《经济学人》劝华盛顿：让中国抄美国AI

A few years ago, American AI companies competed over whose model was smarter; today, they talk more about who is copying whom.

Anthropic and OpenAI have already asked the US government to step in and stop Chinese labs from copying their AI models.

Anthropic goes further, hoping Washington will impose real penalties, not just verbal representations.

At the heart of the dispute is a practice called distillation.

It bypasses the most expensive part: no need to train from scratch, just take the output of a large model as teaching material and feed it to a much smaller model.

This way, the chaser saves computing power, and time as well.

Both companies' terms of service explicitly prohibit this practice.

On Capitol Hill, federal Senator Bill Hagerty introduced a bill authorizing the government to impose sanctions over "model theft" from abroad and to place the companies involved on the entity list.

Treasury Secretary Scott Bessent put it more bluntly: the US can sanction China over "this kind of theft."

In September, the US side accused six Chinese companies of copying American AI products.

China responded that these accusations are groundless.

The material presented at the corporate level is far more specific.

OpenAI said users linked to Moonshot AI tried to extract its reasoning process.

Such activity heated up sharply before being blocked in late July, involving more than 4,000 users and about 16,000 requests.

Anthropic, in written material submitted to the Senate Banking Committee, called Alibaba's conduct "the largest known distillation attack."

At this same time the company was also dealing with another lawsuit; it agreed to pay about 1.5 billion dollars to settle a class action brought by a group of writers.

And it is precisely here that another judgment emerges.

Some commentary argues that what Washington really should do is precisely not to ban distillation.

The reason is a figure given by experts: Chinese labs are only three to six months behind the American frontier.

Letting them catch up along this road will instead make China dependent on American output in the long run, rather than building a separate system.

By this judgment, even completely cutting off distillation would only set China back a few months, and nothing more.

What's more, these companies have argued in court that training models on copyrighted works constitutes fair use.

But what machines write is itself not protected by copyright; the same logic, turned around, applies to their own output as well.

The commentary ends with a counterintuitive conclusion: copying does not equal catching up.

The Soviet Union once built chips from stolen American blueprints, and still ended up behind.

By this reasoning, distillation may instead lock China into second place.

So Washington faces two paths: sanctions, or letting it be; and those who advocate letting it be argue that letting it be is itself a form of containment.