Provenance: where did this data come from?

Lesson 1 of 5 in Data Governance and Documentation Artifacts for AI.

Every AI governance question eventually becomes a data question. Is the model biased? — against whom was the data skewed? Can we honour a deletion request? — is that person in the training set? Did we have the right to train on this? — where did it come from, under what terms? Organisations that cannot answer where did this data come from and what are we allowed to do with it cannot answer anything downstream of it.

Two disciplines carry the answer. Provenance is the origin story of a dataset: source, collection method, legal basis, licence, consent scope, collection date. Lineage is its journey through your pipelines: every join, filter, transformation, label pass, and sampling decision between the source and the training run. Provenance without lineage tells you the ingredients but not the recipe; lineage without provenance documents a pipeline processing data of unknown right. High-stakes AI needs both, captured as metadata that travels with the data — not as tribal knowledge in a departed engineer’s head.

Key terms: data provenance, data lineage, training data, text and data mining exception, web scraping, chain of custody

Lineage with governance hooks: what gets captured at every hop

  1. Sources

    Captured here: origin (internal system, vendor, public web, partner), legal basis and consent scope, licence terms, collection dates, known gaps. This is provenance — everything downstream inherits it.

  2. Ingestion & combination

    Captured here: which sources were joined (dataset matching is an EDPB DPIA criterion), deduplication rules, records excluded and why. Combining datasets collected for different purposes is where purpose-limitation problems are born.

  3. Cleaning & filtering

    Captured here: filter criteria, outlier handling, imputation rules — and their demographic side-effects. “We removed incomplete records” can silently remove a population whose records are systematically incomplete.

  4. Labelling

    Captured here: who labelled (employees, vendors, crowdworkers), instructions given, inter-annotator agreement, spot-check results. Label quality is invisible in the data and decisive in the model.

  5. Train/test assembly

    Captured here: sampling strategy, class balancing, augmentation, synthetic additions (flagged as synthetic!), the exact dataset version hash the model will be trained on.

  6. Model training

    Captured here: which dataset version trained which model version — the link that makes deletion requests, bias investigations, and Annex IV documentation answerable.

  7. Outputs & feedback

    Captured here: whether production outputs and user feedback flow back into future training — the loop that quietly turns your customers into unconsented data subjects of the next model version.

Web-scraped data is where provenance meets enforcement. The cautionary tale is Clearview AI, which scraped billions of face images from the public web to build a biometric search engine — and collected GDPR fines of roughly €20 million each from France, Italy, and Greece, a €30.5 million fine from the Dutch DPA in 2024, and a US settlement with the ACLU under Illinois’ BIPA restricting its sales. The lesson regulators keep repeating: publicly accessible does not mean freely usable. Personal data on the open web is still personal data; biometric data still carries special-category protections.

The scraping-friendly precedent people cite — hiQ v. LinkedIn — is narrower than its reputation: the Ninth Circuit held that scraping public profiles likely does not violate the US anti-hacking statute (CFAA), but hiQ ultimately lost on breach-of-contract grounds and folded. In the EU, text-and-data-mining runs through the DSM Directive’s two TDM exceptions: Article 3 (research organisations — cannot be contracted away) and Article 4 (anyone — but rights-holders can reserve their rights in machine-readable form, and the EU AI Act’s Art 53(1)(c) obliges GPAI providers to maintain a copyright policy that honours those reservations, wherever in the world the training happened).

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.