Executive Summary

Editor's note. This piece follows Doing Science with a Tool You Cannot Reproduce. Where that piece dealt with models being closed, this one deals with what happens when they are open and the choice still splits along national lines. It runs on a different axis from America's Strongest Open-Weight Model — So Why Does It Trail China?, written around the same time: that report watched the race to build models, while this one measures the distribution of who uses them. Both belong to the Sovereign AI series.

The four numbers this report rests on

The first three give the current share of open-weight use and the size of the tilt behind it; the last one shows that the tilt is not about openness. The values look similar to one another, but each rests on a different denominator and a different condition, so the sections below walk through them one at a time.

Source: Dunivin, Z. O. (2026), Who Uses Open-Weight Models? China and the Shifting Geography of AI in Science, arXiv:2608.11090v1. All 2026 figures cover January through June only.

This study needs stating precisely. Who Uses Open-Weight Models? China and the Shifting Geography of AI in Science, posted to arXiv on 11 August 2026, is a v1 preprint written solely by Zackary Okun Dunivin of the Institute for Social Sciences at the University of Stuttgart. It has not been peer reviewed. What it does have is a reproduction repository with code, seeds, frozen classification decisions and checksums. Every number in this report was verified against that v1 text and its appendix.

Most summaries of the paper open with the phrase "an analysis of 21 million papers." Open the sample-construction table in the appendix, though, and 21 million turns out to be not the analytic sample but the size of the corpus that was screened. The difference matters. It inflates the scale by a factor of more than 130, and more importantly it points at the wrong contribution: what this study is good at is not how much it looked at but what it threw away.

1.1The funnel from 21 million to 157,000

The starting point is S2ORC, the Semantic Scholar full-text corpus. After cleaning up duplicate records left by a schema migration, 21,338,177 unique papers remain. Five filtering stages then fix the analytic sample. Watching how much falls away at each stage shows where the work went.

Source: arXiv:2608.11090v1, appendix Table S1.

The accurate phrasing is this: 21 million papers were screened to isolate 157,446. And the regression that actually computes national differences in choice runs on 48,129. It is the steepness of that funnel, not the size of the corpus, that makes the study credible.

1.2Used, or merely mentioned

The two stages in that table with the largest drop-off are the study's real contribution. A model name appearing in a paper does not mean the paper used the model. "We evaluated GPT-4 on three benchmarks" is use. "Prior work has evaluated GPT-4" reports someone else's use. "GPT-4 is multilingual" is a statement about the world. All three contain the same name, and only the first belongs on a map of scientific instruments.

Dunivin handles these judgements by fine-tuning SciBERT, a language model pretrained on scientific text. The identity classifier reaches 0.921 accuracy; the author-use classifier reaches 0.89 accuracy with an F1 of 0.93 on the use class. During training the target model name is masked behind a [TARGET_MODEL] placeholder, which forces the classifier to read context instead of memorising particular family names. Silver labels came from DeepSeek-V4-Flash, and against 150 human-labelled passages the pipeline reports F1 of 0.970 for identity and 0.943 for author use.

1.3Removing "we used ChatGPT to polish the text"

A third classifier targets the AI-use disclosure statements that now sit at the end of many papers. In a hand audit of 150 such statements, 71% of the model mentions were about manuscript editing, 21% about other parts of the workflow, 11% reported no use at all, and only 9% described use as a scientific instrument. A single mention can point at several purposes at once, so these categories are not mutually exclusive and the percentages sum to more than 100. This filter removed 11,080 papers, and 72.3% of the dropped paper-family pairs were ChatGPT.

1.4What "open" means in this paper

One definition needs nailing down before anything else. In this study, open-weight means only that a release lets you download the trained parameters or run them locally. Dunivin states in the text that the label does not imply that training data, training code, filtering procedures or licence terms meet any open-source standard. On the other side, proprietary means parameters are not released and access stays under the provider's control. Families that mix downloadable and controlled releases are not automatically counted as open-weight; they are resolved only by release-level rules or paper-level evidence.

