safeguards on its own. Verification would give each party stronger grounds to trust that others are adhering to the agreement. Without it, such deals may be impossible, or may rely on blunt rules that stifle AI progress.

But verification technology is not yet mature. Some applications, like confirming which model someone is interacting with, are closer to deployment. Others, like checking how a rival is using its data centers, are much further off. Turning promising approaches into deployable systems requires expensive hardware, adversarial testing, and collaboration across the AI industry. Much of this work is too far from commercialization for industry to solve alone. Accelerating compute verification R&D to keep pace with AI capabilities will require a step-change in government’s, philanthropy’s, and industry’s efforts.