The Economist published an analysis on October 25, 2026, detailing how Chinese AI labs are catching up to the American frontier: not through independent training, but through 'distillation'—training models on the outputs of other models. According to estimates cited in the article, this mechanism keeps the lag of Chinese models within 3–6 months; without it, the gap would have nearly doubled. Anthropic is responding firmly: it closed the channel for leaking Claude's step-by-step reasoning and began charging even for rejected requests, causing data collection by distillers to shrink by 90–99%. We examine what this shift means for the industry and for readers choosing models.
What happened
On October 25, 2026, The Economist in its Science & Technology section analyzed the growing dependence of Chinese AI labs on distillation—building models on the outputs of other models instead of independent training. According to Saif Khan of the Center for Technology and National Security (Washington), Chinese models lag the American frontier by 3–6 months, and without distillation, the gap would have nearly doubled. Anthropic's September report revealed the scale of the leak: five Chinese firms collected over 180 million messages from Claude models in five months, and in some cases, Chinese models identify themselves as Claude or ChatGPT. The lab's response has taken concrete form: a vulnerability found in August allowing access to Claude's step-by-step reasoning has been closed, and rejected requests are now being charged. As a result, the volume of data distillers collected via Claude dropped by 90–99%; the cost of this policy for Anthropic itself is up to 5–7% of revenue.
Context
Distillation is training a student on the answers of an already prepared model, and technically it transfers an imitation of the teacher's behavior, not the ability to independently generate something new. Hence the caveat to any figures on the lag: the lag of copying models is a replication lag, not a parity of capabilities, and such models typically combine strong 'standard' answers with failures on new tasks. The most striking quantitative confirmation in the article is the DeepSeek case: a small model grew from 14.6% to 45.8% on a popular math test after training on 800,000 examples. Two structural processes work further. First: for American frontier labs, quality growth now comes almost entirely through reinforcement learning—more than half of frontier computing power is spent on it, compared to about 1/20 in early 2024, because internet text as a source of new growth is exhausted, and training is conducted on the labs' own trajectories, which distillation by design will only copy with a delay. Second: Yimin Zhang's argument from ByteDance—copying reduces the gap but does not provide an overtake; on a two-year horizon, the market will likely split into labs with their own reinforcement learning infrastructure and 'distillation followers' with a persistent lag.
Why this matters for the industry
For companies building AI products, the mechanics of cheap growth are changing: pipelines that pumped data from closed APIs have degraded—after Anthropic's countermeasures, the volume of material collected by distillers has nearly vanished, and charging even for rejected requests turns mass dumping of others' answers into a significant expense item; Anthropic itself consciously sacrifices part of its revenue for such a policy, meaning rules will tighten, not loosen. Hence practical consequences right now: recalculate the economics of data generation schemes through other models, account for rejected requests as paid events and limit retries, conduct an audit of dependencies—if product logic relies on mass collection of other models' answers, a Plan B with a change of source or provider is needed. Strategically, value shifts from the model itself to data, integrations, and distribution: demand is growing for model routing, multi-provider abstraction, and budget control of agent pipelines, while products in the form of a thin wrapper over one foreign model are becoming more expensive and less stable, because their cheap source is no longer guaranteed and can be cut off at any moment.
Why this matters for users
For the reader, this is primarily a calibration of expectations when choosing a model. Fresh Chinese models are better read as a 'version of the frontier with a lag,' not an equivalent replacement: they are strong in imitation and confidently provide standard answers at the level of Claude or ChatGPT, but are weaker where it is necessary to independently generate something new. A useful quick guide: if a model suddenly identifies itself as Claude or ChatGPT, this is a typical fingerprint of training on others' outputs—a sign that is easy to check independently, without relying on lab reports. The predictive part: if American labs' countermeasures persist, the lag of copying models, according to the cited estimate, may increase significantly, so it is not worth assuming that the next Chinese release will automatically reduce the gap. For tasks where generating something new is valued over imitating others' answers, this profile should be considered when directly comparing models.
What is still unknown / limitations
Confirmations of the method exist, but it is too early to call it 'proven': the only quantitative confirmation is the DeepSeek case, where the math test is not named, and the methodology is not disclosed, and the reproducibility of the result cannot be verified from available sources; the correct formulation is 'the mechanism is convincingly shown in one case.' Saif Khan's estimate of a 3–6 month lag and the nearly twofold growth of the gap without distillation is an expert assessment, not a measurement; the forecast of further lag growth is an interpretation. The figure of more than 180 million collected messages comes from Anthropic's own September report and has not undergone independent verification in available sources. Finally, the spread of the policy of paid refusals and closed distillation channels to other labs is a trackable hypothesis: today this is Anthropic's practice, not an established industry norm.
Sources
- China increasingly depends on copying American AI — The Economist (Science & Technology)
- China increasingly depends on copying American AI — Hindustan Times (full reprint of The Economist article)
- Discussion: China increasingly depends on copying American AI — Hacker News
Author
Look at AI, editorial team