Research opportunity · Concentration, Stack & Sovereignty
Does concentration at the model layer make apps fail alike? A test on China's register
In short: has this been done?
Related work: Partly addressed. Related studies cover parts of this question; what they leave open is set out below. Only a first-pass search has been done, so it is not yet known whether a fuller search would find a study that does the whole design.
How far checked: First-pass screen only. A first-pass AI web search on 16 October 2026. No deeper check yet. Every search and check so far was done by AI; no person or domain expert has checked them yet.
The searches, the nearest work found and what the search could not reach
Across debates: it also draws on Regulation-in-Practice: Rights, Compliance & Developmental Steering, What kind of thing is generative AI, and does the answer predict anything? and Individual gain, collective narrowing? Does AI make science, creative work and shared knowledge more alike?, and is listed on each of their pages.
This page shows RO-LD11-07 version v1, whose record never changes. It is the current version; if the opportunity is revised, the newer version appears at its main page.
Question
In China's register of generative-AI services, where every launched application must display the filed service or model it uses, how concentrated is applications' dependence on a few base models, and on a fixed set of test cases are the errors, refusals and biased outputs of applications that share a base model more strongly correlated than those of applications built on different models, beyond the correlation already found between models from different providers?
Gap type: Causal mechanism not isolatedA PhD projectLow to medium costRelated work: Partly addressed
Why it matters
LD-11 holds that concentration at one layer of the stack does not by itself establish power; a mechanism has to be shown. One mechanism is proposed in LD-01 and partly tested in LD-09. The foundation-model account predicts that homogenization "creates single points of failure". LD-09 records that models' errors are correlated on benchmarks, even across providers, and that applicants screened by one hiring vendor's software were recommended for rejection by every position they applied to more often than chance, on the authors' account; these are the software's recommendations, not observed hiring decisions. No study links which deployed generative-AI products depend on which base model to whether they fail on the same cases. Elsewhere that dependence is usually hidden, so a related opportunity on LD-01 (RO-LD01-03) asks the question without a way to know which products share a model. LD-12 records that China's regulator requires launched applications to name the filed service or model they use, which makes a dependence map public for one large market.
Evidence that motivates it
- Concentration claims in AI must identify the stack layer, market definition, metric, geography and period; concentration at one layer does not establish concentration or power at another. (LD11-HQ-C01 v1.0)
How this study uses it: Concentration claims must name the layer, and concentration at one layer does not establish power at another; this design tests one route from the model layer to downstream outcomes. - Menlo Ventures estimated late-2025 enterprise LLM API shares of Anthropic 40%, OpenAI 27% and Google 21%, compared with OpenAI 50% and Anthropic 12% in 2023. (LD11-HQ-C13 v1.0)
How this study uses it: A venture investor's modelled estimate, from a survey of US enterprise decision-makers, put most enterprise language-model API use with three providers in late 2025, which suggests shared dependence; it is not an audited or global share, and the investor has an interest. - China's CAC notices require launched generative-AI applications or functions to display the filed/registered service or model they use, including the model name and filing or launch number. (LD12-HQ-C13 v1.0)
How this study uses it: China's regulator requires launched generative-AI applications to display the filed service or model they use, with its name and filing number, which supplies the dependence map. - China's CAC reported 1,112 generative-AI services filed and 731 applications/functions registered by 31 August 2026, up from 988 and 598 at the end of June. (LD12-HQ-C12 v1.0)
How this study uses it: By 31 August 2026 there were 1,112 filed services and 731 registered applications, enough for a comparison between applications that share a model and those that do not. - Bommasani and colleagues named “foundation models” for their “critically central yet incomplete character,” argued that their significance lies in “emergence and homogenization,” warned that homogenization “creates single points of failure,” and predicted that the barrier to entry for developing them “will continue to rise.” (LD01-SL-C28 v1.0)
How this study uses it: The foundation-model account warns that homogenization "creates single points of failure" but gives no measurement. - Across more than 350 language models, when two models both answer a leaderboard question wrongly they give the same wrong answer 60% of the time on one dataset, and larger, more accurate models have highly correlated errors even across different architectures and providers. (LD09-SL-C25 v1.0)
How this study uses it: Across more than 350 models, errors are correlated on leaderboard questions, highly so among larger models even across providers; this is the baseline a shared-model effect must exceed, and the claim does not reach effects on real decisions. - In field data from one hiring-software vendor covering 3 million applicants, 4% of applicants who applied to 10 positions were recommended for rejection by all of them, a rate the authors report as higher than expected by chance. (LD09-SL-C24 v1.0)
How this study uses it: Applicants screened by one hiring vendor's software were recommended for rejection by all the positions they applied to more often than chance, according to the authors; these are algorithmic recommendations, not hiring decisions, and the evidence is at the level of a shared vendor, not a shared base model.
Each claim is shown as it read when this version was published.
A possible design
Build a dataset from the regulator's successive public lists. Because an application may display a filed service rather than a base model, add a mapping step that traces each displayed service to its underlying base model from filing records and providers' own documentation, and record where it cannot be traced. Measure concentration of dependence (share of applications per base model, by sector). Draw a stratified sample of applications in two or three sectors where outputs can be checked (customer service, education, health information), sized so that a difference in error correlation between same-model and different-model pairs can be detected against the high baseline correlation reported across providers. Put the same fixed set of test cases to every sampled application, including cases where the answer is known, cases that should be refused and paired cases that differ only in a group attribute. Compare the correlation of errors, refusals and group differences within same-model and different-model pairs, controlling for sector and firm size, and repeat after a base-model update to see whether failures move together.
