Distillation via API: What really gets copied, and what the defenses achieve

According to a Reuters investigation, training your own AI models on answers from US models is now routine in China. We explain how this kind of harvesting through an API works technically, what research says the defenses achieve, and what distillation is good for: cheap specialist models, yes, but hardly a complete frontier model.

Analysis Nov 17, 2026 9 min read Original analysis by our editors
Illustrative image, AI-generated

What this is about

On November 17, Reuters published its analysis of more than 300 Chinese documents: research papers, patents, company reports, job ads and government papers. The finding: distilling US models is no longer a university experiment in China but standard industry practice. In distillation, a strong "teacher model" generates answers that are used to train a smaller "student model". Beijing encourages it. A Washington official told Reuters that industrial-scale distillation is the "main reason" Chinese models have caught up.

Most coverage stops at the political question. We look at the technology and put forward a thesis: through an API (application programming interface), you can mainly copy narrow capabilities cheaply into small specialist models. You don't get a frontier model that way. To test this, we look at what research says the defenses achieve, how they can be bypassed, and what the documents show was actually distilled.

What Reuters actually documents

According to Reuters, an "AI+Manufacturing" policy package from eight Chinese agencies, issued in January, explicitly recommends training, fine-tuning and distillation for industry. Shanghai and Shenzhen are funding techniques that need less compute. No government document says US models specifically should be distilled. Reuters reports that Chinese developers have asked the regulator MIIT where normal use of foreign models ends and organized harvesting begins. The MIIT hasn't responded. Companies are now building software to automate the process and hiring specialists to prepare the material. A former Alibaba engineer sums up the appeal: one expensive query pays off again and again.

What was distilled, and for what? We compiled the specifically documented cases from several investigations:

Source Teacher Purpose Student
Reuters, Nov. 2024 (Shanghai Jiao Tong University) OpenAI o1, GPT-4o mini 5,000 answers selected and translated, yielding 10,750 bilingual training examples Qwen2.5-72B
Reuters, Apr. 2025 (Alibaba) GPT-4, GPT-4o Training material in languages other than Chinese smaller Qwen models
Reuters, Feb. 2026 (Shanghai Jiao Tong, Meituan) GPT-4o Plans for how AI agents collaborate on coding smaller model
Reuters, July 2026 (military) GPT-3.5 Summarizing military source code lightweight model for internal networks
Reuters, July 2026 (military) Claude 3 Haiku Synthetic data for a classifier that monitors social media classifier
Anthropic, Feb. 2026 Claude DeepSeek: over 150,000 queries (reasoning, rubric-based grading); Moonshot: over 3.4 million (agents, computer use); MiniMax: over 13 million (agentic coding) in-house models

The list shows that most cases involve translation, synthetic training data and narrowly defined tasks for small models. According to SiliconANGLE, Reuters characterized the use in the military papers as an attempt to transfer expensive reasoning capabilities into "smaller systems that you control and run locally". Reuters itself also writes that distillation can't fully replicate a frontier model and doesn't eliminate the need for compute.

Anthropic's numbers look somewhat different. In total, they involve 16 million queries through about 24,000 fake accounts. Do the math, and roughly 81 percent of those come from MiniMax and less than 1 percent from DeepSeek. So the leading labs were targeting specific capabilities such as agentic coding, not "everything". These figures come from Anthropic; the labs haven't confirmed them.

How distillation via API works

In classic distillation, the student sees the teacher's full probability distribution. It doesn't just learn "the right word is X", but also "Y would have been almost as good". Through an API, you get none of that. For reasoning models, OpenAI refuses to return probabilities (logprobs) as a matter of policy, according to its developer forum and the Azure documentation. What's left is the finished text. So the student is trained on example answers, as in normal fine-tuning, which means learning by imitation.

The most valuable part is the reasoning trace (chain of thought), the intermediate steps the model takes to reach its answer. That's exactly what providers withhold. Anthropic says it only shows a summary. According to the Claude documentation, you're still billed for every reasoning token generated internally. The thinking blocks come back encrypted and have to be sent along with the next request. Anthropic measured what the summary costs: a pretrained Claude model reached 80.4 percent on the SWE-bench Verified coding benchmark when trained on Claude Opus 4.8's full reasoning traces, and 71.8 percent with the summarized ones. Our take: that's a significant loss, but the summary still delivers a lot.

