Protocolization of AI Model Security

AI development is quickly going from open source to privatized to secret. When stakes go up, protocols get called in.

Indeed, as perceived stakes escalate, different groups advocate for various protocols.

While the RAND report proposes a security protocol, the “Right to Warn” open letter calls for a whistleblower protocol.

I find the number of anonymous signatories to the open letter somewhat unsettling. Although I acknowledge that the danger of retaliation is real, anonymity seems to run counter to the transparency the letter is calling for. On the other hand, some degree of anonymity may be necessary, at least initially, to establish a space for dissent and critique.

A common thread connecting both protocols is the need for open discussion - an open learning protocol.

Personally, I suspect that the right approach to AI governance will require a combination of both weak and strong protocols.