The dictionary groups 159 model roots into 98 families. Access regime and developer nationality were coded first by an AI agent and then verified by the author, and one ordering decision matters for the design: access-regime coding happens after name recognition. Assumptions about which models are open are kept from contaminating the extraction step itself. As Section 5 revisits, this narrow definition carries straight through to the conclusion. The single fact that weights are downloadable lets the paper claim less than you might expect.

The most widely quoted number from this study is 44.0%. It is the share of the 16,125 single-family papers published between January and June 2026 that used an open-weight model. A single-family paper is one that used exactly one model family. It is less likely to be foundational work benchmarking several models against each other and more likely to be applied work picking one tool to solve a problem, which makes it the right place to read "what did the researcher choose." Note also that the corpus stops at June 2026, so every 2026 figure covers half a year. None of them should be read as an annual total.

The proxy is a rough one. The only evidence Dunivin offers for it is a difference in field composition. Among papers with an OpenAlex topic assignment, 55.9% of single-family papers have computer science as their primary topic, against 70.9% of multi-family papers. Papers that line several models up and measure them cluster harder inside computer science. He goes no further than saying the contrast "is consistent with the assumption that single-family papers are more likely to represent applied than foundational research." Read it as a signpost pointing toward applied work, not as a classification of it.

2.1Unfold it by family and a different picture appears

That 44.0% is the union of every open-weight family. Break the total apart and you can see what happened inside it. Below is each family's share of single-family papers in the first half of 2026.

Inside that 44.0%, Qwen alone accounts for 22.0%. Among papers that chose open weights at all, 49.9% chose Qwen, and widening to Chinese-built models of any kind gives 60.5%. In the discussion Dunivin rounds this to "about 61 percent, of which 83 percent is Qwen." Calling the last three years a shift toward open weights is therefore only half right. Closer to what was measured is a shift toward Qwen.

2.2GPT's lead and the open-weight rise are both true

In the same table, the GPT family sits at 44.2%. That is close enough to the 44.0% open-weight total to invite confusion, but the two numbers are mutually exclusive groups sharing a denominator. A single-family paper uses one family by definition, so papers that used GPT and papers that used open weights do not overlap. Applied research in the first half of 2026, in other words, split almost evenly between the GPT camp and the open-weight camp. The near-identical values are not a coincidence; the split really is about half and half.

GPT has not been displaced, though. In 2023, the GPT family held 80.9% of single-family papers. It has halved in three years and is still the largest single family, and in multi-family papers it sits at 64.6%, effectively tied with Qwen at 62.1%. Foundational work running benchmarks puts GPT and Qwen side by side; applied work picking one tool divides between them. Both pictures hold at once.

2.3Llama and Mistral lost ground with their weights still open

Saying that Chinese families absorbed the open-weight gains is another way of saying Western open weights lost ground. They did. Putting single-family and multi-family shares next to each other shows how far.

Denominators: 16,125 single-family papers and 29,000 multi-family papers. A multi-family paper uses several families, so that column sums past 100. Source: arXiv:2608.11090v1, Table 1.

Llama peaked at 12.0% of single-family papers in 2025 and has come down to 8.6%. In multi-family papers it slid from 58.7% in 2024 to 41.3%. Mistral fell further: from 20.6% to 12.1% in multi-family papers, and down to 0.8% in single-family ones. The floor-level shares of gpt-oss (0.7%) and Kimi (0.1%), by contrast, are lag rather than decline. Both arrived in the second half of 2025 or later, and given publication lag there has been no time for them to show up. Dunivin writes that he expects both to rise.

The regression runs on the 48,129 single-family papers where the first author's and corresponding author's institutional country could be identified. Within that sample, 28.9% of papers come from Chinese institutions, 11.5% have Chinese-origin author names but institutions outside China, and 59.6% have no China connection at all. The outcome variable is whether the paper chose open weights, and every model absorbs timing effects with a natural spline on publication date at four degrees of freedom.

3.1Why 2.87 fell to 2.13

Dunivin runs three models and reports all three. The order is the point.

