In 2019, Mitchell et al. published "Model Cards for Model Reporting" in the proceedings of the ACM Conference on Fairness, Accountability and Transparency. The paper proposed a structured documentation framework for machine learning models: a concise artifact covering intended use cases, performance benchmarks, evaluation data, ethical considerations and training data composition. The proposal was practical, specific and technically achievable. It has since become the most-cited framework in AI transparency discourse. It has also been largely ignored in the one place it matters most: the disclosure practices of the largest foundation model developers operating in 2026.
This post audits what Anthropic, OpenAI, Google and Meta actually document about their training data. It measures that documentation against the Mitchell et al. specification. And it examines why the gap between the framework and current practice is not a technical oversight but a structural condition that affects data sovereignty, consent architecture and legal accountability for every individual whose data was ingested without knowledge.
What Model Cards Were Designed to Do
The Mitchell et al. framework specified six reporting sections: model details, intended use, factors (demographic and environmental), metrics, evaluation data and training data. The training data section was not decorative. It asked developers to document the datasets used, the motivation for those choices, any preprocessing applied and the provenance of the underlying sources.
The word provenance is doing significant work in that specification. Provenance means origin tracing: where the data came from, who produced it, under what conditions it was collected and whether appropriate consent or licensing existed at the point of collection. This is not a vague aspiration. It maps directly to established data governance concepts formalized in ISO/IEC 5259 (the data quality series) and to lineage tracking requirements embedded in GDPR Article 14, which obligates controllers to disclose the source of personal data when it was not collected directly from the data subject.
Mitchell et al. envisioned model cards as living documents, updated as models were fine-tuned or retrained. They were explicit that transparency about training data was foundational to any meaningful fairness or accountability evaluation. You cannot audit bias in outputs without knowing what went into training. You cannot assess copyright or consent exposure without understanding the composition of training corpora.
That was the design. What followed in practice was something else entirely.
The Disclosure Gap: What Major Labs Actually Publish
An audit of publicly available model documentation from the four largest foundation model developers reveals a consistent pattern: performance metrics are published with precision while training data provenance is disclosed with deliberate vagueness.
OpenAI's documentation for the GPT-4 series, including the GPT-4 Technical Report released through arXiv (arXiv:2303.08774), dedicates substantial space to benchmark evaluations across standardized tests. On training data, it states the model was trained on data from "the internet" up to a knowledge cutoff and supplemented with licensed data and RLHF data from human contractors. No dataset names are listed. No licensing structures are disclosed. No methodology for filtering personal data is described. The report explicitly declines to detail training data composition, citing competitive sensitivity.
Anthropic's model documentation for the Claude series follows a similar pattern. Published safety evaluations and Constitutional AI methodology papers (arXiv:2212.08073) describe the RLHF and reinforcement learning from AI feedback pipeline in detail. Training data sources for the base pretraining corpus are not publicly documented. Anthropic has described its approach to responsible scaling but has not published a training data card that satisfies the Mitchell et al. specification.
Google's technical documentation for the Gemini model family, published through DeepMind, describes a multimodal training approach and references web documents, books and code. The Gemini Technical Report (arXiv:2312.11805) notes the use of SentencePiece tokenization and describes data quality filtering at a methodological level. Specific dataset provenance, licensing status of web-scraped content and consent basis for personal data inclusion are absent from public documentation.
Meta's LLaMA series represents partial exception. The LLaMA 2 paper (arXiv:2307.09288) names specific datasets in the pretraining corpus including CommonCrawl, C4, GitHub, Wikipedia, Books and ArXiv. This is meaningfully more transparent than the other three developers. The disclosure still does not address individual data subject rights, consent basis under applicable privacy law or the methodology used to identify and handle personal data within those sources.
Training Data Provenance: The Missing Layer
Naming a dataset is not the same as disclosing provenance. CommonCrawl, which appears in the training corpora of most major foundation models, is itself an aggregation of web crawls containing content scraped without individual consent. The fact that CommonCrawl is named does not tell an affected data subject whether their blog post, medical forum comment or professional profile was included in training. It does not document the legal basis under which that content was ingested.
This distinction matters under GDPR Article 6, which requires a lawful basis for processing personal data. Legitimate interest, the basis most commonly invoked for web-scraped training data, requires a balancing test against the rights and freedoms of affected individuals. That test cannot be independently evaluated when the training corpus is undisclosed.
The W3C Data Catalog Vocabulary (DCAT, W3C Recommendation) and the PROV-O ontology (W3C Recommendation, PROV-DM specification) provide existing technical infrastructure for machine-readable provenance documentation. Neither has been adopted by any major foundation model developer as a standard component of model cards. The vocabulary exists. The will to implement it does not.
In the Personal Data Asset Origination System (PDAOS) framework developed at Own Your Data Inc, provenance tracking is treated as a first-order requirement rather than an optional enhancement. The argument, elaborated in Dr. Patrick Fisher's work and in Volume 6 of The Invisible Series, is that data without documented origin is data without accountability. A model trained on provenance-free corpora inherits that unaccountability at scale.
