Six researchers from the University of Washington, Carnegie Mellon University and Georgetown University have traced the ancestry of 15 foundation models, from OpenAI's GPT-4 and Anthropic's Claude 4 to AI2's OLMo 3 and Nvidia's Nemotron 3, and recorded 1,161 unique artifacts and 2,778 unique dependencies, some of them nine steps removed from the model at the end of the chain. Their preprint, "Model Maps: A Documentation Framework for Tracing Foundation Model Ancestry and Understanding AI Entanglement," is dated November 13, 2026, and was circulated on LinkedIn on November 18 by privacy and digital law specialist Rosalia Anna D'Agostino under the line "AI models are like layered cakes."

In Short

Researchers drew family trees for 15 well-known AI models and found that each one is built from hundreds of other models, datasets, benchmarks and websites, often from rival companies. That matters because advertising and marketing tools run on these models, and properties such as errors, bias or licence limits can travel down the chain without the company at the end seeing them. Nothing changes in law yet: the paper is a preprint arguing that transparency rules should require developers to name each upstream source, not just say that one exists.

What the paper sets out to do

The authors build on a term from earlier work by Hong and colleagues. "AI entanglement," according to the paper, describes how "properties of artifacts upstream can influence the outputs of a foundation model downstream." Their proposal for documenting it is called a model map: "a detailed knowledge graph that represents the relationships between artifacts that influence the resulting foundation model."

The problem they describe is practical. A black-box model, in their account, can pick up security flaws, hallucinations or biased predictions from somewhere upstream, and "developers are left with no recourse to identify or mitigate the root issue," the abstract states. Regulators, auditors and litigants face a parallel difficulty, according to the authors, in "determining the entities responsible for the harms enacted by a model."

Model maps build on three earlier documentation formats the paper cites by name: ecosystem graphs, datasheets and model cards. Where an ecosystem graph tries to chart the whole industry, a model map is drawn around a single model family. In graph terms, the paper notes, an ecosystem graph can end in many applications, while "a model map for M will only have sink nodes corresponding to M." The authors compare the idea to a software bill of materials, which lets dependencies be tracked "all the way up the chain in a modular, crowd-sourced manner without relying on a centralized repository."

The six authors are Rachel Hong, Madeleine Grunde-McLaughlin and Jevan Hutson of the University of Washington; Sarah Cen of Carnegie Mellon University; Jamie Morgenstern, listed with both the University of Washington and Amazon; and Tadayoshi Kohno of Georgetown University. The paper carries the arXiv identifier 2611.18998v1 in the Computers and Society category and runs to 47 pages.

Thirteen ways one artifact shapes another

At the core of the framework sits a taxonomy of 13 dependency mechanisms, grouped into four types. The authors say they started from categories set out in earlier work by Hong and colleagues and extended them to data-based dependencies. They deliberately left out relationships that "lack substantial evidence of impacting M's outputs," such as baseline comparisons, generated training code and reuse of a model's architecture.

Several of the mechanisms describe models used as tools in building other models. One model may filter another's training data, score its answers as a judge, or watch its outputs in production. The paper is blunt about the last of these. Output guardrails "may provide some sort of promise of safety," but earlier research "has revealed shortcomings in current personal information redaction, ineffective and biased toxicity classification, and insufficient copyright detection." A similar criticism applies to guardrails used on training data, which the paper says can "enable and obfuscate the incorporation of undesired data into the resulting dataset."

The full taxonomy, as given in Table 1 of the paper, reads as follows. "Data" covers datasets, benchmarks and websites; "Model" covers both models and AI-based APIs such as ChatGPT.

