Between Regulation and Accessibility: How Chinese University Students Navigate Global and Domestic Generative AI

Fei Cheng; Ming Li; Qin Xie

REVIEW 3 major objections 5 minor 10 references

Between Regulation and Accessibility: How Chinese University Students Navigate Global and Domestic Generative AI

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Chinese undergraduates treat global and domestic generative AI as complementary tools, routing around access barriers with VPNs and shared accounts while relying on domestic models for Chinese-language and culturally embedded tasks.

desk verdict Vivid interviews, thin sample: a useful exploratory study that overgeneralizes and includes a non-reproducible benchmark figure. read the letter →

arxiv 2506.14377 v1 pith:F3J4ODGJ submitted 2025-06-17 cs.CY

Qin Xie , Ming Li , Fei Cheng This is my paper

classification cs.CY

keywords generativeAIhighereducationChineseuniversitystudentsdigitaldivideUTAUT2qualitativeinterviewsChatGPTdomestic

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper uses semi-structured interviews with 15 undergraduates at one elite Chinese university to establish that students treat global and domestic generative AI as complementary tools rather than direct substitutes. The central claim is that engagement is shaped by accessibility, language proficiency, and cultural relevance: students reach ChatGPT and other global tools through VPNs, shared accounts, and paid intermediaries, while turning to domestic models such as Kimi, Doubao, Wenxin Yiyan, and Qwen for Chinese-language and culturally aligned coursework. The paper reads this behavior through the technology-acceptance lens of UTAUT2 and notes that domestic tools, though easily accessible and strong in Chinese, are limited by content filtering. If the claim holds, AI adoption in Chinese higher education cannot be read from model rankings alone; it is a product of regulatory, economic, linguistic, and social forces that universities and policymakers must address together.

What carries the argument

The argument is carried by a strategic task-allocation mechanism observed in interviews: students choose a tool based on the language, cultural context, and perceived strengths of the model for each assignment, rather than loyalty to one platform. This mechanism is read through UTAUT2 (Unified Theory of Acceptance and Use of Technology 2), whose seven constructs—performance expectancy, effort expectancy, social influence, facilitating conditions, hedonic motivation, price value, and habit—supply the interview themes and coding categories. The study also builds a five-dimension benchmark comparison (English, Chinese, math, reasoning, coding) from public evaluations to frame the perceived versus measured gaps between global and domestic models.

What would settle it

A representative survey or observational log study of Chinese undergraduates that found most students never use VPNs, shared accounts, or intermediaries to reach global AI, and that showed no systematic relationship between assignment language or cultural content and tool choice, would contradict the paper's central claim.

Watch

Extended reading notes

Core claim

The paper's central discovery is a strategic task-allocation pattern: Chinese university students do not choose between global and domestic generative AI on raw capability alone; they allocate tasks based on access, language, and cultural fit. Students reach ChatGPT, Claude, and Gemini through VPNs, shared accounts, and paid intermediaries, accepting instability and privacy risk for tasks where they perceive global models as stronger, particularly English, STEM, and coding work. They turn to domestic tools such as Kimi, Doubao, Wenxin Yiyan, and Qwen for Chinese-language interaction, traditional Chinese culture, and coursework on socialism with Chinese characteristics, where domestic models are seen as more fluent and context-appropriate, though content filters block politically sensitive queries. Under the UTAUT2 lens, accessibility, effort, price, social influence, and hedonic motivation jointly explain this split, and a supplementary five-dimension benchmark comparison suggests global and domestic models are close on English, math, reasoning, and coding, while domestic models lead on Chinese.

Load-bearing premise

The load-bearing assumption is that 15 self-recruited undergraduates at one unnamed elite university in a first-tier city, recruited through the authors' personal WeChat networks, stand in for Chinese university students generally.

