Skip to main content

The provenance file: your data map an auditor could follow

The short answer

The provenance file is the followable map

The deliverable of this topic is a single document that lets a stranger trace any output of your system back through every dataset and transformation to its origin, its legal basis, and its owner, without asking you a question. The Royal Free had the data and the good intentions; what it lacked was this map, and the map's absence is what the regulator found.

What you will be able to do

  • Assemble a data provenance file for a real AI system that names every dataset it uses and, for each one, records source, legal basis, original collection purpose, transformations, and a named owner.
  • Trace a single output of your model back through every dataset and transformation that produced it, so an auditor can walk the same path without your help.
  • Record the legal basis and the original collection purpose for each dataset in plain language, and flag where the use you now make of the data differs from the purpose it was collected for.
  • Document an honest "provenance unknown" entry for any dataset whose origin you cannot establish, rather than leaving a silent gap the auditor will find.
  • Compile the findings from Topics 2.1 through 2.5 into the corresponding fields of the file, so the module's work becomes one inspectable artifact rather than five loose notes.
  • Construct the file so a reader who has never met you or your system can follow the data trail end to end, applying the same plain-language glance test you would to any learner-facing document.
  • Assign a named owner and a review date to every dataset entry, so no data source sits in the file unowned and no entry silently goes stale.
  • Connect each provenance entry to the downstream artifact that will consume it, naming the conformity file (Module 5) and the evidence annex (Module 10) as the places this file comes back.
  • Distinguish what a provenance file is responsible for (the followable data map) from what belongs to the data protection impact assessment (Module 10), handing the adjacent material to its owner topic.
  • Proportion the depth of each entry to the stakes of the decision the data feeds and the sensitivity of the data, so effort concentrates where a wrong or unlawful output would harm people most, while every dataset still clears the seven-question floor.
  • Adapt the legal-basis field to the jurisdiction that governs each dataset, flagging where a basis is established under one country's law but still open under another's that the data or system also touches.

The ideas, one by one

One entry per dataset, the same seven questions

Name, source, legal basis, original collection purpose, transformations and lineage, retention and destruction, and owner with a review date. A vague paragraph names nothing an auditor can check; a dense pipeline diagram answers the engineering question and not the governance one. The structured entry is what a stranger can walk.

Legal basis and original purpose are where files sink

These two fields force the questions the Royal Free never answered on the record: are we lawfully allowed to use this data, and is the use we now make it of the use it was collected for. A blank legal-basis field on data already in production is the single finding most likely to become an enforcement action.

Lineage makes the trail followable forward

Recording what happened to the data from source to output is how you trace a bad output to a bad input, prove a deletion reached every downstream copy, and catch synthetic data laundering a bias upstream. A lineage gap is exactly where a harmful output later proves untraceable.

"Provenance unknown" is an entry, not an omission

Every real organization has datasets with incomplete provenance. The honest entry says so, names what is known and unknown, and assigns an owner and a date to close the gap. An auditor who finds a flagged gap confirms your honesty; one who finds a hidden or guessed-at gap doubts your whole file.

The file is mostly assembly, if you did the module honestly

Topics 2.1 to 2.5 already produced the audit, the consent history, the retention decision, the synthetic-data call, and the vendor verification. This topic places each finding into its field. An empty field signals an earlier investigation you have not finished.

The file is living, owned, and dated

Data changes, so the file changes with it; a stale provenance file describing datasets you no longer use is trusted and wrong. Every entry carries a named owner and a review date, and a change to the data is not complete until the file reflects it. The named owner is the structural answer to "who was answerable for this?"

It is the map, not the risk assessment

The provenance file records where the data came from; the data protection impact assessment (Module 10) assesses what could go wrong for people. The file feeds the assessment; you cannot assess the risk of data you cannot trace. Build the map here; the assessment reads it later.

It grounds everything after

The provenance file is lifted into the conformity file in Module 5 as its data-governance evidence and linked into the evidence annex in Module 10; when your file is attacked in Module 11, the provenance entries an adversary probes first are the gaps and the guesses. Module 2 exists so the rest of the program stands on data you can actually account for.

Provenance travels; the basis field must too

The file is jurisdiction-neutral, but the legal-basis field is not: data collected in one country, processed in a second, and deployed in a third can face three rulebooks. Build the basis field to hold more than one answer where the data crosses borders, and flag the jurisdictions where a basis is established and where it is still open. The cross-border shipping decision later in the program rests on exactly this.

Depth is proportioned to risk

Every dataset gets an entry answering the seven questions, but the sensitive data feeding a high-stakes decision earns the deepest, fully evidenced entry, while low-risk data earns a lighter one. Proportion keeps the file finite and honest, and recording the proportion reasoning stops an auditor mistaking a deliberately light entry for a neglected one.