The three stack variables one at a time. Model 1 puts only Chinese-institution affiliation on top of the publication-date spline. Model 2 adds random intercepts for 83 subfields, absorbing "where did this get published." Model 3 splits the China connection into three non-overlapping groups, pulling papers with Chinese-origin names but non-Chinese institutions out of the comparison baseline. The small rise from 2.13 to 2.23 in Model 3 is therefore better read as the reference group narrowing to papers with no China connection at all than as the effect getting stronger.

n = 48,129. Brackets show 95% confidence intervals. Source: arXiv:2608.11090v1, Table 2.

Without subfield controls, Model 1 puts the odds ratio at 2.87. Add random intercepts for 83 subfields and it drops to 2.13. The meaning of that drop is unambiguous: researchers at Chinese institutions were publishing disproportionately in the subfields that favour open weights. Roughly a quarter of the raw gap came from field composition rather than nationality. That Dunivin reports all three models instead of hiding the drop says something about the paper's intellectual honesty.

With the name variable added, Model 3 gives a final odds ratio of 2.23, confidence interval 2.12 to 2.34. Papers with Chinese-origin names at institutions outside China sit at 1.23 — real, but far smaller. This is association, not causation. Nothing in the paper supports writing that nationality determined model choice. Dunivin himself locates the decision not in nationality but in platforms, markets, and sociocultural and geopolitical context.

3.2Who produced the 23.1-point rise

The growth is more interesting than the level. In the regression sample, open-weight adoption climbed from 12.8% in 2023 to 35.8% in the first half of 2026. Decomposing that 23.1-point rise by group splits it as follows.

The decomposition has two axes. One is the rise in adoption rate within each group; the other is the change in each group's share of the sample. Overall, within-group increases account for 23.9 points and compositional change for −0.8 points. Practically all of the growth came from the same groups choosing open weights more often, not from any group publishing more papers. The bars below show each group's net contribution once both axes are combined.

Papers from Chinese institutions contributed 10.2 points. Divide that by the total rise of 23.1 points and you get 44.0% — the "44 percent of the increase" that the abstract reports. This is also where readers most often slip, because the value 44 turns up in three different places.

Add the 3.7 points from papers with Chinese-origin names outside China and the China-linked contribution reaches 13.9 points, or 60.2% of the rise. That group makes up 38.9% of the sample. Two-fifths of the sample produced three-fifths of the growth. Note that the 35.8% denominator behind this decomposition is a different quantity from the 44.0% in the previous section: the regression sample keeps only papers with an identified country, and as the next subsection shows, affiliation data is missing more often for open-weight papers, which biases the adoption rate downward.

3.3The author puts the openness story to the test and rejects it

At this point one reading suggests itself: that Chinese researchers care more about openness. Dunivin turns that reading into something testable and puts it to the data. If openness itself drove the pattern, papers from Chinese institutions should also reach for non-Chinese open weights more often. To check, he fits a multinomial model with three outcomes: proprietary, non-Chinese open weight, Chinese open weight. Adjusted probabilities, standardised over the subfield distribution, come out as follows.

For papers from Chinese institutions, the adjusted probability of choosing a Chinese open-weight model rose from 2.2% in 2023 to 37.1% in the first half of 2026. Over the same period, papers with no China connection went from 0.3% to 9.2%. The gap widened from 1.9 points to 27.9 points. This 9.2% shares a value with the 9.2 points in the decomposition above but is a different quantity: that one is a contribution to growth, this one is a probability of choice. So far, no surprises.

3.4Affiliation data goes missing most in open-weight papers

The most interesting limitation in this study is not on the model side but the metadata side. Author country comes from OpenAlex, and that field is missing at different rates depending on what is being observed — most of all exactly where the finding lives.

Here is how often the first author's and corresponding author's country can be resolved, by type of paper.

