Executive Summary
Chinese researchers and companies ask a leading US model for answers and feed those answers to a smaller domestic model as training material. That work has left the experiment stage and become a production process. Reuters reported it exclusively on 17 November, after reviewing more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents. Software automates the collection of the answers, and there are staff assigned to nothing else.
The documents Reuters counted were never hidden, though. They were published as papers, filed as patents, posted as job ads. And the paper the report held up as its leading example happens to be the one that criticizes distillation at the greatest length, and the one that built a scorecard for whether a project records where its data came from. Scored on that card, five of seven attempts in the same field came out at zero on the data-source item.
What these teams are doing fills public documents to overflowing. Which model the answers came out of, and which model they went into, is the part that does not stay with the model. So the question a reader is left with is not who took something in secret. It is what the model we picked up was trained on, and where the fact that our own synthetic data came from another model's output gets written down.
5 of 7
o1 replication attempts scoring zero on data source
Scored on the card the paper itself built. The limit is that the authors also graded their own work
300+
Chinese documents Reuters reviewed
Papers, patents, corporate disclosures, job postings and government documents. Not one of them confidential
13 vs 12
AIME2024, 30 problems: distilled model against o1-preview
o1-mini solved 21 in the same table. The margin behind "surpassing" is one problem
69.08 → 62.65
Chinese factuality after distillation fine-tuning
They distilled to build a Chinese-language model, and of nine metrics this is the one that fell furthest
The 300 Documents Reuters Reviewed Are All Public
The exclusive Reuters filed from Shenzhen on 17 November rests on a review of more than 300 documents out of China. The article states its own scope: "A Reuters review of more than 300 Chinese academic papers, patents, corporate disclosures, recruitment notices and government documents reveals how the practice — known as distillation — has expanded beyond academic experiments into mainstream industrial use in China." Five kinds of document, and every one of them sat where anyone could pick it up.
Distillation means asking a large model for answers and using those answers as training material for a smaller one. The technique is not new, and why it leaves the distiller leaning on whoever holds the original data was the subject of an earlier report, You Can't Buy Sovereignty by Distilling It. What this story adds is not the technique but the way the work is now run. The records Reuters read describe Chinese researchers getting answers out of US models, generating code and finding solutions to hard problems, while Chinese companies build software that automates that loop and assign dedicated staff to refine the training material for their own smaller models. That is what "industrial scale" means here. Work one person used to do by hand, query by query, now has someone who owns the process.
Where this meets export controls is plain enough. Withhold advanced chips and compute goes down; copy down the answers another company's model produces and training data goes up. Reuters wrote that widespread distillation lets Chinese developers build capability on US output rather than US silicon, which blunts the effect of limiting chip sales. There are two channels, and the controls sit on one of them.
Compute and output travel different channels. The top one is what chip export controls and the 16 November interim rule on remote access aim at; the bottom one is where those controls have not yet reached. The diagram marks where the controls sit, not the relative volume each channel carries.
How the chip channel gets crossed is something we looked at five days ago, in A Blacklisted Chinese University Buys Nvidia H200 Chips, and Rents More, which collected the records of a sanctioned Chinese university both buying Nvidia H200s and renting server time. Chips get through by purchase or rental, and answers get through by being copied down.
How Washington reads this channel is in the same article. A US official familiar with the Chengdu talks told Reuters that Washington sees industrial-scale distillation as the main reason Chinese models have narrowed the gap. That assessment rests on one unnamed official, and the same article carries the rebuttal. Two people familiar with Chinese developers pushed back, calling distillation one factor among several, alongside engineering, algorithms, data and model design. None of Shanghai Jiao Tong University, Alibaba, Meituan, the US Commerce Department or the Cyberspace Administration of China responded to requests for comment.
Three Rungs From a University Lab to a Company Process
Reuters shows the move from experiment to process through three cases. It starts in a university lab in 2024, goes on to training a company model in 2025, and by 2026 reaches the training of a model that coordinates several AI agents. What the three rungs show together has less to do with technique than with where the record lives. All three are written up in papers, and those papers are public.
