China Is Copying U.S. AI Models at Industrial Scale. Can Anyone Stop It?

The White House and two of the biggest names in U.S. AI, OpenAI and Anthropic, are sounding alarms about Chinese entities using a technique called distillation to copy American AI models at what they describe as industrial scale. Letters to government officials from both companies cite tens of thousands of fake accounts allegedly deployed by at least one Chinese competitor. The problem: experts and lawyers say distillation is a normal part of AI development, copyright law is a poor fit, and U.S. courts may be nearly powerless to act.

What happened

Detail	Fact
Who raised the alarm	OpenAI (ChatGPT) and Anthropic (Claude)
Method alleged	Distillation using tens of thousands of fake accounts
Government action so far	White House Office of Science and Technology memo, spring 2026
Legal prognosis	Lawyers told the New York Times it would be “very challenging” in U.S. court
Perceived AI gap	Industry players now call it “very slim”
Named Chinese competitors gaining ground	DeepSeek, Z.ai, Alibaba

This spring, a memo from the U.S. Office of Science and Technology laid out the core accusation: unnamed foreign groups were attempting to “expose proprietary information, systematically extract capabilities from American AI, and exploit American expertise.” The memo set out four next steps: brief domestic tech companies, establish defensive best practices, provide resources to help those companies protect themselves, and explore ways to hold foreign actors accountable for large-scale distillation campaigns.

Since then, according to recent letters from OpenAI and Anthropic to government officials, the situation appears to have gotten worse, not better. Both companies say they believe at least one Chinese competitor has been running tens of thousands of phony accounts to pull data from their systems, replicate their capabilities, and repackage the results under different branding.

What is distillation and why is it so hard to police?

Distillation, in the AI context, means training a smaller or newer model partly on the outputs of a larger, more capable one. This is a standard technique across the industry. Many companies have shared model outputs with each other by agreement over the years, and it is widely accepted with open-source software. The controversy here is about doing it without permission, at scale, against proprietary closed systems.

The legal picture is murky. Lawyers speaking to the New York Times in June described applying copyright law to these cases as “very challenging.” There is genuine disagreement about whether the outputs of a model, rather than its underlying code, can even be protected. One lawyer told the Times directly: “This narrative that all of the capabilities of these models are coming from Anthropic is not as true as people say it is,” pushing back on the idea that distillation alone explains how Chinese competitors have closed the gap.

Adding another wrinkle: China’s leading AI players, including DeepSeek and Alibaba, largely operate on open-source principles. That openness has made some Chinese model components foundational in AI systems globally, possibly including systems built in the U.S. Blocking Chinese AI on IP grounds while U.S. developers rely on Chinese open-source work is a genuinely awkward position.

How close is the race actually?

People inside the industry now describe the U.S.-China AI development gap as “very slim.” Startups like DeepSeek, Z.ai, and Alibaba have produced models that industry observers say are roughly as capable as leading U.S. products, but cheaper to build and cheaper to run. That cost advantage matters enormously for adoption, particularly in markets that are already skeptical of U.S. technology. Much of the world, according to reporting cited in the source, now perceives China as leading the sector.

A separate trend worth watching: Chinese tech professionals who trained in the U.S., through education or work experience, are reportedly returning home in growing numbers, taking that expertise with them. That is a slower-moving but potentially more durable knowledge transfer than distillation.

If you want to follow how these AI developments are reshaping the competitive landscape, our AI news coverage tracks the story as it moves.

Why it matters

For businesses evaluating which AI tools to build on, the distillation debate is a signal about how quickly capability gaps close. A model that was a clear leader eighteen months ago may now have a cheaper, comparably capable alternative built by a competitor who partially bootstrapped on the original’s outputs. That is good for buyers in the short term (more options, lower prices) but raises real questions about which providers can sustain investment in genuine R&D when competitors can shortcut years of work.

It also matters for anyone thinking about AI integration inside their own products. If the underlying models from major U.S. labs are being reproduced and redistributed under different names, knowing what you are actually running and where the data goes becomes more important, not less.

Our take

The framing of this as a clean case of theft versus innovation is too simple. Distillation is genuinely a shared practice, and the line between “learning from” and “stealing from” a model’s outputs is not settled in law or in technical norms. OpenAI and Anthropic have a real grievance, but they also have an incentive to push for regulation that protects their moats.

What is harder to argue with is the cost dynamics. If Chinese models deliver similar results at lower inference costs, the market will use them, regardless of how they were built. The more practical question for most businesses is not who to root for geopolitically, but which model actually performs best on their specific task at the price point they can afford. Run your own benchmarks. The answer changes faster than any policy memo will.

What to do about it

1. Audit which AI providers your products and workflows currently depend on, and check their terms of service around distillation and data use.
2. Run head-to-head tests comparing U.S. and Chinese model outputs on your actual workload, not synthetic benchmarks, before committing to a vendor.
3. Watch the open-source landscape closely. Models like those from Alibaba are available now and may already be competitive for your use case.
4. If IP protection matters to your own outputs, talk to a lawyer about what your API usage agreements actually cover before assuming you are protected.

The policy fight will take years. Your infrastructure decisions cannot wait that long, so test with specifics and stay flexible about which providers you rely on.

Source: Bing News · Anthropic