Country resolves for 77.1% of all papers, rises to 82.1% among papers using proprietary models, and falls to 65.9% among papers using open weights. For papers using Qwen it is 49.6%, and for Qwen papers published in 2026, 39.5%. The same pattern shows up year by year. In 2023, coverage for papers with Chinese-origin names was 92.9%, slightly above the 91.1% for everything else; by 2026 that has flipped to 62.8% against 77.2%.

Put plainly, this study's instrument is blindest on the papers that are newest, most open-weight, and most Chinese. Dunivin does not hide this. He states that the data is not missing completely at random and warns that the negative compositional term in the decomposition (−1.9 points) should not be read as evidence that Chinese-affiliated papers shrank as a share of the population. He also runs a sensitivity analysis: filling the gaps by allowing any author's last-listed affiliation moves the odds ratio from 2.23 to 2.19, and leaves the 10.2-point contribution unchanged. The honest summary is that the conclusion holds while the precision has a ceiling.

Uneven coverage in scholarly metadata is not unique to this paper. Pebblous has hit the same seam while working with the 300 million papers OpenAlex now gives away. The difference here is that this author measured where his own instrument goes dark and published the numbers.

Before asking what a researcher chose, it helps to ask what there was to choose from. The three years in which open-weight use grew were also three years in which usable open-weight models poured out. They poured out unevenly.

4.1Weight releases clustered on one side

Below are the major public releases that overlap the study's observation window. Putting the Chinese and Western sides on one axis makes the difference in density visible.

The densest stretch is the eight months from September 2024 to April 2025. Qwen2.5, DeepSeek-V3, DeepSeek-R1 and Qwen3 landed one after another, with Qwen3 under Apache 2.0 and DeepSeek-R1 under MIT. Over the same three years, the comparable Western weight releases number three: Gemma, Llama 4 and gpt-oss. The H1 2026 distribution from Section 2 reflects only the earlier part of this timeline, which is also why gpt-oss and Kimi are still sitting on the floor.

Each entry was re-checked against official announcements as of 13 August 2026.

One caution belongs here. The Qwen catalogue is mixed. Apache 2.0 coexists with the source-available Qwen License and the non-commercial Qwen Research License, and the top-end models are API-only. This is exactly why the paper's rule — never assign a mixed family to open-weight automatically, resolve it release by release — was necessary. The market has already produced a situation in which a family name tells you nothing definitive about openness.

4.2Opening the weights was not enough

Llama's weights were open. So were Mistral's. And as Section 2 showed, both lost ground in the literature. Dunivin takes the contrast head-on in the discussion.

"The contrasting trajectories of Qwen and DeepSeek on one hand and Llama and Mistral on the other suggest that releasing weights translates into scientific use only when attached to a competitive technical and economic package. Open release is itself an industrial strategy."

Dunivin (2026), Discussion

The diagnosis continues. Researchers do not choose openness or closedness in the abstract. They choose among systems that bundle capability, price, language performance and institutional factors. Which means the growth of open weights may reflect competition among model producers rather than any change in scientists' methodological values.

4.3The terms of choice differed by family

Licence terms differ substantially across families. The Llama Community License has carried a set of restrictions since Llama 2. Companies exceeding 700 million monthly active users in the preceding month must request a separate grant from Meta, which Meta may refuse at its discretion, and the acceptable use policy excludes individuals and entities domiciled in the EU. On top of that come a "Built with Llama" attribution requirement, a rule that derivative model names begin with Llama, a clause barring the use of outputs to improve competing foundation models, and a patent-retaliation clause. The OSI does not recognise Llama as open source. Qwen3 and gpt-oss, by contrast, are Apache 2.0, and DeepSeek-R1 is MIT.

None of this establishes that the asymmetry caused Llama's decline. The study does not include licence as a variable, and what can be confirmed here stops at the observation that the terms of what a researcher could choose differed by family. The view of weight releases as a question of data sovereignty is treated separately in Open-Weight Kimi K3 Makes Data Sovereignty a Hardware Question.

The case for using open weights rested on reproducibility from the start. Hold the weights in your hand, the expectation goes, and you can get the same result again. Dunivin examines in the discussion whether that expectation actually holds.