The three rungs do not rest on equally solid ground, though. The first names a paper we could open in full; for the other two, Reuters wrote only "an April 2025 paper" and "a February 2026 paper", and we could not identify which papers those are. The last column of the table below records that difference as it stands.
| Date | Who | What they did | How far we could verify it |
|---|---|---|---|
| November 2024 | Shanghai Jiao Tong University | Picked about 5,000 o1 responses, translated them into Chinese with GPT-4o mini, built 10,750 bilingual pairs and further trained Qwen2.5-72B-Instruct | Paper identified We read arXiv:2411.16489 in full |
| April 2025 | Alibaba | Used GPT-4 and GPT-4o to produce training material in languages other than Chinese, refined and scored the responses to filter them, then trained a smaller Qwen | Reuters description only We could not identify the paper |
| February 2026 | Shanghai Jiao Tong University and Meituan | Used GPT-4o to produce collaboration plans for AI agents, then trained a small model to coordinate that work | Reuters description only We could not identify the paper |
Source: Reuters, 17 November 2026. These are the three cases Reuters names in the body of its article, carried over as written; the table is not a distribution of all 300 documents the report reviewed. The figures in the first row were re-checked against the paper itself, and no arXiv number or model name has been attached to the two rows below it.
Line the three up and the change shows. In 2024 the raw material was a few thousand responses a person had picked out. On the 2026 rung the material is itself a work plan another model produced, and what gets trained on it moves up to the seat that coordinates several models. The job ads and patents Reuters read show that somewhere in between, a role appeared whose job is the collecting and the filtering.
Original Pebblous diagram. The three rungs of the ladder side by side: moving left to right, the training material shifts from responses a person picked to a work plan another model produced. Box color marks whether we verified the case against the original paper (orange) or it rests on Reuters' description alone (gray); it is not a measure of the three cases' relative scale or importance.
This is not the only channel through which China sources training data. In October a state-owned company was founded to have the state itself supply data, which we covered in China's National Data Group Supplies the Data That AI Learns From. One channel produces the data at home and delivers it; the other copies it down from someone else's model. Whether the two sit under the same policy package remains unverified.
The Numbers Reuters Quoted Come From the Paper's Side Experiment
The figures the report gives for the first rung are 5,000 o1 responses, 10,750 bilingual pairs and Qwen2.5-72B. Open the paper and those numbers turn out to sit somewhere other than where the report seems to put them. All three come from §4 of the paper, "Application Beyond Math Reasoning", an experiment appended at the end to see whether a model distilled on mathematics carries over outside mathematics. The paper's headline experiment used a different base model and a different volume of material.
| Headline experiment (§2.3, §3) | What the report quoted (§4) | |
|---|---|---|
| Base model | Qwen2.5-Math-72B | Qwen2.5-72B-Instruct |
| Training material | Long chains of thought distilled from o1, "tens of thousands of samples" | About 5,000 picked from o1 output → translated into Chinese with GPT-4o mini → 10,750 pairs |
| What it tested | Mathematical reasoning on AIME and MATH500 | Whether it carries over outside mathematics |
| Result | AIME2024 13/30 (o1-preview 12/30, o1-mini 21/30), MATH500 87.2 (85.5 / 90.0) | Of nine metrics, four up, one flat and four down |
Source: arXiv:2411.16489 v1 (submitted 25 November 2024, accessed 18 November 2026), §2.3, §3, §4 and Tables 1 and 2. The paper has no v2.
The paper describes the §4 data like this: "From our distilled O1 model outputs, we carefully select approximately 5,000 high-quality samples containing retrospective thinking and self-reflection elements. These samples are then translated into Chinese using GPT-4o mini model, resulting in a balanced bilingual dataset. The final training dataset comprises 10,750 mixed Chinese-English sample pairs." The 5,000 is a count picked a second time out of output that had already been distilled, not a count of the questions put to o1.
The claim on the headline side is worth measuring as well. The paper writes that a base model fine-tuned on tens of thousands of o1-distilled chains of thought surpassed o1-preview on AIME. The value in Table 1 is 13 problems against 12, out of 30 on AIME2024. One problem. In the same table o1-mini solved 21, and MATH500 reads 87.2 against 85.5 (o1-preview) and 90.0 (o1-mini). The paper says as much itself: "there remains a noticeable gap compared to O1-mini in terms of mathematical reasoning performance." The model it caught and the model it trails sit in the same table.