What research says the defenses achieve

Defense Result Caveat
Paraphrasing answers (DistillGuard) Student barely gets worse –
Poisoning the training data (DistillGuard) Mainly hurts conversational quality Domain skills largely survive
Withholding the reasoning trace (DistillGuard) Math: 67.8% → 31.4% No effect on coding
Rewriting the reasoning trace (Ma et al.) Student up to 61.3% worse on GSM8K/MATH, teacher even improves Small students (1–3B parameters)
Watermarking reasoning traces (Ma et al.) Detection mostly 98–100%, almost no false positives Only measured on GSM8K
Anthropic's full package 90–99% fewer harvested reasoning tokens Self-reported

The DistillGuard paper by Bo Jiang tests nine defense variants. The teacher is Qwen3-14B, the student Qwen2.5-7B, and the attacker isn't especially sophisticated. The author's conclusion: today's methods aren't enough to "broadly prevent" knowledge theft. How well they work depends heavily on the task. Rewriting the reasoning traces, as in the paper by Xinhang Ma and colleagues, has a stronger effect. But it was tested with a small teacher (DeepSeek-R1-Distill-Qwen-7B) and very small students.

Anthropic relies on three levers. First, it controls where reasoning traces can go. API accounts created on or after August 31, 2026, can no longer modify Claude's earlier conversation history, and older, less protected models can no longer read the reasoning traces. Second, a two-stage classifier screens requests, with a false-positive rate below 0.1 percent. Third, there are bans and stricter rate limits, and since September 24, blocked requests are billed too. According to eSecurity Planet, the 90 to 99 percent figure only measures how many reasoning tokens reached suspected harvesters on Anthropic's own platform. Our take: anyone who goes undetected doesn't show up in that statistic. Anthropic itself admits it hasn't yet seen a "sophisticated, persistent" evasion attempt, and says it's unclear whether its lead will hold.

The counterattacks

Two new papers show that the defenses have gaps. The EMNLP paper by Hongyu Zhu and colleagues (Shanghai Jiao Tong) presents a method called SIFT. It targets "antidistillation sampling", where the provider deliberately injects perturbations as each word is generated. According to the authors, these poisoned traces only hide the learnable signal behind a few particularly hard-to-predict tokens. Filter out those spots, and you get up to 21 percent more than with standard training. Their advice: a good defense has to reduce how much can be learned from a trace in the first place, rather than just injecting as much surprise as possible.

Even more direct is the paper "Stealing Reasoning Traces from Proprietary LLM APIs" (Panfilov, Andriushchenko et al., August 2026). The encrypted thinking blocks from Anthropic, OpenAI and Google could be fed into a weaker model from the same provider, which then output them in plain text. According to Developers Digest, the authors checked whether the recovered texts were genuine mainly like this: on 120 Codeforces problems, they compared the reasoning-token count reported by the API with the length of the recovered text. Our take: the same length doesn't mean the same content. That's not enough to prove a verbatim reconstruction. Security researcher Johann Rehberger independently reproduced the attack against OpenAI. According to the Cloud Security Alliance, the published tricks stopped working from mid-August. Anthropic's move to bar older models from reading reasoning traces starting August 31 fits that timing and content exactly. Whether it was a direct response, Anthropic doesn't say.

The "main reason" China caught up?

Anthropic's October report supports Washington's thesis. It says Zhipu deliberately distilled cyber capabilities from US models for GLM-5.3. For narrow tasks, distillation demonstrably works very well. DeepSeek distilled its own reasoning chains into Qwen-32B using 800,000 examples. According to the model card, the result beats OpenAI's o1-mini on AIME 2024 (72.6 to 63.6) but trails it on Codeforces coding (1691 to 1820). But DeepSeek had full access to its own teacher, which a harvester doesn't.

A finding from Berkeley argues against the thesis. The paper "The False Promise of Imitating Proprietary LLMs" showed that imitation mainly transfers style. More imitation data doesn't close the gap; what matters is a better base model. On the accusation against Moonshot, Braden Hancock (Laude Institute) said, according to Implicator, that the two weeks between the launch of Claude Fable and the preview of Kimi K3 were too short for distillation. Researcher Nathan Lambert considered its contribution small. The White House didn't present any evidence. Moonshot credits its progress to its own architecture.

The bottom line

Our take: the Reuters documents mainly support a narrower claim. In China, distillation is a cheap tool for breadth: translation, data generation, small models for factories, government agencies and the military. The documented cases don't prove that China's frontier models are essentially copies. That's no reason to relax, though. Specifically harvested capabilities such as agentic coding or cyber know-how can shortcut in-house development, and those are exactly the capabilities that matter for security.

On the technical side, withholding the reasoning trace is what works best, and that's exactly where providers are focusing. But every layer of protection already has a published countermeasure. And the success figures so far only measure what was detected. The legal line is also unresolved. A US bill would exempt training that complies with the terms of service. Chinese developers, meanwhile, are still waiting for their regulator to tell them what's allowed.

#distillation#china#anthropic#openai#security#research

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.