"Downloadable parameters enable local execution, preservation, adaptation, and reduced exposure to unilateral API changes, but they do not make models transparent in the ordinary scientific sense."

Dunivin (2026), Discussion

He gives three reasons. Training data and development procedures often remain undisclosed; model behaviour is extremely hard to explain; and the technical possibility of reproduction neither creates an incentive to reproduce nor makes reproduction happen. The last condition stings the most. That even published code fails to run about half the time is something we already documented in Released Paper Code: AI Agents Ran It, and Half of It Failed.

5.1The same weights, a different result

The sentence "we used an open-weight model" will not get you the same result three years from now. Too much between the weight file and the output is left free to vary. These are the items a methods section has to carry for even minimal reproduction to be possible.

• Checkpoint identifier and revision. Models are quietly updated under the same name more often than you would like. Only a repository revision hash pins it down.
• Quantisation. A run at 4-bit and a run at 16-bit do not agree, even from identical weights.
• Serving stack and version. Which inference engine, at which version, determines the kernels and the numerics.
• Decoding parameters. Temperature, top-p, repetition penalty, max tokens. Writing down "temperature 0" alone raises reproducibility substantially.
• Random seeds. Needed for both sampling and data splits.
• The prompt itself. Not a summary of it — the exact string that went into the model.

None of this depends on whether the weights are open. It is equally necessary for a proprietary API call and for open weights run locally. What open weights give you is the possibility of keeping this record, not the record itself.

5.2The paper is its own worked example

Pleasingly, the study demonstrates the answer to the problem it identifies. The public repository holds the extraction dictionary, a text-free relational dataset of 3,619,386 candidate occurrences, the LLM prompts and structured annotations, hand-coded evaluation data, fixed train/validation splits, and software specifications alongside random seeds. Three seeds are stated explicitly: 13 for SciBERT training and fold assignment, 20260710 for the 500 draws behind the adjusted probabilities, and 20260630 for the 2,000-replicate decomposition bootstrap.

Two details deserve more attention. First, a lightweight workflow starts from the frozen classification decisions, regenerates the statistical results, and automatically checks the resulting sample sizes, tables, figures and model outputs against the reported values. Second, the three fine-tuned SciBERT checkpoints ship as versioned release assets with checksums and model cards, together with the exact revision of the base SciBERT model. Even under copyright constraints that forbid redistributing paper full texts, document identifiers and in-document positions were retained so that the measurement process itself can be re-examined.

5.3Even when it is open, the problem remains

In Doing Science with a Tool You Cannot Reproduce, Pebblous dealt with verification breaking down because the model was closed. The diagnosis then was that closure was the problem. What this study shows is the next phase. Open weights only convert into reproduction when a record follows, and which open weights get chosen is settled by ecosystems rather than by openness.

Another study asks a related question from a different angle. The Nature paper by Evans and colleagues, cited as reference 8 in this preprint, drew from 41.3 million papers the conclusion that AI tools have made individual scientists stronger and science itself narrower. One counted how AI changed the breadth of science; the other counted where that AI came from. Lay them on top of each other and a further question appears. If the topics of science are narrowing while the instrument that measured the narrowing is itself narrowing to one ecosystem, what then?

Korea's absence from this map is by design rather than accident. The study's axis is the United States and China, and the name-based classifier includes a rule that suppresses Korean surnames to reduce false positives — a guard against names like Kim, Lee and Park being read as Chinese-origin. Dunivin is explicit that name classification is a noisy indicator and states that it does not identify nationality, citizenship, ethnicity, self-identification or residence. Korean researchers, in other words, are lumped into the not-China side of this instrument without a coordinate of their own.

6.1Which models Korean papers mention

Applying the study's method to Korea would require the full-text corpus and both use-versus-mention classifiers, which public APIs cannot reach. A much cruder substitute is available: the share of a country's AI papers in which a given model name appears anywhere in the record. Queried on 13 August 2026, OpenAlex returns the following.

