OpenAI, Google, and Anthropic Unite Against Chinese AI Distillation: IP Theft or Industry Bullying?

On April 7, something unusual happened in the AI world.

OpenAI, Anthropic, and Google - three companies that normally compete fiercely against each other - suddenly stood on the same side.

Through an organization called the Frontier Model Forum, they issued a joint statement targeting three Chinese AI companies: DeepSeek, MiniMax, and Moonshot AI.

The accusation: using “adversarial distillation” to steal American AI model capabilities.

What Is Distillation, Exactly?

Knowledge distillation is a legitimate machine learning technique. Simply put: you use a large model’s (teacher) outputs to train a smaller model (student), transferring some of the larger model’s capabilities.

This technique has been used in academia for years, completely legal. Using GPT-4’s API to generate high-quality training data for your own model - that’s technically distillation.

But the US companies are alleging “adversarial distillation” - not normal API usage, but systematic capability extraction through fake accounts and automated scripts.

The Numbers Are Staggering

Anthropic’s investigation report cited specific figures: approximately 24,000 fake accounts initiated over 16 million abnormal interactions with Claude.

16 million.

If that number is accurate, it’s clearly beyond what “normal API usage” can explain. This looks more like organized, industrial-scale data harvesting.

OpenAI moved first - submitting a special memorandum to the US Congress in February 2026, naming DeepSeek specifically. Google followed, reporting similar anomalous attack patterns.

Three companies sharing risk monitoring data, unified messaging, maximum firepower.

IP Protection or Industry Bullying?

There are two completely different ways to read this.

Reading one: legitimate intellectual property protection. If Chinese companies truly extracted model capabilities through abnormal means at scale, that’s technical theft and should be pursued. API terms of service explicitly prohibit using outputs to train competing models.

Reading two: a pretext for a new form of tech blockade. The timing of the three giants’ joint action is suspicious - right before DeepSeek V4’s launch, during a period of rapid global market share growth for Chinese AI companies. Using distillation accusations to suppress competitors isn’t unprecedented in business.

My personal lean: the truth lies somewhere between the two.

Questions Worth Asking

First, where’s the line for distillation? Using API outputs to train models is something virtually every AI company does. The difference is scale and method. 24,000 fake accounts is clearly excessive, but what about doing the same thing with legitimate accounts at normal volumes?

Second, who defines “abnormal”? 16 million interactions sounds like a lot, but for a large AI company’s R&D team, that volume isn’t outrageous. The key is the pattern - whether it’s clearly systematic data harvesting. Currently, that judgment is entirely defined by the US companies themselves.

Third, what about open-source models? DeepSeek V4 is open-source. If an open-source model’s capabilities happen to resemble a closed-source model, how do you prove it was distilled rather than independently developed? Technically, this is nearly impossible to prove.

The Bigger Picture

This can’t be separated from the broader US-China tech competition.

Chip bans, entity lists, export controls - the US has already systematically restricted China’s AI industry at the hardware level. Now they’re opening a new front at the software level, using distillation accusations to limit Chinese companies’ access to model capabilities.

From the US perspective, this is protecting IP. From China’s perspective, this is tech bullying.

Both narratives have merit. Both have holes.

My Take

As someone who’s written plenty of code and used plenty of APIs, my stance on distillation is: the technique itself isn’t wrong, but the method matters.

If 24,000 fake accounts were really used, that’s not a good look. But attributing all of Chinese AI’s technical progress to “stealing” massively underestimates China’s AI R&D capabilities.

DeepSeek V4’s sparse attention architecture, Ascend optimization, MoE innovations - you can’t distill those.

This debate won’t be resolved anytime soon. But it reminds us of one thing: in the AI race, technical competition and geopolitics are now deeply intertwined.

As developers, all we can do is write good code, use compliant tools, and not let narratives lead us around.