Data provenance explained for AI teams

Data provenance used to be a research concern. It is now a commercial and legal one, because AI teams have to show where their training data came from to regulators, rights holders and their own customers.

What is data provenance?

Data provenance is the record of a dataset's origin and history. It answers four questions: where the data came from, who created or collected it, under what rights it was obtained and what was done to it before it reached you.

A provenance record is evidence, not a description. A sentence saying the data is licensed for commercial use is a claim. A copy of the license, the date it was signed and the list of files it covers is provenance. See the glossary entry for data provenance for a short definition.

Why does provenance matter for AI training data?

Provenance matters for AI training data because training is hard to undo and the rules on what can be trained on are tightening. A dataset without provenance may be cheap to acquire and expensive to defend.

The scale of the problem is documented. The Data Provenance Initiative, a team of legal and machine learning researchers, audited more than 1,800 text datasets and reported in Nature Machine Intelligence in 2024 that popular hosting sites had license omission rates above 70% and error rates above 50%. The authors describe this as a crisis of misattribution.

Regulation has followed. Annex XI of the EU AI Act requires providers of general-purpose AI models to document information on training data, including its type and provenance, curation methods and how it was obtained and selected. Article 53(1)(d) separately requires a public summary of training content.

“information on the data used for training, testing and validation, where applicable, including the type and provenance of data and curation methodologies”

What does a provenance record contain?

A provenance record contains fields that describe the source, the rights and the processing history of a dataset. The Data and Trust Alliance proposed eight standards in November 2023, built by experts from member companies including IBM, Mastercard, Pfizer, UPS and Walmart, and they make a practical checklist.

The standards also call for a unique provenance metadata ID for each dataset. At launch, the Alliance said only generation date was consistently surfaced in existing metadata.

How is provenance different from lineage and chain of title?

Provenance is the overall record, while lineage and chain of title are two parts of it. Lineage tracks technical changes to the data. Chain of title tracks the legal transfer of rights.

The chain of title and data manifest glossary entries go into each term in more detail.

Dataset provenance is documented through a few widely used formats. None is mandatory, but buyers increasingly expect a seller to supply one of them.

• Datasheets for datasets: a question set proposed by Timnit Gebru and co-authors on arXiv in March 2018, covering why a dataset was created, what it contains, how it was collected and what it is suited for.
• Croissant: a metadata vocabulary from MLCommons built as an extension to schema.org, describing a dataset's attributes, provenance and usage restrictions. Its responsible AI extension adds a machine-readable summary of the attributes found in data cards.
• Data and Trust Alliance standards: the eight cross-industry fields in the table above, aimed at enterprise data rather than research datasets.
• Dataset cards: the documentation files that hosting platforms attach to each dataset, often based on the datasheet questions.

MLCommons says any dataset with Croissant metadata is discoverable through Google Dataset Search, which covers more than 700,000 datasets across Hugging Face, Kaggle, OpenML and other sites. See our dataset card entry for how cards relate to provenance.

How do you verify a seller's provenance claims?

You verify a seller's provenance claims by asking for the document behind each one and testing it against the data. A claim you cannot trace to a document should be treated as unverified.

Phyllo's September 2026 guide adds two checks that are easy to miss: rights documentation that traces to individual assets rather than a blanket assurance for the whole corpus, and per-layer clearance for complex content, such as the music, images and trademarks inside video.

What provenance should you keep after you buy?

After you buy, keep the provenance file for as long as any model trained on the data is in use. You may need it to answer a rights holder, a regulator or a customer's procurement team years after the purchase.

• The signed license and any amendments
• The data manifest and file hashes for each delivered version
• The seller's diligence file, including privacy notices and de-identification reports
• Your own processing log from receipt to training
• A list of the models and model versions trained on the data

If you provide a general-purpose AI model in the EU, this file is what lets you complete the Commission's training data summary template. Licensed datasets fall into a separate section of that template, with lighter public disclosure than scraped or public data. Our guide to the EU AI Act and training data transparency explains how.

How does provenance work for licensed company data?

Provenance for licensed company data is usually simpler to establish than for public or scraped data, because there is one owner and one set of source systems. The work is in documenting it properly before a sale.

Every seller we work with passes an eight-point vetting standard that includes ownership and chain of title, third-party content, de-identification method and manifest accuracy, before any buyer sees their data. You sign the license directly with the data owner, which keeps the rights trail short. Read the due diligence checklist for the full document list, or book a buyer consultation.