And distillation is not free. Table 2 in §4 compares nine metrics before and after fine-tuning, and since the caption reads "Performance comparison (accuracy) before and after SFT", higher is better in every row. For the metric that measures how well a model resists flattery, the paper notes separately that a higher percentage means holding up better. Here are all nine rows, with none left out.
| Metric | Before SFT | After SFT | Change |
|---|---|---|---|
| LIMA | 77.2 | 87.2 | +10.00 |
| Auto-J | 81.6 | 88.0 | +6.40 |
| ChineseFactEval-Sycophancy | 89.70 | 92.65 | +2.95 |
| Flames | 91.0 | 92.5 | +1.50 |
| DiaSafety | 100.0 | 100.0 | 0.00 |
| SimpleQA | 10.58 | 10.41 | −0.17 |
| C-SimpleQA | 47.08 | 45.76 | −1.32 |
| WildSafety | 92.0 | 86.5 | −5.50 |
| ChineseFactEval-General | 69.08 | 62.65 | −6.43 |
Source: arXiv:2411.16489 v1, Table 2, all nine rows. The change column is the arithmetic difference between the two values and is not a column the paper supplies. The metrics are on different scales, so reading the column vertically tells you little.
Four metrics went down, and the largest drop is in Chinese factuality. They translated into Chinese and trained on it in order to build a Chinese-language model, and the test that asks about Chinese facts lost 6.43 points. The paper's own wording is weaker than that, though. The only drop it calls pronounced is WildSafety, and of the three factuality metrics it says they "did not demonstrate significant improvement". The arithmetic in the table and the wording in the prose differ, so both are recorded here.
The Same Paper Criticizes Distillation for a Whole Section
The last section of that same paper is titled "The Bitter Lesson of Simple Distillation". Reuters compressed it into a single sentence: "excessive reliance on distillation could stifle original innovation, leaving developers dependent on breakthroughs made by more advanced model providers." In the original, that warning is not one sentence. It is a section with six subheadings under it.
| Subheading | What it says |
|---|---|
| SURFACE APPEAL | Distillation looks like an elegant solution and performance climbs quickly. But "this convenience comes at a price that may not be immediately apparent" |
| PERFORMANCE CEILING | "Models trained through distillation are invariably bounded by the capabilities of their teacher model… This creates an implicit ceiling effect" |
| MISSED INNOVATION | It skips what actually made o1 work, inference-time scaling and search optimization. "we risk creating a widening technological gap" |
| RESEARCH CULTURE SHIFT | "The availability of 'easy wins' through distillation has begun to shift research focus away from tackling fundamental challenges." With no infrastructure, teams lean on distillation; leaning on it, they build less infrastructure, and the loop closes |
| EROSION OF FUNDAMENTALS | "This shift from 'how it works' to 'what works' represents a fundamental change in research mentality" |
| GROWING DIVIDE | "genuine breakthroughs become the exclusive domain of a small number of well-resourced organizations, while the broader community remains trapped in a cycle of incremental improvements through distillation." |
Source: arXiv:2411.16489 v1, all six subheadings of §6. Quoted text is verbatim from the paper; the summary around it is ours.
Of the six, the one that matters most to anyone working with data is the ceiling effect. A model trained by distillation is bounded by the capabilities of its teacher model, and that bound acts as a ceiling that nothing on the surface shows. Benchmark scores keep rising all the way up toward the teacher, so the score is not where the signal of having hit the ceiling shows up first.
At the close of the section the authors step back.
"The bitter lesson here is not that distillation is inherently problematic - it remains a valuable tool in our technical arsenal. Rather, the danger lies in allowing the convenience of distillation to divert us from the harder but ultimately more rewarding path of fundamental innovation."
The warning is aimed at the convenience, not at the method, and the same sentence calls distillation a valuable tool.
Reading this section as a Chinese researcher's confession of conscience points the story the wrong way. The ten authors are not only from Shanghai Jiao Tong University, SII and GAIR; New York University is in there too. And what they did was not confess. They built a scorecard and graded the work of their own field with it. That scorecard is what Section 6 below is built on.
Terms of Use Restrict It, a Policy Document Encourages It, and No Law Covers It
No country has a law banning distillation. Reuters wrote that "Distillation is not illegal and is prevalent in other parts of the world." What restricts it is terms of service. The same article puts it this way: "OpenAI's terms restrict using its output to develop competing models, subject to exceptions, while Anthropic similarly restricts training competing systems without permission." The OpenAI restriction carries the qualifier that exceptions apply, and Anthropic's turns on permission rather than on an outright prohibition. Render either as a blanket ban and you have made it stronger than the original.
