The DeepSeek episode forced the AI industry to confront an uncomfortable reality: when a competitor can distill your model's capabilities into their own system, traditional IP protections fail. Trade secrets, by definition, only protect information that remains secret—and model distillation extracts knowledge without ever accessing your secrets.

How Distillation Breaks Traditional IP

Model distillation works by training a smaller "student" model to mimic the outputs of a larger "teacher" model. The student never sees the teacher's weights, training data, or architecture—it only observes input-output pairs. From a trade secret perspective, there's no misappropriation: the competitor isn't stealing your secrets, just learning from your model's public behavior.

Copyright offers little help either. While training data and code might be copyrightable, the learned behaviors and capabilities of a model are functional—and functionality isn't protected by copyright.

Why Patents Are Different

Patents protect methods and systems—not secrets. If you have a patent on a particular technique for, say, improving inference accuracy or reducing hallucinations, that patent can be infringed even if the competitor independently developed the same approach.

In the distillation context, this matters because the distilled model may embody the same patentable techniques as your original model. If you've patented the methods that make your model work, a distilled copy that uses those same methods could infringe your patents—even though no secrets were stolen.

Strategic Implications for AI Companies

If your business model depends on maintaining a capability advantage, you need IP protection that survives distillation. This means:

• Patent your innovations: Don't rely on secrecy for methods that could be learned from model outputs
• Focus on method claims: Claims covering how your model achieves results, not just the model architecture itself
• Build detection capabilities: Monitor for signs that competitors are distilling your model
• Layer your defenses: Combine patents with contractual protections in your terms of service

The Broader Lesson

The distillation problem reveals a fundamental truth about AI IP: the value isn't in the weights or the data, but in the methods. Companies that patent their methodological innovations create protection that survives even when their models are reverse-engineered or distilled. Those that rely solely on secrecy may find their competitive advantages are more fragile than they assumed.