Provenance is operational, not just compliance

The file is what lets you act in minutes when a public dataset is found poisoned, a vendor is found unlawful, or a person demands deletion. An organization without it meets every data incident as an unanswerable panic. The file is for the incident that will come, not only the auditor who might.

Listen to it

The same lesson, talked through by two hosts as episode 16 of the podcast.

Real cases

These are real, documented practices and events that show the discipline of a followable data map, done well and done badly. The Royal Free and DeepMind case anchors the immersive scenario in Section 5; these show the wider pattern across jurisdictions.

Example 1: The GDPR record of processing activities (Article 30, European Union). The GDPR requires most organizations to maintain a written record of their processing activities: what personal data they hold, why, on what basis, who it is shared with, and how long it is kept (GDPR, Article 30; established). This is the legal ancestor of the provenance file. The lesson for you is that "write down where your data comes from and why you are allowed to use it" is not a novel governance idea invented for AI; it is an existing legal obligation that AI systems make sharper, because AI reuses data at a scale and for purposes the original collection rarely contemplated. A provenance file is a Article-30 record built to be walked, not just filed.

Example 2: The EU AI Act's data-governance requirement (Article 10). For high-risk AI systems, the EU AI Act requires providers to apply data-governance practices to training, validation, and testing data, including documenting the data's origin and, for personal data, the original purpose of collection (Regulation (EU) 2024/1689, Article 10; established). The same article also requires the data to meet quality criteria, relevant, sufficiently representative, and, to the best extent possible, free of errors and complete for the system's intended purpose (Article 10(3); established), and permits processing special-category data, under safeguards, specifically to detect and correct bias in a high-risk system (Article 10(5); established). This moves the provenance file from good practice toward legal requirement for a whole class of systems. The point marked established: the fields this topic teaches, source, legal basis, original purpose, are becoming the fields a regulator can demand, and increasingly the quality and bias evidence behind them too. The deep treatment of high-risk conformity, including these quality obligations, belongs to Module 5 (see Topic 5.6); the point here is that the data map you build is increasingly the thing the law expects to see.

Example 3: Datasheets for datasets (Gebru et al., 2018 onward). A research proposal, "Datasheets for Datasets," argued that every dataset should ship with a document stating how it was collected, what it represents, its known gaps and biases, and its intended and unintended uses. The proposal has been widely adopted in AI research and by some data providers. The relevance is direct: a datasheet is the provenance entry for a single dataset, written by the people who made it. Where a dataset you use came with a datasheet, it populates your provenance entry; where it did not, the datasheet's absence is itself a provenance gap you record. The discipline of documenting a dataset's origin runs from the people who create data to the people who deploy it, and your file is the deployment end of it.

Example 4: The LAION-5B dataset takedown (2023). Researchers at the Stanford Internet Observatory found that the LAION-5B image dataset, used to train widely deployed image generators, contained thousands of links to suspected child sexual abuse material, and the dataset was withdrawn. The deep treatment of this case, as quiet poison in a dataset's provenance, belongs to Topic 2.1 (see Topic 2.1); the point here is what it shows about provenance files. An organization that had trained on LAION-5B and kept an honest provenance entry could, when the finding broke, immediately see that this dataset fed their model and act. An organization with no provenance file could not even answer whether they were affected. Provenance is what turns a third-party data scandal from an unanswerable panic into a known, bounded task.

Example 5: Clearview AI and the provenance of scraped data (multiple regulators). Clearview AI built a facial-recognition database by scraping billions of images from the public web, and multiple data-protection regulators across Europe and Australia found the practice unlawful for lack of a valid basis and notice to the people whose faces were taken. The deep case is owned elsewhere; here it illustrates the provenance question at the collection end. A provenance file that honestly recorded "source: scraped from public websites; legal basis: none established; the people were not told" would have flagged the problem before deployment, not after enforcement. The file does not make an unlawful source lawful; it makes the unlawfulness visible early enough to stop, which is the whole point of writing provenance down before you build.

Example 6: Content provenance standards (C2PA, emerging). The Coalition for Content Provenance and Authenticity (C2PA) publishes an open standard for attaching tamper-evident provenance metadata to media, so that an image or video carries a signed record of where it came from and how it was edited. This is provenance at the level of a single artifact rather than a dataset, and it is being adopted by major camera makers, software vendors, and platforms. Marked emerging: standards and tooling for automated provenance are advancing fast. The lesson for your file is that provenance is becoming machine-readable and portable, but the governance judgment, is this basis valid, is this purpose compatible, is this gap acceptable, still requires the human owner your file assigns.

