A model can outlast the person who built it. The proof of what it did must outlast the model. That changes how you think about keys, succession, and trust.
Most artificial intelligence systems are designed to be trusted in the present tense. But a model can run for decades, and the record of its actions has to be verifiable long after the keys, the company, and the author are gone. This is an argument for signing today for a verifier who has not been born yet.
Originally published on mickai.co.uk. This is a cross-post; the canonical version, with the full body, footnotes and references, lives on the mickai.co.uk article page.