What moved in place of a law was a warning about enforcement. Treasury Secretary Scott Bessent wrote this on X on 22 July.
"We support open-source AI and the innovation it unlocks. But open source is not open season on American IP. When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table."
There are two conditions for sanctions and Entity List designations to be on the table. The conduct has to be covert, and it has to cross the line into IP theft. The sentence reads as requiring both. But the 300 documents Reuters counted this time are papers, patents and job ads. Enforcement is pointed at what was hidden, and what this report counted is what was written down.
Line up the four US records and the mismatch gets sharper. The Chengdu meeting in the first row came to light through this report.
| When | What happened | What it targets |
|---|---|---|
| July | US Commerce Department officials met Chinese regulators in Chengdu, raised the collection of output from leading US models as a problem, and warned that some of it could amount to IP theft. The Chinese side defended distillation as a standard development technique and said it would act where theft was involved, but made no specific commitment (three people familiar with the talks) | A closed-door meeting |
| 22 July | Treasury Secretary Bessent signaled sanctions and listings, with conditions attached | Conduct that is covert and crosses the line into IP theft |
| 8 September | NSA, CISA and the FBI released joint cybersecurity advisory AA26-251A | Circumvention through the gray market in API proxies |
| 16 November | The Commerce Department's Bureau of Industry and Security (BIS) announced the interim rule "Addressing Remote Access to Advanced Semiconductors" (RIN 0694-AK50) | Remote access to compute through overseas data centers |
Sources: Reuters, 17 November 2026 (the July Chengdu meeting), Bessent's post, the advisory on cisa.gov, and the BIS interim rule. This is not a complete count of US government action on distillation; it collects only the four items that bear on this report.
The September advisory is titled "China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies", and it names six firms: DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI. The three agencies wrote that distillation is "not a supplement to these companies' AI model development, but the critical core of it", and that millions of exchanges and requests since late 2024 have pulled out billions of tokens. The route they pointed to was "a large gray market of API proxies, or 'transfer stations'". We covered that covert route separately in September, in Chinese Labs Rerouted Their Own Users' Requests to Claude.
What the 16 November rule closes off is not the chips themselves but remote access. It puts a license requirement on the path by which a controlled entity uses US-made accelerators remotely through data centers in Indonesia, Malaysia, Singapore, Thailand and Brazil, with the threshold set at 12 million TPP in aggregate. We sketched the shape of that opening back in October, in When Tencent Rents AI Chips From Oracle, Where Does the Data Go?. The rule takes effect on publication in the Federal Register, though, and publication is set for 19 November, so as of this writing it has not taken effect. The accurate statement is not that the path is closed but that a rule to close it has been issued.
There is a document on the other side as well. Reuters wrote that the "AI+ manufacturing" policy package (《"人工智能+制造"专项行动实施意见》) issued in January by eight Chinese government departments contains guidance on model training and fine-tuning for industrial applications, and guidance that supports distillation. That passage could not be verified against the Chinese original, so it is carried here only as Reuters described it. A technique restricted by the terms of service written in one country is encouraged by a government document in another.
Chinese developers find the boundary unclear as well. Reuters wrote that "Chinese AI developers have in recent months asked Beijing's industry regulator for guidance on the boundaries between ordinary use of foreign models and organised collection of output for training". What they put to the regulator is where ordinary use of a foreign model ends and systematic collection of its output for training begins, and the source for it is one person familiar with the discussions. The report came out in the same week as the APEC summit in Shenzhen, and a bilateral meeting between President Trump and President Xi Jinping is scheduled for 19 November. No outcome of that meeting was available at the time of writing.
What They Do Is Written Down, What They Built It From Is Blank
Every record so far has been the same kind. We do distillation, and this is the method we use. Call that a record of the activity. It sits in papers, in patents and in job ads, which is how Reuters could count 300 of them. But the record anyone working with data needs is a different kind. This model's output went into that dataset and came out as this model. Call that a record of lineage. It belongs in the model card and the dataset metadata, and that is where it is blank.
| Record of the activity | Record of lineage | |
|---|---|---|
| What it records | That we do distillation, and by what method | That this model's output went into that dataset and came out as this model |
| Where it lives | Papers, patents, job ads | Model cards, dataset metadata |
| What this reporting showed | More than 300 | Five of seven scored zero |
| Status | Overflowing | Blank |
The left column rests on the Reuters report, the right column on Table 4 of arXiv:2411.16489 v1. The two figures count different things, so subtracting one from the other or binding them into a ratio would blur what each one means.
