Abstract

This disclosure describes how suppliers of AI computing, such as makers of AI accelerators, cloud providers and operators of computing sites, can make their supply conditional on the safety of the AI systems it serves, without having to judge their customers' businesses themselves. Computing is supplied under expiring leases. Each lease names the customer, the AI systems it serves by their lifelong identifiers, the amount and kind of computing, and an expiry, and is renewed only while those systems hold live certification in a neutral public registry, issued by accredited bodies under a published standard. The supplier checks the registry rather than forming its own view. When a system's certification lapses or is suspended, the lease follows a declared graded ladder: notice to the customer, a window in which to restore certification, then restriction of new or large workloads, then stop, with each step signed, shared with the customer and open to appeal to an independent body. Systems that provide essential services, such as health care or utilities, are declared as such and are restricted in stages that keep the essential service running. Accelerators may hold their own signed lease in a secure element, so that above a declared scale of work they run only while a valid lease is present; leases are renewed through the supplier or, offline, by a time-limited token. Accelerators reported stolen or diverted are marked in the registry, their leases are revoked, and they refuse work at scale, so that stolen hardware loses much of its value. Every lease, renewal, step on the ladder, appeal and outcome is recorded in a signed record. The result lets a supplier honour a commitment not to power unsafe AI with a neutral, fair and enforceable process, and narrows the computing available to those who refuse to be certified.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS