[ggm]

I actually meant the reverse: Government and related information, highly strategic information is dispersed, but winds up in the model. Inferences from the information are not uploaded but the model is trained in ways which expose the inferences, and thus the secure-room-only information of strategic value leaks into the model, and becomes visible to people sharing access to the model. If the partitions between government and private models are weak, then I see the risk.

Quite a lot of government secrets are bound up in trade. If the data about trade and goals in private sector enterprises forms the backbone of department of state strategic intent, then to some extent, the model will reinforce "what is the actual intent of the state department in pursuit of these apparent goals over the same data"

"oh, I didn't SAY that we were going to totally onshore VLSI before 2030" is one thing. Being able to draw the inference over land deals, taxation changes, state level decisions, planning decisions, related strategic supply chain decisions, pending cases in review, matters not available for discussion as a collected set over time...

[palata]

Oh right, governments leaking to those companies by using AI. Sure, that's also a concern.