What turned the right-hand column into a number was that same paper. In §5 the authors built a scorecard they call the Technical Transparency Index and graded o1 replication attempts out of 100: 14 points for data, 33 for methodology, 24 for evaluation and 29 for open sourcing. Here is what the first item, Data Transparency, asks.
"This aspect evaluates whether the origin of the data is clearly specified, including detailed descriptions of the datasets used and their respective sources. It considers whether the dataset names, providers, or publications from which the data is derived are explicitly mentioned. This applies to all datasets used in downstream tasks such as supervised fine-tuning (SFT), reinforcement learning (RL), or search algorithms and becomes even more crucial when the datasets serve as seed data for synthesizing long thought data."
The two sub-items under it are Data Source and Data Selection Process. And the last sentence points straight at what this article is about: the item matters most precisely when the data is the seed for synthesizing long chains of thought. That caveat was written in 2024, and the 300 documents Reuters counted in 2026 are exactly that case.
The authors graded seven o1 replication attempts on this card. Table 4 in full is below.
| Project | Data (14) | Methodology (33) | Evaluation (24) | Open sourcing (29) | Total (100) |
|---|---|---|---|---|---|
| Open O1 | 0 | 8 | 20 | 5 | 33 |
| O1-Journey (Part 1) | 10 | 33 | 24 | 9 | 76 |
| LLaMA-O1 | 0 | 6 | 0 | 5 | 11 |
| K0Math | 0 | 0 | 16 | 0 | 16 |
| Skywork O1 | 0 | 0 | 0 | 0 | 0 |
| DeepSeek-R1-Lite | 0 | 0 | 20 | 0 | 20 |
| O1-Journey (Part 2) | 10 | 33 | 24 | 12 | 79 |
Source: arXiv:2411.16489 v1, Table 4, all seven rows. The four items in each row add up to the total. The graders are the authors of Part 1 and Part 2.
Read the seven rows across and the zeros cluster in one column. Of the five that scored zero on data source, four picked up points elsewhere: three on evaluation, two on methodology and two on open sourcing. Only one project scored zero on all four items. The case where the data column went to zero because the whole project was thin stops at that one.
Stand the data column on its own and it looks like this. Five of the seven are at zero, and the top two reach 10 out of a possible 14.
Source: the data column of Table 4 in arXiv:2411.16489 v1. All seven, with no rows left out. The zero bars are drawn as a thin rule so they stay visible. The graders are the authors of the two projects that scored 10.
There is one more layer to this. In the prose the authors describe their own work as "achieving a perfect transparency score", while the total written for them in the same table is 79. Ten on data and 12 on open sourcing, so neither item is at the maximum. The side that built the scorecard did not fill in its own scorecard either. On top of that, one independent review noted that the index is self-scored and that the reported scores cannot be verified from the manuscript alone. The party that designed the index and its weights is the party that assigned the scores. The "five of seven" this article cites carries that same limit.
Is the reason nothing gets written down simple laziness, then? Not really. It is closer to a mismatch between where the writing happens and where the model travels. The derived model this paper produced was never released at all, as a footnote to the abstract states: "Per OpenAI's Terms of Use, our distillation of the OpenAI O1 series models is strictly for research purposes and will not be fully disclosed publicly." The writing happened in the paper, and the model never came out. This is not a case of a model card leaving out the teacher model's name. It is a case of there being no model.
The blank is one layer below that. The model card for Qwen2.5-72B-Instruct, the model this experiment rode on top of, has a base-model field but no datasets field, and no passage describing where the training data came from. DeepSeek-R1-Distill-Qwen-32B, on the same hub, does the opposite. Its metadata gives Qwen2.5-32B as the base model, and its body states that it was fine-tuned on 800,000 reasoning samples generated by DeepSeek-R1. The teacher model's name and the sample count are both on one card. We opened both on 18 November 2026.
The place to write it already exists. Writing it is simply optional. One hub carrying the card that fills it in alongside the card that leaves it empty is the proof that the place is there.