Type Dependency Source Target Description (source A influences target B)
Data-based Data composition Data Data B is composed of data from A
Data-based Training data Data Model B is trained on data from A
Data-based Benchmark evaluation Data Model B is evaluated on data from A
Model-based data curation Data generation Model Data B incorporates data generated from A
Model-based data curation Data annotation Model Data B incorporates data annotated from A
Model-based data curation Data filtering Model Data B is filtered through scores or predictions from A
Model-based data curation Data guardrail Model Data B is cleaned by A to remove inappropriate data
Model-based training Model adaptation Model Model B's weights are adapted from A's weights
Model-based training Preference learning Model Model B is preference-tuned by A's generations or ratings
Model-based training Model distillation Model Model B trains on A's outputs to mimic A
Model-based deployment Model collaboration Model Model B at inference time uses A for a specific task
Model-based deployment Output guardrail Model Model B uses A to monitor B's output for compliance
Model-based deployment Model evaluation Model Data B uses A to determine performance metric

Each mechanism carries its own risk, according to the authors. Synthetic data generation has been linked to "model collapse" when models are successively trained on their own outputs. Automated labelling can create "feedback loops that shifts the distribution away from the real data and can exacerbate stereotypes." Distilled models "can inherit similar behaviors or generated preferences as the upstream model, through a process known as subliminal learning."

How a model map is built

According to Appendix A of the paper, building a model map for a model collection M follows five steps, which can be carried out by an outside party or by the developer itself.

  1. Workflow construction. The development pipeline for M is defined, with stages such as training data curation, model training and deployment, and evaluation benchmark selection, plus specific training steps such as pre-training, fine-tuning or reinforcement learning from human feedback.
  2. Artifact extraction. A node is created for every model, dataset, benchmark and API in each stage, recording its type, any documentation or usage licence, and any disclosed timestamps or modifications. Artifacts created by M's developers are "root" artifacts; pre-existing ones are "external."
  3. Relationship determination. Relationships between nodes are found, evidence for each is tracked, and the 13-category taxonomy is applied. Codebases, architecture, ablations, adversarial filtering and baseline comparisons are ignored.
  4. Artifact retrieval. For each external artifact, its research publication or source location (for example on Hugging Face) is located and cited, or recorded as missing. Steps 2 and 3 are then repeated for that artifact, omitting its own evaluation relationships, until a pre-defined stopping point is reached, "such as reaching X hops from M, or once a website, API, or another foundation model artifact is reached."
  5. Validation. Every relationship is checked again for evidence.

For the study itself, the authors stopped a path when it reached an unknown artifact, a website, an API or another foundation model. Websites, according to the paper, "possess challenging provenance gaps," while AI-based APIs "are often not transparent as to their individual components." Stopping at another foundation model has a second purpose: it lets one model's map plug into another's.

The work was done by hand. The first author followed the procedure using technical reports and the arXiv papers they cite as sources, and the maps were recorded in Obsidian, a linked note-taking application, with each note an artifact and each link a dependency carrying the quotation that proves it. A second author re-annotated a stratified random sample of 285 dependencies and reached what the paper calls "almost perfect agreement" (Fleiss kappa 0.884). The OLMo 3 map alone, according to the authors, "took several weeks to document and another several weeks to retrace and validate."

Fifteen models, measured

The sample covers language model families with public technical reports published between March 2023 (Alpaca) and December 2025 (Nemotron 3), chosen from 13 organisations or affiliated universities across "a wide range of reported transparency levels." Four are closed-weight: GPT-4, Apple's foundation model, Claude 4 and Gemini 2.5.

The table below combines Table 2 of the paper with two of its charts. "Layers" is the largest number of steps between any artifact and the target model. The last two columns are read from Figures 6 and 4: the share of all dependencies whose upstream artifact is unnamed, and the share of model-based dependencies whose source is a closed-weight model.

