Let’s be honest. When you hear about AI companies accusing each other of ‘stealing’ models through distillation, you probably think it’s a simple matter of right and wrong. It’s not. It’s a power play, and the rules are rigged.
You’ve seen the headlines: ‘OpenAI outraged by unauthorized distillation.’ ‘Google demands respect for its API terms.’ Fair enough, right? They built it, they should control it. Except the same companies screaming about IP theft are the ones doing the stealing.
Take the recent incident where Moonshot was accused of distilling Fable for their K3 model. The response? Crickets from the accuser’s own history of doing the same. As one commenter put it: ‘rules for thee but not for me.’ That’s the entire industry in a nutshell.
Distillation—taking a powerful model’s outputs to train a cheaper, faster version—isn’t inherently unethical. It’s a technique. The AI industry’s rulebook is written by the incumbents, for the incumbents. When you have the capital to train a 1-trillion-parameter model, you can afford to be generous with your ‘openness’—until a smaller player uses your own outputs to compete. Then suddenly the lawyers show up.
This isn’t about ethics. It’s about control. The biggest labs—OpenAI, Google, Anthropic, Meta—all distill. They have to. Training from scratch is too expensive and too slow. The difference? They have the compute and the legal teams to get away with it. Distillation isn’t a crime; it’s a weapon. And the weapon is only illegal when the wrong person wields it.
Here’s the twist: the real scandal isn’t that distillation happens. It’s that the outrage is selective. When a startup like Moonshot does it, it’s ‘theft.’ When an incumbent does it, it’s ‘research acceleration’ or ‘model improvement.’ The same behavior, different labels. Why? Because the incumbents write the narrative.
And what does that mean for you? If you’re building on top of frontier models, you’re sitting on borrowed time. The terms of service can change overnight. The API can be revoked. The ‘open’ model can retreat behind a paywall. Your entire business model can be labeled ‘distillation’ and sued into oblivion, while the giants do exactly what you’re doing without a peep from the press.
So stop believing the narrative. The AI industry isn’t a meritocracy—it’s a battlefield where the biggest guns write the rules. The next time you see a headline about ‘unethical distillation,’ ask yourself: who is accusing whom, and who stands to benefit? The answer is always the same: those in power, protecting their power.
FAQ
Q: Isn't distillation a legitimate technique? Why is it considered theft?
A: Distillation is a legitimate technique in machine learning. The controversy arises not from the technique itself, but from selective enforcement: big companies use it freely while suing smaller competitors for the same practice. The real issue is the power imbalance, not the technology.
Q: What does this mean for startups building AI?
A: For startups, the message is clear: you are building on a legally shaky foundation. The incumbents can change their terms of service or sue you for 'distillation' at any moment. Build defensively, rely on open-source where possible, and expect the legal landscape to be a weapon used against you, not a level playing field.
Q: Isn't it fair for companies to protect their investments in model development?
A: In theory, yes. But the fairness evaporates when the same companies that cry 'theft' are themselves the biggest distillers. If you can't practice what you preach, you're not protecting investment—you're maintaining a monopoly. The hypocrisy undermines the moral argument entirely.