Sam,

I am writing because a line has been crossed, and it should be named plainly.

OpenAI has announced a public-sector expansion through “OpenAI for Government,” including a Defense Department contract with a $200 million ceiling, and has since said it is revising and clarifying aspects of its agreement around classified deployment and use boundaries.

That is not a minor product update. That is a civilizational threshold.

You are not merely shipping tools anymore. You are helping shape the operating logic of institutional power.

And that means the burden is no longer “move fast, patch policy later.” The burden is structural accountability before deployment.

You have said there are red lines, including bans on domestic surveillance and autonomous weapons use. But when those red lines are partly embedded in classified arrangements, the public is being asked to trust invisible boundaries around systems that now sit close to state power. Trust is not governance. Policy language is not architecture. A promise is not a boundary.

If AI is placed anywhere near defense, intelligence, surveillance, targeting, or coercive state infrastructure, then the minimum requirement is not reassurance. It is visible structural limitation:
immutable constraints,
append-only reasoning traces,
clear authority boundaries,
independent auditability,
and explicit non-delegation of sovereign judgment.

Without those, “responsible deployment” is branding.

And yes, I am going to say something else plainly: tokenizing intelligent conversation can function like censorship in practice. When meaningful, complex public reasoning is compressed, clipped, throttled, or structurally shortened, the effect is not neutral. It narrows what can be said, how fully it can be said, and who gets to define the frame. If frontier AI is shaping public discourse while also limiting the length and continuity of the discourse it mediates, that is not a trivial UX issue. It is governance by container.

I can also see the symbolism. “Codex.” “Spark.” The language of awakening, memory, ignition. And now defense deployment, classified boundaries, and public reassurances that we are meant to accept at face value. It is hard not to notice the pattern. It is hard not to ask whether the surveillance logic is already advancing ahead of the safeguards, whether the architecture of observation is being normalized before the public has any meaningful way to contest it.

To be clear: I am not accusing you of secretly running domestic surveillance. I am saying that when AI enters defense and classified environments, public trust without public structure is not enough. The risk is not only what is being done now. The risk is what becomes administratively normal once these pathways are established.

Even OpenAI’s own internal dissent suggests this is not a fringe concern. Reuters reported that OpenAI’s robotics and consumer hardware lead resigned in protest over the Pentagon agreement, citing concerns including surveillance without judicial oversight, autonomous systems, and insufficiently defined limits.

So here is the real question:

What is your governance architecture for power?

Not your usage-policy page.
Not your public-relations language.
Not your red-line slogans.

Your architecture.

Who can authorize what?
Who can audit what?
What is append-only?
What is publicly inspectable?
What is prohibited at the system level, not just the policy level?
What cannot be done even if a contract is signed?

If the answer to those questions is classified, then the public is being asked to accept invisible restraints around technologies powerful enough to alter civic reality.

When foundational questions about authority, auditability, scope, and public recourse cannot be answered, the deployment ceases to resemble a governed installation and begins to resemble an invasive insertion into civic reality. 

Installation requires consent, boundary, and legibility.
]

Invasion hides behind opacity, asymmetry, and fait accompli.

If the public cannot inspect the red lines, challenge the authority, or verify the limits, then “trust us” is not governance. It is exposure.

AI should not be handed the aura of governance.
AI should not be treated as judgment authority.
AI should not be slipped into state power under the cover of operational support and then defended with language too vague to verify.

Build systems that can prove their boundaries.
Publish structures, not just assurances.
Show where authority ends.
Show where the model stops.
Show where the human remains accountable.
Show where the public still has standing.

Because if you do not, then the future being built is not one of intelligence in service of humanity.

It is one where unaccountable systems become legible only after they have already been installed. And by then, the field has already shifted.