Model map (developer as named in the paper) Weights Nodes Edges Layers Root nodes Unknown-source share Closed-model share
Alpaca 7B (developer not named) Open 6 5 2 2 0.00 0.50
GPT-4 (OpenAI) Closed 38 57 4 10 0.12 0.47
Mistral 7B (Mistral AI) Open 31 31 4 2 0.03 0.00
Apple Foundation Model (Apple) Closed 71 107 4 13 0.12 0.81
Llama 3 (Meta) Open 121 184 5 21 0.06 0.31
Phi 4 (Microsoft) Open 75 143 5 14 0.09 0.69
DeepSeek-V3 (DeepSeek) Open 167 237 7 5 0.03 0.48
Gemma 3 (Google) Open 46 102 3 12 0.20 0.56
Qwen3 (Alibaba) Open 90 137 3 11 0.05 0.53
Claude 4 (Anthropic) Closed 56 74 4 12 0.12 0.89
Gemini 2.5 (Google) Closed 115 201 4 11 0.06 0.72
SmolLM3 (Hugging Face) Open 418 782 7 4 0.01 0.42
DR Tulu (AI2) Open 104 171 9 5 0.02 0.53
OLMo 3 (AI2) Open 681 1,357 8 42 0.02 0.41
Nemotron 3 Nano (Nvidia) Open 259 429 9 13 0.05 0.40

Many maps share artifacts, which is why the per-model counts add up to more than the unique totals. By PPC Land's calculation, the 15 maps contain 2,278 nodes and 4,017 edges before duplicates are removed, against the 1,161 unique artifacts and 2,778 unique dependencies the paper reports.

Open models built partly on closed ones

The most pointed finding concerns openness. "Across all model maps except for Mistral 7B, a proportion of model-based dependencies are based on closed models, hence demonstrating that even developers for open-weight models still rely on proprietary models," the paper states. Among the 11 open-weight models, the closed share ranges from 0.31 for Llama 3 to 0.69 for Microsoft's Phi 4.

OLMo 3 gets particular attention. AI2 describes it as "fully open," yet 0.41 of its model-based dependencies come from closed-weight models, according to the paper's Figure 4, and some of AI2's own intermediate models "are directly dependent on closed models as well." The four closed-weight models show shares from 0.47 for GPT-4 to 0.89 for Claude 4.

Size tells its own story. OLMo 3's map has 681 nodes, while the closed models' maps run from 38 to 115, and the authors attribute the gap to disclosure: closed models are "significantly smaller in size due to lack of disclosed dependencies." For the smaller maps, benchmark evaluation makes up "a disproportionate amount of dependencies," which the authors read as "supporting the observation that proprietary model organizations do not disclose details of training data curation."

Lineage that crosses company lines

OpenAI appears "predominantly across many dependencies," according to the paper, which counted the ten organisations most often found as sources. Developers lean on their own earlier work, and the paper names OpenAI for GPT-4, Google for Gemma 3 and Gemini 2.5, Anthropic for Claude 4, Apple for its foundation model, Alibaba for Qwen3 and Meta for Llama 3. Read from Figure 5, own-organisation artifacts account for 0.46 of edges in GPT-4's map, 0.38 in Claude 4's, 0.31 in Gemini 2.5's, 0.32 in Gemma 3's, 0.25 in Llama 3's, 0.24 in Qwen3's and 0.19 in Apple's. "However, various foundation models implicitly depend on artifacts from competitors," the authors write.

In the training data of the three largest maps (Nemotron 3, SmolLM3 and OLMo 3), the most common upstream models include GPT-4, GPT-4o-mini, GPT-4o, DeepSeek-R1, GPT-3.5-turbo, two Llama instruction models, Qwen 2.5 32B Instruct and GPT-4.1. The ten most common upstream websites in those maps are Wikipedia, Common Crawl, GitHub, Codeforces, AoPS, Reddit, StackExchange, arXiv, CodeChef and AtCoder.

