When the Government Decides Which AI Gets Released

Something fundamental changed in AI deployment this week, and most people missed it.

OpenAI announced it will release GPT-5.6 only to government-approved customers before any public rollout. Not as a courtesy. Not as a voluntary safety measure. As the new normal. According to an internal memo from CEO Sam Altman, this follows President Trump's executive order requesting federal review of powerful AI models—though OpenAI stressed the order is technically voluntary.

Meanwhile, Anthropic received explicit permission from the Commerce Secretary to redeploy its Mythos 5 cybersecurity model to over 100 organizations after suspending access in June due to security concerns. The company had to implement additional safeguards and wait for government authorization before flipping the switch back on.

This isn't about whether government oversight of AI is good or bad. It's about recognizing that we've crossed a threshold. The era of "build it, release it, deal with consequences later" is ending for frontier AI models. What's replacing it is messier, slower, and arguably more complex than anyone anticipated.

The voluntary nature of these reviews creates an interesting dynamic. Companies aren't legally required to seek approval, but the political and reputational cost of proceeding without blessing from federal agencies may be too high to bear. It's regulation by implication—powerful because it's ambiguous, concerning because it lacks clear boundaries.

Consider the downstream effects. If OpenAI needs 30 days of government review before releasing a model, what happens to their competitive position against companies operating from other jurisdictions? Does this create a two-tier system where approved models serve US markets while unapproved alternatives proliferate elsewhere? California's new AI job loss tracker suggests states are already preparing for economic disruption, but are they prepared for the regulatory fragmentation that might follow?

The cybersecurity angle adds another layer. Anthropic's Mythos suspension came from the company's own security assessment, not government mandate. Yet government approval was required to turn it back on. This suggests an emerging protocol: if you build something powerful enough to matter, you'll need permission to deploy it—even if you're the one who identified the risk in the first place.

We're seeing the birth of what might be called "approval infrastructure" for AI. Not quite regulation, not quite voluntary compliance, but something in between that will shape how innovation happens for years to come. The technology sector has operated for decades on the principle that permission is easier to seek after the fact than before. That principle is being rewritten.

The question isn't whether this is the right approach. The question is whether this ad-hoc system of executive orders, voluntary reviews, and case-by-case approvals can scale to handle the dozens of frontier models that will emerge in the next few years. Because if every major AI release requires Commerce Secretary sign-off, we're going to need a much bigger Commerce Department.

Or maybe that's exactly the point.