Direct query against the OpenAlex Works API, 2026-08-13. Population: papers from 2023–2026 carrying an AI concept tag — 72,425 for Korea and 875,729 for China. GLM and Solar were excluded because their strings collide with "generalized linear model" and with solar-energy research.

Korean AI papers mention Llama at roughly twice China's rate, and their rates for Chinese-built models are comparable or lower. That can be read as a directional signal that Korea does not show the home-model concentration the preprint found in China. Three things stop it from being pressed harder. First, this is mention, not use — precisely the distinction that the preprint needed two classifiers to draw. Second, affiliation is attributed differently: the preprint looks only at first and corresponding authors, while the OpenAlex query matches the institutions of every author. Third, domestic-model mention rates fall below 0.05%, so the basis for comparison is thin. There was also no way to cross-validate. It is a number that can only be used with the caveat that it is a single-source, coarse indicator.

6.2Present in output, absent in compute

On other measures Korea does appear on the map. According to Chapter 1 of Stanford HAI's AI Index 2026, notable AI models in 2025 numbered 59 from the United States, 35 from China and 8 from Korea, followed by Canada, France, Hong Kong, Singapore and the United Kingdom at one each. In the same report, AI patents per 100,000 people in 2024 put Korea first in the world at 14.31, ahead of Luxembourg at 12.25, China at 6.95 and the United States at 4.68.

Yet in Epoch AI's tally of national shares of AI supercomputer performance as of May 2025, the United States holds 74.5% and China 14.1%, and Korea does not appear. Present in the output and density measures, absent from the compute-share table. If open-weight choice is entangled with resource constraints, that contrast is hard to wave away. It is inference, though, not measurement, and worth marking as such.

6.3The most capable Korean model carries a non-commercial licence

Apply Section 4's "terms of choice" to domestic models and another texture appears. Below are four major Korean models with released weights, arranged by their commercial-use conditions. Looking at the threshold between downloading something for research and putting the result into a product shows how differently the same word — "open" — can be scoped.

Published licence terms as of 13 August 2026. Terms for EXAONE successors released after that date need separate verification.

That contrast repeats, at national scale, the structure Section 4 described. Opening a set of weights and getting those weights widely used are different projects. The strategic version of this problem is collected in the Sovereign AI series.

Pebblous works on data in the state it is in before it reaches a model. Which is why the question this study raises does not sound like someone else's problem. It puts one more question in front of the work of preparing data. What was that data read with?

7.1Model choice belongs in the lineage, not the experimental setup

Inside a data pipeline, an LLM is no longer a tool but a component. It labels, classifies, adjudicates quality, generates synthetic data. And yet the provenance of the dataset gets recorded meticulously while the provenance of the model that processed it ends at "we used GPT-4." For data lineage to mean anything, the coordinates of every model that touched the data have to be preserved at the same precision. Move model choice out of the experimental-settings box and into the lineage. The checklist in Section 5 is the minimum unit.

7.2Record the holes instead of filling them

The missingness structure in Section 3 is a textbook case for data quality work. Because missing affiliation clusters in recent papers, open-weight papers and papers with Chinese-origin names, the instrument's field of view changed with the object being observed. Accept the regression coefficients without knowing that coverage is uneven and those coefficients cannot tell you what they failed to see. Dunivin measured the skew and published the numbers, then showed with a sensitivity analysis how far the conclusion moves. That is exactly the job of data quality diagnosis: the holes and the direction they lean are part of the result, not something to be papered over.

7.3"We use open weights" is not an answer to an audit

Once regulators and auditors start asking for model lineage, this becomes a documentation problem. Faced with an AI bill of materials requirement or the documentation obligations in the EU AI Act, "we use open-weight models" is not a response. What is being asked for is which weights, at which version, in which runtime, and that answer cannot be reconstructed after the fact. Not written down at run time means gone for good. Openness makes such a record possible; it does not substitute for one.