Editorial extensions

If this is right

Chinese students' real AI usage is broader than official domestic-platform statistics suggest, because VPNs, shared accounts, and intermediaries route a share of use to global tools.
Domestic AI gains adoption not primarily through technical superiority but through effortless access, Chinese-language fluency, and cultural and political alignment; content filtering then constrains what students can explore.
Students' willingness to pay diverges by tool: ChatGPT subscriptions are seen as worth about $20 per month, while free domestic tiers satisfy most local needs, and payment friction itself deters global subscriptions.
Coursework language and cultural content predict tool choice: English, STEM, and Western-context assignments push students to global AI, while Chinese, humanities, and political-cultural assignments push them to domestic AI.
Students see generative AI as time-saving but not grade-raising, and their job anxiety is field-specific, with each discipline viewing the other as more exposed to automation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the workarounds documented here are common beyond this sample, usage data collected only from domestic platforms will undercount AI-assisted learning in China and mislead educators about which models students actually rely on.
The task-language allocation pattern implies a testable signature: in a representative sample, the probability of choosing ChatGPT should increase with the Englishness and Western-context content of an assignment, and choosing domestic tools should increase with Chinese cultural or political content; vignette experiments could confirm or falsify this.
Students' perception that domestic models are weaker in English and stronger in Chinese may persist even where public benchmarks show the gap narrowing, meaning purely technical improvements in domestic models will not shift usage until user beliefs change.
Because all interviewees attend one elite first-tier city university, the same mechanisms may not hold for students at less selective or rural institutions, where VPN payment channels, foreign bank cards, and peer account-sharing networks are less available; a comparative multi-site study is the natural next test.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Plain English: this is a small qualitative study of 15 Chinese undergraduates at one elite university, interviewed about how they use global versus domestic generative AI. The interview material is the best part. You get concrete, memorable details: students sharing ChatGPT accounts with ex-boyfriends, buying access through Taobao, switching languages by task, and perceiving domestic models as better for Chinese culture and politics. That is new and worth having on record. The authors use UTAUT2 as a framing device, describe a credible coding procedure, give named quotes, and state in the conclusion that this is a single-site study. Credit where it is due: the empirical account is plausible and the authors are honest about the sample's limits.

The soft spots are real and match the stress-test note. The sample is 15 self-recruited students from one unnamed elite university in a first-tier city, recruited through the authors' WeChat contacts. That is acceptable for an exploratory qualitative study, but the abstract and discussion repeatedly generalize to 'Chinese university students' and even 'students in similar contexts worldwide.' Those claims outrun the data. The authors acknowledge the single-site limitation only in the final paragraph, which is too little, too late. They need to qualify every generalizing sentence or bring in multi-site data.

Second, Figure 1 is a genuine problem. It presents a benchmark comparison of three global and three domestic models across five dimensions, but there are no raw scores, no error bars, no model version dates, and no source table. The text says 'no significant performance gap' as if it were a statistical finding, but it is just a hand-drawn radar chart. This figure should be cut or replaced with a proper table of cited scores. Right now it is not reproducible.

Minor: the interview protocol is built around UTAUT2 constructs, so the deductive coding somewhat invites the constructs that later appear as findings. That is a mild circularity concern, not a fatal one. The reference list also has a few small errors, but nothing that changes the analysis.

Verdict: this is a useful, plausible exploratory account of an underexplored population. It deserves serious peer review, but the referee should demand tighter generalizing language and a fix or removal of Figure 1. I would send it to a qualitative-research-friendly education or information-science venue. I would not cite it as evidence about Chinese students generally, but I might cite it as an example of how students navigate the dual-AI environment.

Referee Report

3 major / 5 minor

Summary. This qualitative study investigates how Chinese undergraduate students engage with global and domestic generative AI in their learning, using semi-structured interviews with 15 self-recruited students at one elite Chinese university. The authors analyze the transcripts through UTAUT2-informed thematic coding and report that students' engagement is shaped by accessibility constraints and workarounds (VPNs, shared accounts), language proficiency and task type, cultural/ideological fit, and attitudes toward AI as human-like or machine-like. The paper also includes a benchmark-based radar chart comparing global and domestic models across English, Chinese, math, reasoning, and coding, arguing that the performance gap has narrowed. The discussion connects the findings to digital divides, non-English AI development, cultural awareness in AI, and human-centered design.

Significance. If the findings were robustly generalizable, the paper would make a useful contribution to the under-studied area of generative AI adoption in non-Western higher education, particularly by documenting task-based division of labor between global and domestic tools and the socio-technical nature of access workarounds. The topic is timely, and the qualitative material is often vivid and analytically suggestive. The paper is also transparent about its interview procedure, coding approach, and participant recruitment, and it includes named participant quotes that illustrate the main themes. Its central limitation is external validity: the sample is small, self-selected, single-site, and recruited through the authors' personal networks, so the abstract's population-level claims outrun the evidence. The quantitative benchmark comparison, which is used to contextualize the qualitative findings, is not currently documented well enough to be verified.

major comments (3)