Data needed
The regulator's published filing and registration lists over time; filing records and documentation to map displayed services to base models; access to the sampled applications as an ordinary user; a validated test set in Chinese with known answers and paired cases; Chinese-language research capacity.
Feasibility
A PhD project. Low to medium; the lists are public and the applications are used as an ordinary user would, but mapping services to base models and building a Chinese test set take researcher time. Applications may wrap a base model in their own prompts, retrieval and filters, which may decorrelate failures; that is part of what the study measures. The displayed model is what the application declares, not an audited fact. Work on Chinese platforms may need a local partner.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 16 October 2026; not yet checked in depth).
The component-sharing hypothesis has been tested on benchmarks, where shared training data reliably homogenized outcomes and a shared foundation model gave mixed results depending on how it was adapted. Error correlation across hundreds of language models has been traced partly to shared architectures and providers, and correlated failure has been shown across commercial prediction services and for one hiring vendor. China's registration data has been analysed descriptively by a commercial research firm. No study was found that uses declared base-model dependence among deployed generative-AI products, in China or elsewhere, to test whether products that share a base model fail on the same cases. The search was web-based and partial.
How strong this evidence is
- Who searched and checked: no checker is recorded beyond the first-pass screen, which was part of the AI drafting (Claude (SocioLens Project Lead), AI-drafted from the LD-11, LD-12, LD-01 and LD-09 debates). Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 6 of the 6 nearest works below were judged from an abstract, summary or search listing only; each is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web searches and arXiv records on the date shown, with two further works added by the independent review of 16 October 2026; Chinese-language literature and Chinese databases were not searched, and work on the filing register may exist there. Absence of a hit is not proof that the work does not exist.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Bommasani, Creel, Kumar, Jurafsky and Liang, "Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?", NeurIPS 2022 Judged from an abstract, summary or search listing only
Abstract read. Proposes and tests the component-sharing hypothesis on fairness benchmarks; shared training data reliably increases homogenization, while sharing a foundation model gives mixed results that depend on the adaptation method. Benchmarks, not deployed products. - Kim and colleagues, "Correlated Errors in Large Language Models", ICML 2025 (a Source of LD-09) Judged from an abstract, summary or search listing only
Abstract read. Over 350 models; identifies shared architectures and providers among the drivers of error correlation, while larger models are highly correlated even across them. Its abstract backs claim LD09-SL-C25. - Toups, Bommasani, Creel and others, "Ecosystem-level Analysis of Deployed Machine Learning Reveals Homogeneous Outcomes", NeurIPS 2023 Judged from an abstract, summary or search listing only
Seen in search results only. Across commercial prediction services in three modalities, some users are misclassified by every available model; not generative-AI products, and their shared components were not observed. - Bone, Stephany and del Rio-Chanona, "Monocultural Biases: Correlated biases in large language models lead to unequal systemic exclusion rates in hiring", arXiv:2609.22169 (2026) Judged from an abstract, summary or search listing only
Abstract read. Ten models' hiring decisions become more correlated after post-training, raising systemic exclusion in simulated callbacks; a test of models, not of deployed products that share a base model. - Bommasani and colleagues, "Algorithmic Monocultures in Hiring", FAccT 2026 (a Source of LD-09) Judged from an abstract, summary or search listing only
Abstract read for LD-09, where it backs claim LD09-SL-C24. Field data from one hiring-software vendor; shared vendor, not shared base model. - Trivium China, "Seeking the next DeepSeek - What China's generative AI registration data can tell us about China's AI competitiveness" Judged from an abstract, summary or search listing only
Seen in search results only. A descriptive analysis of the registration lists (counts by province and business model); whether it maps applications to base models was not checked, and it does not test outputs.
Search queries used
correlated failures downstream applications sharing same foundation model empirical audit homogeneous outcomesChina CAC generative AI filing registered applications base model dependence analysisalgorithmic monoculture outcome homogenization foundation models deployed systems ecosystem-level analysisToups Bommasani "Ecosystem-level analysis of deployed machine learning reveals homogeneous outcomes"LLM hiring screening monoculture same applicants rejected across models systemic exclusion 2025Trivium China CAC generative AI registration data analysis base models applications
Cite this page
SocioLens project. (2026, October 16). Does concentration at the model layer make apps fail alike? A test on China's register [Research opportunity RO-LD11-07, version v1]. SocioLens. https://sociolens-web.pages.dev/opportunities/ro-ld11-07/v1/
@misc{sociolens_ro_ld11_07_v1,
author = {{SocioLens project}},
title = {Does concentration at the model layer make apps fail alike? A test on China's register},
year = {2026},
date = {2026-10-16},
howpublished = {\url{https://sociolens-web.pages.dev/opportunities/ro-ld11-07/v1/}},
note = {Research opportunity RO-LD11-07, version v1}
}The link points to this version, whose record never changes. For a finding, cite the original study as well: SocioLens summarizes research and does not replace it.
Provenance
RO-LD11-07 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-11, LD-12, LD-01 and LD-09 debates on 2026-10-16. Not yet reviewed. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.