When regulation arrives before innovation does

I've been thinking about something for a few days that feels pretty significant, and I wanted to write it down before I forget it.

Looking back, practically every technology that's changed how we live has followed the same pattern: innovation arrives first, people start using it freely, and only afterwards do governments step in with regulation, usually trying to rein in excesses or protect society from potential risks. It happened with the internet, it happened with crypto, it's happening with biotech. And yes, sometimes regulation has gone too far (in Europe I sometimes feel like there's more red tape than innovation per square metre), but the order has always been the same: use first, regulate later.

With AI we're seeing the exact opposite, and a few weeks ago gave us a perfect example.

On June 26th, OpenAI unveiled its new model family, GPT-5.6, split into three tiers: Sol (the most capable), Terra (balanced for everyday work), and Luna (fast and cheap). Nothing unusual there: something new comes out every few months. What's interesting is that, at the request of the US government, access was initially limited to a small group of around twenty vetted organisations, citing national security. No public waitlist, no open access. And it wasn't an isolated case: something similar happened to Anthropic with Mythos 5, its most advanced model, restricted since June 12th to around a hundred companies and government agencies.

It's the first time a government has asked a company to limit the rollout of its technology before it reaches the general public. Before, the sequence was: people get access, then the government decides what to do about it. Now the government decides before people can even try it.

And honestly, I don't know exactly what to think. On one hand, it's reassuring: maybe there's a solid reason behind it, maybe it's preventing a real cybersecurity risk or the kind of employment impact these models could have if released with zero oversight. On the other hand, it worries me, because unlike the internet, this technology isn't particularly open or distributed: the servers, the data centres, and the most powerful models are in the hands of a handful of companies. Before, anyone with an internet connection could build something. Now, if you don't have access to the most cutting-edge AI, how much can you really compete with someone who does?

And that's where the real question comes in for me: if access to the most advanced technology of our time is decided by a government and a handful of companies, is this still a democracy in the fullest sense? Or is that exactly the point (protecting us) and therefore it makes sense that it works this way? Honestly, I don't have the answer. But it feels like too big a shift not to stop and think about it.

That said, and even though it might sound contradictory, this week also brought good news. Anthropic launched Claude Science, a workspace built specifically for researchers: it connects to more than sixty scientific databases and has specialised agents for genomics, structural biology, chemistry, and more. It's not a new model (it runs on the same Claude models that already exist), but it's a very concrete way of putting AI to work for science: helping researchers move faster on projects that often get shelved due to lack of time, resources, or simply because they weren't profitable enough for pharma companies or labs. In fact, applications are open to fund research projects using this tool until July 15th, so if you know a researcher, it might be worth mentioning.

So I'll leave it there: a technology moving faster than ever, under more government control than ever, with real potential to help as much as to exclude, depending on who gets access to it.

What do you think? Does it seem reasonable for regulation to arrive before public access, or do you think this could end up concentrating too much power in too few hands?

Take care.

Update (July 17th): by the time this went up, both restrictions had already been lifted. OpenAI opened general access to GPT-5.6 on July 9th, and Anthropic did the same with Mythos 5 (publicly released as Claude Fable 5) on June 30th. The underlying reflection, that for the first time regulation arrived before public access, still holds. The timing I told it with, less so.