[Sampling (p. 11); Discussion (p. 25); Conclusion (p. 28)] The abstract and introduction state as a general finding that engagement with global and domestic generative AI is shaped by accessibility, language proficiency, and cultural relevance, and the Discussion extends this to 'students in similar contexts worldwide.' This generalization rests on 15 self-recruited undergraduates from one unnamed elite university in a first-tier city, recruited through the authors' personal WeChat networks, with two computer-science participants explicitly reporting 'legal access' through AI-industry internships and U.S. exchange study. Such a sample is likely to overrepresent students with VPN access, overseas exposure, and strong English skills, so the documented behaviors cannot support population-level claims about Chinese university students or comparable students elsewhere. The authors' own concession in the Conclusion that this is a single-site study requiring comparative mixed-methods work should be reflected in the abstract and Discussion; please reframe the central claims as exploratory and site-specific, or provide a clearly bounded transferability argument with explicit sample-composition analysis.
[Figure 1 and 'Technical Comparison' (pp. 6–8)] The radar chart is the only quantitative support for the statements that domestic models 'consistently outperform' global models in Chinese C-Eval and that 'there is no significant performance gap between global and domestic models at present.' The figure gives no numeric scores, no axis scale, no benchmark version or date per data point, and no per-source citation; the note merely says scores were collected from publicly available sources. Because the five dimensions use different benchmarks (MMLU, C-Eval, MATH, reasoning, coding), the radar normalization is opaque and the 'no significant gap' conclusion is not verifiable. Please add a supplementary table reporting each model's score on each benchmark, with the exact benchmark version, release date, and source, and state how the five dimensions were normalized to a common scale.
[Results (pp. 12–24)] The Results section reports several quantities (e.g., 'all 15 participants identified Chinese as their first language,' 'the majority (10 participants) used both Chinese and English,' 'approximately 11 individuals lack confidence in score improvement') without a participant summary table or a coding-frequency appendix, so these counts cannot be audited. For a qualitative study that makes numeric claims, the authors should provide a pseudonymized participant table (gender, major, year, global/domestic tools used, paid/free status) and, where counts are reported, a theme-by-participant matrix that shows how the counts were derived from the coded data.

minor comments (5)

[Literature Review (p. 9)] The text says UTAUT integrates 'three key determinants' but immediately lists four: Performance Expectancy, Effort Expectancy, Social Influence, and Facilitating Conditions; this should be corrected and the sentence reconciled with the actual UTAUT structure.
[Conclusion (p. 27)] The Conclusion states that the study reveals 'four key shaping factors' but then enumerates only three: access limitations, linguistic capabilities, and cultural familiarity; the count should be changed to three, or a fourth factor should be explicitly identified.
[Figure 1 (p. 7)] The figure legend distinguishes solid and dash-dot lines but does not identify individual models; since the text refers to GPT-4o, Claude-3.5, Gemini-Exp, DeepSeek-v3, Doubao-1.5-Pro, and Qwen2.5, the chart should make the per-model lines identifiable.
[Results (p. 17)] The phrase 'As Maoxin shared' appears twice in the same paragraph; the duplicated attribution should be removed.
[Results (p. 22)] One participant is quoted under the pseudonym 'Xunfei,' which is also the name of a well-known Chinese AI company (iFlytek); if this is a pseudonym, consider choosing a different one to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: interview-based findings stand on reported data, not on fitted parameters or self-citation chains.

full rationale

The paper contains no equations, fitted parameters, or derived quantities; its central claims are qualitative inferences from 15 semi-structured interviews. The UTAUT2 framework is used as an interpretive lens and its constructs are invoked to organize emergent themes, but the interview excerpts (e.g., account sharing after a breakup, payment barriers, Chinese painting aesthetics, ideological-course assignments) supply independent empirical content that is not entailed by the framework. No prediction is computed from a fitted input, and no result is imported from the authors' prior work as a load-bearing premise; self-citations (Xie et al., 2024; Li et al., 2025a, 2024, 2025b) appear only in the literature review and are not used to justify the empirical findings. The benchmark comparison is drawn from public sources and is ancillary to the interview analysis. The explicit single-site limitation is a generalizability caveat, not a circularity. Therefore no circular step is identifiable.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper makes no mathematical claims, so there are no fitted free parameters and no invented entities. The load-bearing assumptions are qualitative: UTAUT2 applies to this setting, interviewees report accurately, the regulatory picture is accurate, and the public benchmark scores used in Figure 1 are trustworthy. All are domain assumptions rather than equations, and the benchmark assumption is the weakest because no raw data are provided.