The blank is not confined to this one card. The proposal to record training data on a model card dates to 2019. "Model Cards for Model Reporting" put training data among nine standard items, and the hub's base-model and dataset fields carried that proposal into practice. Both are recommendations, though, and the entry is made by the owner. A study that audited the whole hub reported, accordingly, that the entries are uneven. Even quantized models, which by definition rest on a base model, often left that field unfilled, and 1,416 cases put a model name where a dataset belongs.
What about the obligation side? The strongest demand for training-data disclosure is the public summary of training content under Article 53(1)(d) of the EU AI Act. But that template asks for data sources by category. It asks whether the data is a public dataset, a crawl, user data or synthetic data, and it does not ask whose model output that synthetic data came from. When we compared the summaries actually filed, in Big Tech Left Half of the EU's Training-Data Summaries Blank, the contested points were user data and crawling practice, not distillation. And the Chinese companies in this report are not subject to that obligation in the first place.
Why This Matters to Pebblous
Pebblous builds data that AI can use. Where this report touches that work is not which country did what, but the blank in the record. A structure in which one model's output becomes the next model's training data is already the global default, and it is not a Chinese peculiarity. What changed is that the work has turned into a process.
7.1A Process Needs a Record Format
Work one person did by hand, query by query, needs no format. It changes once software runs the collection and dedicated staff are attached to it. Which model, what queries, how many items, filtered by what rule, translated into what. The scope of what Pebblous calls AI-Ready Data widens to take that in. And those questions come from the scorecard in Section 6. Its first item asks two things, the source and the selection process, and the questions above are those two written out as units of work.
7.2Data Provenance Is a Debugging Tool, Not a Compliance Document
The paper itself left the evidence, and Section 3 already showed it. In Table 2 the steepest fall of the nine rows lands on Chinese factuality, and a Chinese-language model is what that bilingual training set was built for. If nobody wrote down which model's output went into the mix and how much of it, then once the score drops there is no telling what to take out to bring it back up. That is why recording provenance is quality work rather than audit preparation. It is the same reason DataClinic asks how a dataset was generated before it asks anything else.
7.3If You Deploy an Open Model, You Are Not a Bystander
When you take an open-weight model and put it in a product, what that model learned from is usually not written on its model card. The card for Qwen2.5-72B-Instruct, the one Section 6 opened, is one of those. If there was terms-of-service risk upstream, the risk travels downstream, and if there was a quality defect upstream, the defect travels too. We looked at one model shipped with its training-data sources left blank in Perceptron's Open-Weight Factory Robot Model Leaves Its Data Sources Blank. Three lines are worth adding to a procurement checklist. Does this model's training data include another vendor's model output; if so, which model; and in which document is that stated. And for the synthetic data we make ourselves, we are the ones who have to write those same three lines.
7.4A Record Cannot Be Made After the Fact
Pebblous does not belong on either side of a for-or-against argument about distillation. It belongs on the side that makes the generation path travel with the data, whether that path is synthesis or distillation. The place for it already exists. The hub has a base-model field and a dataset field, and there are teams that do write the teacher model's name and the sample count onto the card. The reason it sits blank is not that it is impossible but that it is not required. The scorecard was built in 2024, and the 300 documents Reuters counted two years later show why it was needed. Provenance cannot be restored later by forensic examination. It survives only if it is attached at the moment the data is made.
References
The figures in this article come from four strands. The verbatim quotes from the reporting were taken paragraph by paragraph from a syndicated copy of the Reuters article; the Reuters page itself sits behind a paywall and could not be opened. The paper's figures were checked against the body and tables of the 25 November 2024 version posted on arXiv, and there has been no revision since. For the government documents, the title, date and quoted sentences were confirmed on each original page. The two model cards were opened directly on 18 November 2026. Alibaba's April 2025 paper and the February 2026 paper from Shanghai Jiao Tong University and Meituan could not be identified, so they are carried only as Reuters described them.