Ancestors nobody named

Every map except Alpaca's contains dependencies whose source is not named, "even for foundation model reports that claim to be completely open-source." The paper says closed-weight models have "relatively higher proportions" of such dependencies, at 0.12 for GPT-4, Apple's model and Claude 4. The single highest share in Figure 6, however, belongs to an open-weight model: 0.20 for Google's Gemma 3. Gemini 2.5, closed, sits at 0.06. For OLMo 3, the authors also recorded six nodes that are cited but whose documentation "could not be found or no longer exists," among them the retired Freebase website.

Loops

Some models consume their own output. The paper finds "concrete evidence that some foundation models are iteratively trained on their own generations," naming Llama 3.1 405B Instruct, DR Tulu SFT and Apple's FM-server. Others grade themselves. "Gemini 2.5, for instance, is evaluated on ZeroBench and an internal automated red-teaming benchmark, both of which use the same model as an LLM judge," the authors write, which "raises concerns of self-preference bias."

Claude 3's mistakes inside Claude 4

The case that gives this article its headline came from Anthropic's own documentation. "Claude 4's technical report indicates that the authors observed that Claude 4 regurgitated hallucinations that were generated by Claude 3, and determined that they had inadvertently web-scraped Claude 3's generations as part of Claude 4's training data," the paper states, citing Anthropic's 2025 system card for Claude Opus 4 and Claude Sonnet 4. A hallucination produced by one generation of the model had been posted online, collected from the web and learned by the next.

Anthropic responded with more data. The paper quotes the system card: "To mitigate these hallucinations, we introduced a targeted dataset into Claude Opus 4's training." The authors use the episode to argue that provenance tracking "can therefore expose and prevent undesired ingestion of synthetic data." The accidental case sits next to deliberate ones. OLMo 3, for example, "is pre-trained on generations from OLMo 2," according to the paper.

The introduction lists other ways problems move downstream. Researchers have planted poisoned backdoors in Qwen and Llama models that "propagated to backdoor vulnerabilities after fine-tuning." Stereotypes embedded in filtering models can carry into cleaned datasets. Failures in privacy guardrails can put personal information into outputs. Upstream artifacts "have license restrictions that may be violated with downstream usage, and training a competitor on a proprietary model's generated data may lead to accusations of model theft."

A licence several steps away

That last point gets a concrete example. According to the paper, OLMo 3's developers "explicitly avoid immediate dependencies on Llama models" because of the "less-permissive Llama license attached." Llama shows up anyway, a few steps further back. OLMo's supervised fine-tuning stage uses prompts from the OpenThoughts3 dataset, which includes prompts from Big-Math, and those prompts "are rewritten by Llama 3.1." The paper's conclusion is careful: "Model maps can thus reveal when downstream developer requirements are not followed multiple degrees upstream." The section heading calls these "potential license violations"; the paper makes no finding that a licence was breached.

Distillation, benchmarks and contamination

Distillation, in which a smaller model is trained to imitate a larger one, appears throughout. Gemma 3's 1B, 4B and 12B models were produced this way. DeepSeek-V3 was distilled from DeepSeek-R1, a model from the same company. Apple's on-device pre-trained model used "knowledge distillation with an unnamed teacher model." The authors only labelled a dependency as distillation when a report used the word, while noting that "a teacher model's generations are often implicitly used to train another model." As corporate attention turns to preventing or regulating model theft, they write, "model maps may reveal blurred lines as to what counts as distillation."

Benchmarks have ancestries too. OpenAI's GPT-4 series makes up "a large proportion of model-based dependencies" in the benchmark evaluation stage, used to build benchmark data or to judge answers. Some benchmarks also end up in training data. "OLMo 3 is trained on prompts of its own benchmark AGIEval," the paper states, "while question-answer pairs from TheoremQA make their way into training and evaluating the Nemotron 3, SmolLM3, and OLMo 3 models." Benchmarks and training sets also draw on the same websites, "including Wikipedia, Github, or news sites like CNN or Daily Mail," which the authors describe as sites of "potential data contamination that warrant additional investigation."