assumptions (4)

domain assumption UTAUT2 constructs (Performance Expectancy, Effort Expectancy, Social Influence, Facilitating Conditions, Hedonic Motivation, Price Value, Habit) validly explain generative AI adoption in this Chinese undergraduate setting.
The Theoretical Framework section adopts UTAUT2, and the interview protocol and coding are structured around it; if the model does not transfer, the results' framing does not hold.
domain assumption Interview self-reports accurately reflect actual AI usage.
All findings rely on what 15 participants said in one-hour Zoom interviews; there is no observational or log data to verify usage frequency, payment, or access claims.
domain assumption Publicly sourced benchmark scores in Figure 1 are accurate, comparable, and current.
Figure 1's note says scores were collected from public sources but provides no raw values or links; the claim that global and domestic models show no significant performance gap depends on this assumption.
domain assumption China's restrictions on global AI services and the availability of domestic alternatives are accurately characterized.
The Introduction and Technical Overview rely on secondary sources, including the Cyberspace Administration of China and press reports, for the regulatory environment that motivates the study.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Between Regulation and Accessibility: How Chinese University Students Navigate Global and Domestic Generative AI." pith.science (2026). https://pith.science/paper/F3J4ODGJ

@misc{pith2026250614377,
  author       = {Pith},
  title        = {Pith review of: Between Regulation and Accessibility: How Chinese University Students Navigate Global and Domestic Generative AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F3J4ODGJ}},
  note         = {Machine review of arXiv:2506.14377}
}

read the original abstract

Despite the rapid proliferation of generative AI in higher education, students in China face significant barriers in accessing global tools like ChatGPT due to regulations and constraints. Grounded in the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) model, this study employs qualitative interviews to investigate how Chinese university students interact with both global and domestic generative AIs in the learning process. Findings reveal that engagement is shaped by accessibility, language proficiency, and cultural relevance. Students often employ workarounds (e.g., VPNs) to access global generative AIs, raising ethical and privacy concerns. Domestic generative AIs, while offering language and cultural advantages, are limited by content filtering and output constraints. This research contributes to understanding generative AI adoption in non-Western contexts by highlighting the complex interplay of political, linguistic, and cultural factors. It advocates for human-centered, multilingual, domestic context-sensitive AI integration to ensure equitable and inclusive digital learning environments.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

10 extracted references · 6 canonical work pages

[1]

1 Between Regulation and Accessibility: How Chinese University Students Navigate Global and Domestic Generative AI Qin Xie, University of Minnesota, USA Ming Li, The University of Osaka, Japan Fei Cheng, Kyoto University, Japan Abstract Despite the rapid proliferation of generative AI in higher education, students in China face significant barriers in acc...

work page 2023
[2]

It depends on the type of homework I’m doing. If it’s English-related, such as programming or working with English documents, I naturally think in English and turn to ChatGPT first

Figure 2 Language Preferences in Generative AI Usage 17 As Maoxin shared, “When brainstorming, I share my thoughts with generative AI, even with ChatGPT, in Chinese because it’s more convenient. After all, it’s my native language. When expressing abstract ideas, using Chinese often feels easier and more natural” Maoxin shared. As their mother tongue, Chin...

work page 2024
[7]

Digital Neocolonialism

and “Digital Neocolonialism” (Arora et al., 2023), highlighting concerns over its cultural biases. These biases can exclude students from diverse cultural backgrounds by providing examples that fail to align with their lived experiences and environments (Mollema, 2024). This form of cultural dominance risks marginalizing non-Western perspectives in educat...

arXiv 2025
[9]

13537–13547)

(pp. 13537–13547). ELRA and ICCL. Suzgun, M., Scales, N., Schärli, N., Gehrmann, S., Tay, Y., Chung, H. W., ... & Wei, J. (2022). Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261. Tacheva, J., & Ramasubramanian, S. (2023). AI Empire: Unraveling the interlocking systems of oppression in generative AI'...

arXiv 2022
[21]

& Hardiah, M

Zaim, M., Arsyad, S., Waluyo, B., Ardi, H., Al Hafizh, M., Zakiyah, M., ... & Hardiah, M. (2024). AI-powered EFL pedagogy: integrating generative AI into university teaching preparation through UTAUT and activity theory. Computers and Education: Artificial Intelligence, 7, 100335. Zhang, X., Li, D., Wang, C., Jiang, Z., Ngao, A. I., Liu, D., Peters, M. A....

arXiv 2024
[1006]

https://doi.org/10.1057/s41599-024-03526-z Davis, F. D. (1989). Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Quarterly, 13(3), 319–339. Dodds, W. B., Monroe, K. B., & Grewal, D. (1991). Effects of price, brand, and store information on buyers. Journal of Marketing Research, 28(3), 307–319. Du, L., & Lv, B...

arXiv 1989
[2012]

Previously, 9 they have some explanatory models of individual acceptance of information technology in psychology and sociology (Ayaz & Yanartaş, 2020)

are the widely used theoretical frameworks to explain user intentions to adopt and use technology (Xia & Chen, 2024). Previously, 9 they have some explanatory models of individual acceptance of information technology in psychology and sociology (Ayaz & Yanartaş, 2020). UTAUT synthesizes several earlier models, including the Theory of Reasoned Action (Davi...