7.4The sovereignty debate has only counted who builds

Sovereign AI has been argued mostly as a question of production capacity. This study adds a layer on top. Who uses them has already diverged, and diverged in a measurable form. If the indicator of sovereignty is not how many domestic models a country has built but what its researchers actually run as instruments, then Korea does not yet have that indicator. That is the gap Section 6 leaves. When we talk about data sovereignty, we would rather the evidence be a measured distribution of use than a policy declaration.

Every figure in this piece was verified directly against the primary source, arXiv:2608.11090v1, including its appendix. What follows groups that primary source, the scholarly literature the preprint cites for its methods and argument, the policy and statistical material used in Section 6, and adjacent Pebblous pieces on the same subject.

Primary source

• 1.Dunivin, Z. O. (2026). Who Uses Open-Weight Models? China and the Shifting Geography of AI in Science. arXiv:2608.11090v1 [cs.CY], 2026-08-11. Institute for Social Sciences, University of Stuttgart.
• 2.Dunivin, Z. O. (2026). who-uses-open-weights. Reproduction repository (model dictionary, candidate-occurrence data, prompts, seeds, classifier checkpoints and checksums).

Methods and data infrastructure

• 3.Lo, K., Wang, L. L., Neumann, M., Kinney, R., & Weld, D. S. (2020). S2ORC: The Semantic Scholar Open Research Corpus. ACL 2020, 4969–4983.
• 4.Priem, J., Piwowar, H., & Orr, R. (2022). OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. arXiv:2205.01833.
• 5.Beltagy, I., Lo, K., & Cohan, A. (2019). SciBERT: A pretrained language model for scientific text. EMNLP-IJCNLP 2019, 3615–3620.

Openness and reproducibility

• 6.Liesenfeld, A., & Dingemanse, M. (2024). Rethinking open source generative AI: Open-washing and the EU AI Act. ACM FAccT 2024, 1774–1787.
• 7.Kapoor, S., et al. (2024). On the societal impact of open foundation models. ICML 2024.
• 8.Hao, Q., Xu, F., Li, Y., & Evans, J. A. (2026). Artificial intelligence tools expand scientists' impact but contract science's focus. Nature, 649, 1237–1243. Cited as reference 8 in the preprint.
• 9.Bommasani, R., et al. (2023). The Foundation Model Transparency Index. arXiv:2310.12941.
• 10.Open Source Initiative. (2024). The Open Source AI Definition 1.0.
• 11.National Academies of Sciences, Engineering, and Medicine. (2019). Reproducibility and Replicability in Science. The National Academies Press.

Policy and statistics

• 12.Meinhardt, C., et al. (2025). Beyond DeepSeek: China's diverse open-weight AI ecosystem and its policy implications. Stanford HAI / DigiChina.
• 13.Chahal, H., et al. (2026). Open models, soft power, and the spectrum of U.S.–China artificial intelligence competition. RAND, PEA4686-1.
• 14.OECD. (2025). AI openness: A primer for policymakers. OECD AI Papers No. 39.
• 15.Stanford HAI. (2026). AI Index Report 2026, Chapter 1: Research and Development. Notable AI models by country (2025); AI patents per 100,000 people (2024).
• 16.Epoch AI. AI supercomputers performance share by country. As of May 2025.
• 17.Direct query against the OpenAlex Works API (2026-08-13). Counts of model-name mentions in AI papers by country. Query conditions are noted in Section 6.1.

Adjacent Pebblous pieces

• 18.Pebblous. Doing Science with a Tool You Cannot Reproduce. The predecessor to this piece.
• 19.Pebblous. AI Made Individual Scientists Stronger, Made Science Itself Narrower. On the Nature study the preprint cites.
• 20.Pebblous. 300 Million Papers, Now Free. Should You Swallow Them Unchecked? On coverage problems in scholarly metadata.
• 21.Pebblous. America's Strongest Open-Weight Model — So Why Does It Trail China? The axis of building models.
• 22.Pebblous. Open-Weight Kimi K3 Makes Data Sovereignty a Hardware Question.
• 23.Pebblous. Released Paper Code: AI Agents Ran It, and Half of It Failed.
• 24.Pebblous. Sovereign AI series hub.