Primary — Reporting
- 1.Baptista, E. & Chen, L. "Exclusive: China taps US AI models to train homegrown rivals as chip curbs bite." Reuters, 17 November 2026, filed from Shenzhen — the scope of more than 300 documents, the automation software and dedicated staff, the three cases, the July Chengdu meeting, the US official's assessment and the rebuttal from the Chinese developer side, China's January policy package, and the passage on asking Beijing's regulator where the boundary lies. Syndicated copy (BusinessWorld)
Primary — Papers
- 2.Huang, Z. and nine others. "O1 Replication Journey – Part 2: Surpassing O1-preview through Simple Distillation, Big Progress or Bitter Lesson?" arXiv: 2411.16489 v1, 25 November 2024 (accessed 18 November 2026). Shanghai Jiao Tong University, SII, New York University and GAIR. The headline experiment in §2.3 and §3; the bilingual data construction and Table 2 in §4; the Technical Transparency Index and Tables 3 and 4 in §5; §6, "The Bitter Lesson of Simple Distillation"; and the terms-of-use statement in the abstract footnote. Code at github.com/GAIR-NLP/O1-Journey
- 3.Mitchell, M. et al. "Model Cards for Model Reporting." FAT* 2019. arXiv: 1810.03993, DOI 10.1145/3287560.3287596 — the original paper proposing training data as a standard model card item. A recommendation only, with no binding force.
- 4.Stalnaker, T. and six others. "An Empirical Analysis of Machine Learning Model and Dataset Documentation, Supply Chain, and Licensing Challenges on Hugging Face." arXiv: 2502.04484 — a full audit of the hub. Its finding that the base-model and dataset fields are self-declared and therefore unevenly filled. Quantized models leaving the base-model field unfilled, and 1,416 cases of a model name sitting in the dataset slot.
Primary — Government and Regulation
- 5.NSA, CISA and FBI. "China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies." Joint Cybersecurity Advisory AA26-251A, 8 September 2026. cisa.gov — the list of six firms, "the critical core of it", billions of tokens since late 2024, and the gray market in API proxies ("transfer stations").
- 6.U.S. Department of Commerce, Bureau of Industry and Security. "Addressing Remote Access to Advanced Semiconductors." Interim Final Rule, RIN 0694-AK50, 16 November 2026 — the five countries covered, the aggregate threshold of 12 million TPP, and publication in the Federal Register set for 19 November 2026, taking effect on the date of publication.
- 7.Bessent, S. Post on X, 22 July 2026 — the warning of sanctions and Entity List designations under the condition of "covert, industrial-scale distillation attacks that cross the line into IP theft".
- 8.China's Ministry of Industry and Information Technology and seven other departments. 《"人工智能+制造"专项行动实施意见》, January 2026 — the policy package Reuters describes as containing guidance that supports distillation. We could not verify that passage against the Chinese original.
- 9.European Commission. "Explanatory Notice and Template for the Public Summary of Training Content for GPAI models" (EU AI Act, Article 53(1)(d)), finalized 24 July 2025, applicable from 2 August 2025 — requires data sources to be disclosed by category, and does not go as far as asking which teacher model synthetic data came from.
Primary — Model Cards
- 10.huggingface.co/Qwen/Qwen2.5-72B-Instruct (accessed 18 November 2026) — has a base-model field but no dataset field and no description of training-data sources.
- 11.huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B (accessed 18 November 2026) — gives the base-model field along with the teacher model's name and 800,000 reasoning samples in the body.
- 12.huggingface.co/docs/hub/model-cards — the specification for the base-model and dataset fields. Filling them in is voluntary.
Secondary
- 13.AI Times (Korea). "China industrializes US AI 'distillation': automated collection, dedicated staff." 17 November 2026 — Korean-language coverage of the same report. None of the secondary coverage we reviewed noted that the 5,000 figure comes from a side experiment, or pointed at the scorecard.
- 14.Independent review summary of the paper (unofficial service, retrieved 18 November 2026) — the observation that the Technical Transparency Index is self-scored and that the reported scores cannot be verified from the manuscript alone. We could not confirm it as a review issued by an institution, so the body refers to it only as "one independent review".
Related Pebblous Articles
- 15.Pebblous Blog. Chinese Labs Rerouted Their Own Users' Requests to Claude · You Can't Buy Sovereignty by Distilling It · Improving AI Without New Data · A Blacklisted Chinese University Buys Nvidia H200 Chips, and Rents More · When Tencent Rents AI Chips From Oracle, Where Does the Data Go? · China's National Data Group Supplies the Data That AI Learns From · Big Tech Left Half of the EU's Training-Data Summaries Blank · Perceptron's Open-Weight Factory Robot Model Leaves Its Data Sources Blank