work page 2012
[2022]

In fact, OpenAI had anticipated the Scaling Law for generative AI models since 2020 (Kaplan et al

ChatGPT, a large language model powered by deep learning, can generate high-quality responses to assist humans in a wide range of downstream tasks (Brown et al., 2020). In fact, OpenAI had anticipated the Scaling Law for generative AI models since 2020 (Kaplan et al. 2020), predicting that model performance would grow predictably with increases in paramet...

work page 2020

Show all 10 references

[2023]

significantly reduce reliance on computational resources. A notable example is DeepSeek, which optimized efficiency-driven techniques to achieve ChatGPT-level performance—despite limitations in computing resources (Guo et al., 2025; Liu et al., 2024). Technical Comparison of G...

2025
[2024]

legal access

Each interview lasted approximately 60 minutes. At the start of each session, the researcher read the consent form, which outlined the study’s objectives, details on data usage, and an overview of the interview questions. Participants were given the opportunity to ask question...

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

1 Between Regulation and Accessibility: How Chinese University Students Navigate Global and Domestic Generative AI Qin Xie, University of Minnesota, USA Ming Li, The University of Osaka, Japan Fei Cheng, Kyoto University, Japan Abstract Despite the rapid proliferation of generative AI in higher education, students in China face significant barriers in acc...

work page 2023

[2] [2]

It depends on the type of homework I’m doing. If it’s English-related, such as programming or working with English documents, I naturally think in English and turn to ChatGPT first

Figure 2 Language Preferences in Generative AI Usage 17 As Maoxin shared, “When brainstorming, I share my thoughts with generative AI, even with ChatGPT, in Chinese because it’s more convenient. After all, it’s my native language. When expressing abstract ideas, using Chinese often feels easier and more natural” Maoxin shared. As their mother tongue, Chin...

work page 2024

[3] [7]

Digital Neocolonialism

and “Digital Neocolonialism” (Arora et al., 2023), highlighting concerns over its cultural biases. These biases can exclude students from diverse cultural backgrounds by providing examples that fail to align with their lived experiences and environments (Mollema, 2024). This form of cultural dominance risks marginalizing non-Western perspectives in educat...

arXiv 2025

[4] [9]

13537–13547)

(pp. 13537–13547). ELRA and ICCL. Suzgun, M., Scales, N., Schärli, N., Gehrmann, S., Tay, Y., Chung, H. W., ... & Wei, J. (2022). Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261. Tacheva, J., & Ramasubramanian, S. (2023). AI Empire: Unraveling the interlocking systems of oppression in generative AI'...

arXiv 2022

[5] [21]

& Hardiah, M

Zaim, M., Arsyad, S., Waluyo, B., Ardi, H., Al Hafizh, M., Zakiyah, M., ... & Hardiah, M. (2024). AI-powered EFL pedagogy: integrating generative AI into university teaching preparation through UTAUT and activity theory. Computers and Education: Artificial Intelligence, 7, 100335. Zhang, X., Li, D., Wang, C., Jiang, Z., Ngao, A. I., Liu, D., Peters, M. A....

arXiv 2024

[6] [1006]

https://doi.org/10.1057/s41599-024-03526-z Davis, F. D. (1989). Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Quarterly, 13(3), 319–339. Dodds, W. B., Monroe, K. B., & Grewal, D. (1991). Effects of price, brand, and store information on buyers. Journal of Marketing Research, 28(3), 307–319. Du, L., & Lv, B...

arXiv 1989

[7] [2012]

Previously, 9 they have some explanatory models of individual acceptance of information technology in psychology and sociology (Ayaz & Yanartaş, 2020)

are the widely used theoretical frameworks to explain user intentions to adopt and use technology (Xia & Chen, 2024). Previously, 9 they have some explanatory models of individual acceptance of information technology in psychology and sociology (Ayaz & Yanartaş, 2020). UTAUT synthesizes several earlier models, including the Theory of Reasoned Action (Davi...

work page 2012

[8] [2022]

In fact, OpenAI had anticipated the Scaling Law for generative AI models since 2020 (Kaplan et al

ChatGPT, a large language model powered by deep learning, can generate high-quality responses to assist humans in a wide range of downstream tasks (Brown et al., 2020). In fact, OpenAI had anticipated the Scaling Law for generative AI models since 2020 (Kaplan et al. 2020), predicting that model performance would grow predictably with increases in paramet...

work page 2020