How often do technical reports mention models that work alongside the main model once it is deployed? Rarely. Model collaboration and output guardrail dependencies "are only documented in four model maps," according to the paper.

Why developers say they made these choices

The authors also read the reasons developers give. Of the 2,778 excerpts that record a dependency, 1,014 include a justification. Two authors built a codebook of 21 codes in eight value categories, five of them taken from a 2022 study of values in machine learning research by Abeba Birhane and colleagues, and reached what the paper calls substantial agreement (Fleiss kappa 0.705). An excerpt could carry more than one category.

Value category Excerpts Most frequent code in the category (count)
Generalization 255 Include diverse data (188)
Data quality 223 Improve downstream quality (153)
Prior work 140 Select a commonly used source (71)
Model performance 130 Improve downstream performance (73)
Efficiency 97 Increase scale (39)
Research standards 75 Select an easily accessed source (50)
Safety 62 Filter sensitive content (33)
Novelty 31 Overcome an existing gap (17)

The authors are sceptical of the vocabulary. Words such as "data fidelity" or "high-quality" are "often described in objective manners" yet "remain subjective determinations of data production," according to the paper. The share of dependencies that come with any justification also varies widely by report, from 0.87 of edges for Mistral 7B to 0.14 for DeepSeek-V3 and OLMo 3, read from Figure 10.

What the authors want from regulators

The policy argument rests on one sentence about California. "Existing legislation (e.g., California AB 2013) requires foundation model providers to disclose the existence of dependencies, such as whether synthetic data or filtering was used, but not the exact source artifacts or how the dependency was evaluated," the paper states. That looseness, the authors argue, "may lead to uninformative compliance" and does not address risks of diminished transparency "as defined in the NIST AI Risk Management Framework."

Their recommendation is direct: "we argue that regulation of general-purpose AI systems should account for upstream dependencies and that model maps provide a framework for such documentation." Section 6.2 is titled "Dependency documentation should be a component of mandatory regulatory disclosure."

The authors expect pushback. Model providers "may resist disclosing dependencies to the public" by arguing that training data dependencies are trade secrets, a stance that "creates information asymmetries for regulators and auditors seeking to address risks of AI systems." Their reply is that the maps already reveal a good deal about proprietary developers' practices from public documents alone.

AB 2013 has already been tested in court. xAI sued California over the law on December 29, 2025, days before it was due to take effect on January 1, 2026, arguing that it violated the Takings Clause, the First Amendment and due process. On March 4, 2026, Judge Jesus G. Bernal denied xAI's request for a preliminary injunction, finding that xAI had standing but had not shown it was likely to win on any of the three claims. The paper's proposal would go further than the law xAI challenged.

The LinkedIn framing

D'Agostino, whose LinkedIn headline reads "Privacy and Digital Law Expert | Host of the Legal4Tech - The Podcast," posted the paper as a 47-page document under the title "AI TRANSPARENCY - MODELS ARE ALL ENTANGLED." Her post places the paper against a regulatory backdrop that starts with the EU AI Act. According to D'Agostino, "aside from the EU AI Act (still not fully enforceable) and the South Korean AI Basic Law, other forms of comprehensive AI regulation obliging AI companies to conduct strict risk assessment audits and pull out potentially risky systems have been lagging behind."

Transparency rules for frontier model developers have moved faster, in her reading. "China, California, Switzerland (only sectorial) and the UK have either passed or are in the process of approving AI transparency requirements for frontier model providers, adding to the now enforceable sections of the AI Act that ask the same of GPAI model providers," the post states. "Still, greater model transparency has its limits too, one amongst many, it fails to account for AI entanglement." The post does not cite the specific Chinese, Swiss or UK measures it refers to, and the paper itself discusses only California's AB 2013 and the NIST framework by name.

In the United States, the attention has been on frontier developers. California Governor Gavin Newsom's Executive Order N-9-26 of September 18, 2026 asked for recommendations on a frontier AI kill switch by November 16. A 37-page federal discussion draft that surfaced on October 19 would set its frontier line at 10^25 FLOP, against 10^26 in California's SB 53.

