The AI Copy Wars: When Innovation Becomes Espionage

24,000 fake accounts. 16 million conversations. This is the scale of what Anthropic calls a coordinated "distillation attack" by Chinese AI firms trying to copy their Claude model. The gloves are officially off in the AI wars.

The Art of AI Theft

Anthropic dropped a bombshell Monday, accusing three Chinese companies—DeepSeek, Moonshot AI, and MiniMax—of running sophisticated campaigns to extract knowledge from Claude. Despite service restrictions blocking commercial access to Claude in China, these firms allegedly used commercial proxy services to circumvent the barriers.

The technique, called "distillation," allows smaller AI models to mimic the performance of larger, more expensive ones by learning from their responses. Think of it as a student copying the smartest kid in class—but at industrial scale.

MiniMax led the charge with over 13 million exchanges, flooding Claude with carefully crafted prompts designed to extract specific capabilities. The Chinese firms then used these responses to train their own models through reinforcement learning—essentially getting years of R&D for the price of API calls.

National Security or Corporate Protection?

This isn't Anthropic's first rodeo. OpenAI fired the opening salvo earlier this month, writing to U.S. legislators about "adversarial distillation" by Chinese firms. Both companies frame this as a national security threat, warning about "authoritarian governments deploying frontier AI for offensive cyber operations, disinformation campaigns, and mass surveillance."

But here's where it gets murky. Anthropic has long advocated for tighter export controls on AI chips to China, consistently framing "compute leadership as a national security priority." As Rui Ma from Tech Buzz China notes, "Whether intentional or not, the narrative of illicit capability transfer strengthens the case for stricter chip restrictions."

The timing is telling. On the same day as Anthropic's statement, Reuters reported that the U.S. found evidence of DeepSeek training its AI model on Nvidia's flagship Blackwell chip, apparently flouting export controls.

The Hypocrisy Question

Here's where things get uncomfortable for the accusers. Online critics quickly pointed out that Anthropic itself uses distillation to train proprietary models. The company even acknowledged that AI firms "routinely distill their own models to create smaller, cheaper versions."

So what's the real difference? When American companies do it, it's called optimization. When Chinese companies do it, it's called theft.

The Chinese firms haven't responded to requests for comment, but their perspective is clear: they're using publicly available APIs and achieving similar performance at "a fraction of the time, and at a fraction of the cost." In any other industry, we'd call that competition.

The Bigger Game

This isn't really about distillation—it's about market dominance. U.S. AI companies are watching Chinese competitors achieve comparable results with significantly fewer resources, and they're worried. The Biden administration seems equally concerned, establishing a new "Peace Corps" initiative last Friday to promote American AI interests abroad.

But there's a fundamental question here about innovation versus protectionism. If the goal is to advance AI capabilities globally, shouldn't knowledge sharing be encouraged? If the goal is to maintain American technological supremacy, then every technique becomes a potential weapon.