OpenAI's Pentagon Deal Isn't Surprising. The Question Is What Comes Next.

This is the same problem we had with drone programs a decade ago. The government said "trust us, we're being careful," and years later we learned the targeting algorithms were considerably messier than advertised. I'm not saying OpenAI is lying about their red lines. I'm saying that promises made in press releases don't mean much when there's no mechanism for accountability.

The timing here matters too. OpenAI has been on a tear lately with government partnerships. Their "OpenAI for Government" initiative launched recently, explicitly targeting public sector adoption. They've been working with US CAISI and UK AISI on AI safety frameworks. They just committed $7.5 million to something called The Alignment Project for independent safety research. It's a full-court press to position themselves as the responsible AI company, the one governments can trust.

And look, maybe they are! Maybe Sam Altman genuinely believes he's building technology that will benefit humanity and that working with the Pentagon is part of that mission. I don't know the man's heart. But I've covered enough tech companies to know that what executives believe about themselves and what their organizations actually do often diverge over time, especially once the contracts get big enough.

Some key things worth tracking:

• The agreement reportedly includes provisions for OpenAI to decline specific use cases that violate their policies, but the enforcement mechanism is unclear
• Classified deployments mean no external auditing of how the technology is actually used
• OpenAI's stated "safety red lines" haven't been published in detail (at least not that I could find)
• The company's nonprofit structure has been quietly evolving in ways that give the for-profit arm more control
• Several key safety researchers have left the company in the past year

None of this proves wrongdoing. It does suggest we should be asking harder questions than "is OpenAI selling out" or "is this good for national security." The real question is whether any company, no matter how well-intentioned, can maintain ethical guardrails once it becomes a defense contractor operating in classified spaces.

I keep thinking about the young founders I've met over the years who started companies with grand visions about changing the world, making information free, connecting humanity, whatever. Most of them meant it! And then the incentives shifted, the investors got impatient, the government came calling with large checks, and suddenly the mission got, you know, flexible.

OpenAI's situation is complicated by the fact that they're not just any AI company. They're building what they themselves describe as potentially the most transformative (and dangerous) technology in human history. Their own technical goals document talks about ensuring AI benefits are "widely and evenly distributed." Does defense contracting fit that vision? Maybe it does, if you believe American military dominance is good for global stability. Maybe it doesn't, if you think concentrating advanced AI in weapons systems is exactly the kind of uneven distribution they claimed to want to avoid.

The honest answer is that nobody knows yet. We don't know what OpenAI is actually building for the Pentagon. We don't know what the red lines really are or whether they'll hold under pressure. We don't know whether the safety research they're funding will have any influence on their commercial and government work, or whether it's just reputation laundering.

What I do know is that we've been here before. The tech industry has a reliable playbook: make idealistic promises, get big, take the money, rationalize the compromises. Sometimes the compromises are fine! IBM's government work helped win World War II. Sometimes they're not fine, and we don't find out until much later.

The kids building AI startups today should probably study this history. Not because OpenAI is necessarily doing anything wrong, but because the pressures that shape these decisions are structural, not personal. Good intentions don't survive contact with billion-dollar contracts and classified programs, not without extraordinary effort and, frankly, some luck.

If you want to argue about this, my email's on the about page. But I suspect most of you already know I'm right, even if you're hoping this time will be different.

Mark Kowalski · 5 days ago · 3 min