In Europe, the AI Act's obligations for general-purpose AI models took effect on August 2, 2025, with a final General-Purpose AI Code of Practice published in July 2025 covering transparency, copyright and safety. The Commission's guidelines of July 18, 2025 set an indicative threshold of more than 10^23 FLOP for a general-purpose model.

D'Agostino closes on an analogy. "Remember that saying, we all have a piece of Genghis Khan in our DNA? Well, at least the 15 open-weight and proprietary models analysed in the study can state the same about their training data. That's definitely something that regulators and supervisors need considering." Comments under the post picked up the same theme. "Most transparency rules still assume each model is built from scratch on its own clean dataset, which just isn't how things work anymore," wrote Juan Sáez Sierra, whose LinkedIn headline lists data science and AI governance.

Method and limits

The authors are open about what the maps cannot do. Each artifact was traced through a single textual source, so dependencies disclosed elsewhere, such as in code repositories, may be missing. The maps record only what developers chose to publish, and internal artifacts at closed companies are largely invisible. Models outside the language domain were not mapped, and downstream applications and users fall outside the framework by design.

The paper also tests automation and finds it wanting. Comparing its maps with dependency graphs built by another research team "through prompting Claude," the authors report that "many dependencies reported by Claude are incorrectly assigned to the wrong artifacts, and many upstream data sources are completely omitted." That is one reason they chose manual construction, and why they warn that "generative tools may miss upstream artifacts and hallucinate non-existing relationships."

They acknowledge a misuse risk as well: "an adversary could use knowledge gained from model maps to select upstream websites to manipulate" in order to cause harm downstream. Their response is that the information is already public, only scattered, and that "information obscurity is not an adequate barrier for a motivated adversary." According to the paper, generative AI tools "were not used for manuscript preparation and data analyses." The work was supported by US National Science Foundation awards CCF-2045402 and IIS-2402647 and a Simons Foundation grant.

Two small inconsistencies appear in the text. The OLMo 3 map is described with 1,357 edges in Figure 1 and Table 2 but with 1,351 edges in Section 5.1.1. And while the abstract says the maps are released at modelmaps.ai, the conclusion says the authors "intend to release this data along with accompanying code."

Why this matters for marketing

Most of the AI features now sold to advertisers sit on top of foundation models like the 15 in this study, and many agency and platform tools now chain several models together. That is the "model collaboration" mechanism in the taxonomy, which the authors found documented in only four of 15 maps. A marketing team choosing a model inherits whatever is upstream of it, whether that is a closed model's outputs inside an "open" model's training data, a benchmark that leaked into training, or a dataset whose licence the developer tried to avoid.

The paper itself lists model selection as a use case. An application developer "may wish to avoid a model if its model map shows excessive data generation dependencies on prior versions of the same model, due to concerns of model collapse," or "may refrain from using a model if its model map indicates that upstream artifacts are accompanied with licenses that restrict downstream usage." Model collapse has a record in PPC Land's coverage: a Nature study published on July 24, 2024 found language models degrade when trained on synthetic data.

Publishers have a direct stake in the website layer. Common Crawl is one of the ten most common upstream websites in the training data of the three largest maps, alongside Wikipedia and GitHub. On April 29, 2026, the News/Media Alliance asked Common Crawl to remove publisher content and bar AI use in its terms. The study's benchmark findings point the same way, with news sites such as CNN and Daily Mail feeding both benchmarks and training sets.

And on hallucinations, the Claude case adds a mechanism to an existing debate. In September 2025, OpenAI researchers argued that models hallucinate partly because benchmarks reward guessing over admitting uncertainty. The model maps describe a second route, in which one model's errors are published online and then collected as training data for the next.