Why This Matters for Data Sovereignty
The failure of model cards to document training data provenance is not an abstract transparency issue. It has direct consequences for individuals asserting data sovereignty over their personal information.
Under the California Consumer Privacy Act (CCPA) as amended by CPRA, individuals have the right to know whether their personal information has been sold or shared. When that personal information exists as a vectorized representation inside a foundation model's weight matrix, the question of whether a right-to-know or right-to-delete claim can be operationalized depends entirely on whether the original data source is traceable. Without model cards that document training corpora at the dataset-and-source level, those rights become practically unenforceable.
The EU AI Act, which entered its phased application schedule in 2026, includes provisions under Article 53 requiring providers of general-purpose AI models to maintain technical documentation about training data. The documentation requirements include information about training data sources and the measures taken to comply with Union copyright law. Enforcement of those provisions will be difficult when major model developers have established a norm of vague disclosure.
Data sovereignty is not a slogan. It is a technical and legal claim that individuals retain meaningful control over information that originates from or about them. That claim collapses when training data is treated as a corporate trade secret rather than a domain requiring structured public disclosure.
Consent Architectures and the Model Card Blind Spot
The Mitchell et al. framework did not explicitly address consent as a reporting field. This is a gap in the specification itself, one that was understandable in 2019 when foundation model scale was not yet the dominant paradigm. In 2026, that omission is no longer defensible as a template for model documentation.
Consent receipts, as specified in the Kantara Initiative Consent Receipt Specification and referenced in ISO/IEC 29184 (Online Privacy Notices and Consent), provide a structured record of what an individual agreed to at the point of data collection. No major foundation model developer publishes a training data consent audit that maps collected data to consent receipts or equivalent authorization records.
The gap is structural. Web-scraped training data is, by definition, collected without individual consent receipts. The developers know this. Regulators know this. What is missing is a disclosure framework honest enough to document the absence of consent alongside the data sources that lack it. An adequate model card, updated for 2026 practice, would include a consent coverage field: for what percentage of training data was explicit consent or clear licensing obtained? That field does not appear in any major developer's public documentation.
At mydatakey.org, the technical infrastructure being developed under Own Your Data Inc explores how cryptographic consent records can travel with data assets as they move through processing pipelines. The model card framework needs an equivalent mechanism: a consent provenance layer that survives from original collection through model training and into deployment documentation.
What Adequate Disclosure Would Actually Require
Measuring current practice against the Mitchell et al. baseline, and extending that baseline to address consent and legal compliance, an adequate model card for a 2026 foundation model would require the following disclosures:
- Named datasets with version identifiers and collection timestamps
- Licensing status for each dataset: open, commercially licensed, scraped without license, public domain
- Consent coverage percentage: the proportion of training data for which explicit consent or equivalent legal basis is documented
- PII detection and filtering methodology, including false-negative rate estimates
- Data subject rights operationalization: how the developer responds to deletion or access requests affecting training data
- Update cadence: when and how the model card is revised following fine-tuning or retraining runs
- Geographic distribution of training data sources and applicable privacy law jurisdictions
- Watermarking or membership inference resistance measures implemented to protect individual data subjects
None of the four major developers audited here publishes documentation covering all of these fields. Meta's LLaMA 2 documentation covers the first field partially. The others cover none of them in the specificity the fields require.
Toward a Provenance-First Model Documentation Standard
The solution to inconsistent disclosure is not voluntary best practice guidelines, which have already failed. It is a technically specified, legally grounded provenance-first documentation standard with mandatory fields and machine-readable outputs.
Several technical building blocks already exist. The Croissant metadata format, developed within the MLCommons working group, provides a structured schema for dataset documentation that can be referenced within model cards. The Data Nutrition Project, developed at MIT, proposed a "dataset nutrition label" concept that maps to many of the provenance fields described above. Neither has been adopted as mandatory infrastructure by any major developer.
The NIST AI Risk Management Framework (AI RMF 1.0, released by NIST and available through nist.gov) identifies measurement and training data documentation as components of responsible AI governance. The framework is voluntary. Its training data provisions have not been operationalized into the disclosure practices of the largest model developers.
What would change this? Enforcement. The EU AI Act's Article 53 documentation requirements, if enforced with specificity against general-purpose model providers, would create a binding disclosure floor. The FTC's authority over deceptive practices, applied to marketing claims about responsible AI development that are contradicted by opaque training data documentation, represents another enforcement pathway.
At Own Your Data Inc, the argument is simpler. Every training dataset is composed of artifacts produced by human beings. Those human beings are data subjects with legal rights and moral claims on how their outputs are used. A model card that does not document the composition of training data is not a transparency document. It is a performance of transparency, structured to satisfy an audience while protecting the information that audience most needs to evaluate.
The Mitchell et al. framework gave the field a precise and actionable standard in 2019. The field's largest actors have had years to implement it. In 2026, the audit result is clear: the gap between specification and practice is not closing. It is widening, scaled by the exponential growth in model capability and training data volume. Closing it requires treating provenance documentation not as a reputational exercise but as a technical and legal obligation owed to the individuals whose data built these systems.
