When Governments Decide Who Gets to Think

There’s a thread doing the rounds about the US government moving to individually approve access to frontier AI models. GPT 5.6, apparently, is now something you need permission to use. I’ve been sitting with this for a few days, turning it over, and I still don’t know whether to be more disturbed by the policy itself or by how unsurprised I am.

Let’s be honest about what’s actually happening here. When a government decides that a software tool is so powerful that ordinary people need to apply for access to it, that’s not safety policy. That’s industrial policy dressed up in safety language. The big players get their licences, their enterprise agreements, their backroom nods from regulators. Everyone else waits in line, or goes without. The moat gets built, and the people who funded the drawbridge get to decide who crosses.

Someone in the discussion put it pretty well: it’s technofeudalism. Peasants don’t get horses. Best you get is a donkey, and only for the master’s field. It’s a bit melodramatic as a framing, but the underlying point isn’t wrong. If access to transformative tools is rationed by government approval, the people doing the rationing will not be neutral about who gets approved.

The cynical read is that this is partly a hype play. Claim the model is so dangerous it needs restricting, generate enormous press coverage about how powerful it must be, then quietly lift the restrictions a few weeks later once the narrative has done its work. I can’t rule that out. AI labs have form on this. The hand-wringing about existential risk has always had a convenient side effect of making the product sound extraordinary.

But I think the more uncomfortable possibility is that the government means it. That they genuinely want to control who accesses these models, which means they’re one step away from controlling what behaviours the models have, what opinions they’re allowed to express, what questions they’ll answer and for whom. That’s not a slippery slope argument. That’s just describing the next logical step.

The open source angle matters here more than people outside the AI space tend to realise. Right now, there are genuinely capable models you can run locally, on hardware you own, without asking anyone’s permission. That situation is not guaranteed to continue. Payment processors get leaned on. Cloud providers get leaned on. App stores already have policies. The infrastructure around open models is more fragile than the “just download it” crowd tends to acknowledge. I watched the conversation about torrent repositories for model weights cycle through the same hopeful proposal and quiet collapse about four times in the last year alone.

From here, watching this from the outer southeast of Melbourne, the immediate practical impact is roughly nil. I can still pull models locally, still use the services I use, still do what I do. But I’ve been in IT long enough to know that the rules you ignore today are the ones that bite you three years from now, usually right when you’ve built something that depends on them not existing.

The geopolitical dimension is real and not particularly comfortable to think about. Chinese open-weight models are good and getting better. If US labs get locked behind government approval processes while Chinese alternatives remain freely available, the outcome isn’t less AI in the world. It’s a different distribution of who controls it. Replacing one set of concerns with a different set of concerns isn’t obviously an improvement.

I don’t have a tidy conclusion here. I don’t think there’s one. The case for some governance of genuinely dangerous AI capabilities is not insane on its face. The case that this particular intervention will do what it claims to do is much weaker. And the case that the people making these decisions are motivated primarily by public safety rather than by protecting incumbent interests is, to put it charitably, unproven.

What I know is that I’m watching the infrastructure of openness get quietly renegotiated, and the people doing the renegotiating are not primarily thinking about me, or you, or anyone who isn’t already at the table. That’s worth paying attention to, even if the immediate impact feels distant.