OpenAI's Pentagon Deal Isn't Surprising. That's the Problem.

But here's the thing that keeps nagging at me: safety research and military contracts aren't mutually exclusive. You can fund alignment research AND build systems for the Pentagon. In fact, that's probably the optimal strategy if you want to maintain public trust while cashing defense checks. The alignment work provides cover, the government work provides revenue. Everyone wins except, maybe, the people who believed the original mission statement.

The voluntary commitments OpenAI made alongside other labs are worth reading. They're full of the right words: safety, security, trustworthiness. But voluntary commitments are exactly that, voluntary. There's no enforcement mechanism. There's no penalty for breaking them. They're vibes, not regulations.

What we don't know

I want to be honest about the limits of what's actually been disclosed here. The specifics of the Department of War contract remain unclear. We don't know the dollar amount. We don't know which systems are being deployed. We don't know what "classified environments" actually means in practice (could be administrative AI for logistics, could be something much more concerning). OpenAI didn't share those details, and I only found the blog post and some secondary reporting to work with.

What we do know is that the company drew certain "red lines" in the agreement. But who defines those lines? Who enforces them? What happens when a military customer pushes back? These questions don't have public answers, and frankly, I'm not sure they have private answers either. It's too early to say whether OpenAI's safety commitments will hold up under the pressure of a major government contract.

The historical parallel nobody wants to hear

You know what this reminds me of? The early days of the internet, when every tech company swore they were building tools for human connection and democratized information. Then advertising happened. Then surveillance capitalism happened. Then we all became the product.

The AI industry is following the same arc, just compressed into about five years instead of twenty. The idealistic founding documents. The pivot to profitability. The government contracts. The gradual abandonment of the original vision in favor of whatever pays the bills. I've covered tech since the 90s, and the pattern is so consistent it's almost boring at this point.

What's different this time is the stakes. Social media made us anxious and polarized. AI deployed in military contexts could, well, let's just say the downside scenarios are considerably worse. And the companies building these systems are moving faster than any regulatory framework can keep up with.

So what

Look, I'm not naive enough to think OpenAI was ever going to remain a pure research nonprofit. That ship sailed the moment they took Microsoft's money. And I'm not even sure that government use of AI is inherently bad, there are legitimate applications in healthcare, infrastructure, even some defense contexts that most reasonable people would support.

But the speed and opacity of this transition matters. The gap between stated mission and actual business model matters. The fact that we're learning about classified military contracts through carefully worded blog posts instead of, say, meaningful public oversight, that matters too.

The kids running these companies (and yes, I know Sam Altman is 40, but compared to me everyone's a kid) seem to genuinely believe they're building something good. I don't doubt their sincerity. But I've watched a lot of sincere founders build things that ended up causing harm they never intended. Good intentions don't scale. Incentive structures do.

OpenAI's incentive structure now includes the Department of War. That's the fact we have to sit with. Everything else is, basically, commentary.

If you want to argue with me about this, my email's on the about page. But what do I know, I'm just a guy who's been watching tech companies make promises they can't keep for three decades now.

Mark Kowalski · 5 days ago · 3 min