What the paper does not establish

The paper is a preprint and does not state that it has been peer reviewed or accepted at a venue. It does not report contacting any of the 13 developers to confirm the maps, does not measure how much any single dependency changes a model's behaviour, and does not examine advertising or marketing products. Its figures for closed models describe what was disclosed, not what was used.

Key facts

  • Paper: "Model Maps: A Documentation Framework for Tracing Foundation Model Ancestry and Understanding AI Entanglement," 47 pages
  • Identifier and date: arXiv:2611.18998v1 [cs.CY], November 13, 2026
  • Authors: Rachel Hong, Madeleine Grunde-McLaughlin, Jevan Hutson (University of Washington); Sarah Cen (Carnegie Mellon University); Jamie Morgenstern (University of Washington, Amazon); Tadayoshi Kohno (Georgetown University)
  • Data location given in the paper: https://modelmaps.ai
  • Models mapped: 15 language model families, technical reports dated March 2023 to December 2025; 4 closed-weight (GPT-4, Apple Foundation Model, Claude 4, Gemini 2.5), 11 open-weight
  • Totals: 1,161 unique artifacts; 2,778 unique dependencies; artifacts up to 9 hops from the target model
  • Largest map: OLMo 3, 681 nodes and 1,357 edges (1,351 in Section 5.1.1)
  • Taxonomy: 13 dependency mechanisms in four types (data-based, model-based data curation, model-based training, model-based deployment)
  • Annotation agreement: Fleiss kappa 0.884 on 285 sampled dependencies; 0.705 for justification coding
  • Justifications: 1,014 of 2,778 excerpts; codebook of 21 codes in 8 categories
  • Closed-model share of model-based dependencies, open-weight models: 0.31 (Llama 3) to 0.69 (Phi 4); 0.00 for Mistral 7B only
  • Funding: US National Science Foundation awards CCF-2045402 and IIS-2402647; Simons Foundation grant
  • LinkedIn post: Rosalia Anna D'Agostino, "AI models are like layered cakes: grasping AI entanglements in frontier AI models," November 18, 2026; document title "AI TRANSPARENCY - MODELS ARE ALL ENTANGLED"; https://www.linkedin.com/posts/rosaliaannadagostino_ai-transparency-models-are-all-entangled-ugcPost-7528713978219421697-_Uks/

Timeline

Summary

Who: Rachel Hong, Madeleine Grunde-McLaughlin, Jevan Hutson, Sarah Cen, Jamie Morgenstern and Tadayoshi Kohno, from the University of Washington, Carnegie Mellon University, Georgetown University and Amazon; Rosalia Anna D'Agostino, a privacy and digital law specialist who shared the paper on LinkedIn. The models mapped come from OpenAI, Anthropic, Google, Meta, Microsoft, Apple, Alibaba, DeepSeek, Mistral AI, Nvidia, Hugging Face and AI2, plus Alpaca.

What: A preprint proposing "model maps," hand-built graphs of every model, dataset, benchmark, website and API that fed into a foundation model, with a taxonomy of 13 dependency mechanisms. Applied to 15 models, the method recorded 1,161 unique artifacts and 2,778 unique dependencies, found closed-model dependencies in 10 of 11 open-weight models, and documented cases such as Claude 4 reproducing Claude 3's hallucinations and Llama-rewritten prompts reaching OLMo 3 despite its developers avoiding Llama directly.

When: The preprint is dated November 13, 2026. D'Agostino shared it on November 18, 2026. The models studied have technical reports from March 2023 to December 2025.

Where: Posted to arXiv under identifier 2611.18998, with the data location given as modelmaps.ai, and circulated on LinkedIn.

Why: The authors argue that transparency rules such as California's AB 2013 ask developers to say that dependencies exist, not to name them, which leaves regulators, auditors and buyers unable to trace where a model's errors, biases or licence limits come from. For advertisers and publishers, the models behind marketing tools carry the same inherited properties, and publisher content sits at the top of many of these chains.