Example 7: ISO/IEC 42001 and data governance in a management system (2023, established). The first international AI management-system standard, ISO/IEC 42001:2023, requires organizations to establish controls over the data used in their AI systems, including understanding its sources and appropriateness for the intended purpose. The deep treatment of the standard as an operating system for governance belongs to Module 6 (see Topic 6.3); the point here is that a certifiable international standard, not just a regional law, expects an organization to be able to account for its AI data. A provenance file is a large part of how an organization demonstrates that control in practice, which is why building it well pays off across legal, standards, and audit demands at once rather than satisfying only one.

Example 8: Records that enabled a fast recall (the good pattern, generalized). The contrast case to the Royal Free is any organization that, when a widely used dataset or model component is found compromised, can answer within hours exactly which of its systems are affected because its provenance records name every place that component flows. This is the everyday, unglamorous payoff of provenance, and it is why data-intensive fields that long predate AI, pharmaceuticals with lot tracing, food supply with traceability rules, aviation with parts provenance, all maintain the equivalent of a provenance file. Marked established as a cross-industry practice: the discipline of tracing what went into a product so you can recall it precisely is mature everywhere the product can harm people, and AI data is now in that category.

Example 9: The cost of not having one (the Royal Free pattern, generalized). The Royal Free and DeepMind case (ICO, 3 July 2017) is the anchor of this topic's scenario, so it is treated in depth in Section 5. Generalized, it is the counter-example to all of the above: a well-intentioned project with a real clinical benefit, undone not by a bad model but by the absence of a followable data map. When the auditor asked the provenance questions, source, legal basis, original purpose, notice to the people, the answers were not written down, so the finding was unlawfulness. Every example above is a version of the same claim: the organizations that survive a data audit are the ones that could walk the auditor through their data before the auditor arrived.

Where people go wrong

  • "Our data is internal, so it does not need provenance." The most common and most dangerous belief. "Internal" is not a source; it is a category that hides the source. Internal data was still collected from someone, for some purpose, under some basis, and reused for a new one. The Royal Free data was "internal" NHS data, and it was still unlawful to hand over without a basis and notice. Provenance is not about foreign or purchased data; it is about every dataset, especially the ones you feel you already own.
  • "The legal-basis field is a job for the lawyers, not for me." Lawyers determine the basis; you record it. The mistake is leaving the field blank until "legal gets to it," which means the data is already in the model before anyone confirmed you were allowed to use it. The provenance file's job is to make the basis question unavoidable at build time, so it is answered before training, not after an inquiry. An empty legal-basis field on a dataset already in production is the exact Royal Free failure.
  • "We can reconstruct the provenance later if anyone asks." You cannot, and this is the trap that catches organizations after a departure or an acquisition. Provenance lives in the memory of the people who handled the data, and those people leave. The transformation nobody wrote down becomes unrecoverable the day the engineer who did it resigns. Provenance is recorded as the data flows, not reconstructed under audit pressure, because by audit time the trail has gone cold.
  • "A pipeline diagram is a provenance file." A data-flow diagram shows how data moves; it does not show where the data came from, on what basis, or for what original purpose. A diagram can be a useful lineage aid, but a file that has boxes and arrows and no legal-basis or original-purpose fields answers the engineering question and not the governance one. The auditor's first question is not "how does the data flow" but "were you allowed to have it," and a diagram is silent on that.
  • "Provenance unknown means we leave it out." Backwards. A dataset with unknown provenance is the most important entry in the file, because it is the one that will surprise you. Leaving it out does not make the risk disappear; it makes the risk invisible until it causes harm. An honest "provenance incomplete, here is what we know and who is chasing the rest" protects your credibility and turns the unknown into an owned task. Hiding it is the failure the file exists to prevent.
  • "The provenance file is the same as the DPIA." They are different documents with different jobs. The provenance file is the followable data map: where the data came from and what happened to it. The data protection impact assessment (DPIA) is the risk assessment that reads the map and asks what could go wrong for people, and it is developed in Module 10 (see Topic 10.4). The provenance file feeds the DPIA; you cannot assess the risk of data you cannot trace. Building the map and assessing the risk are two steps, and this topic is the first.
  • "Once the file is written, we are done." The file is living, not final. Data changes, vendors change, retention periods lapse, models are retrained. A provenance file with an old date describing datasets you no longer use, or missing ones you added, is trusted and wrong, which is worse than no file. Every dataset entry carries an owner and a review date, and a change to the data is not complete until the file reflects it. Evan adds a line the same evening his file is tested; that is the habit.
  • "We only need provenance for the training data." Every dataset that touches the system needs an entry: training, validation, test, the live user inputs the model sees in production, the reference data it retrieves, the third-party feeds it joins. A model can be harmed or made unlawful by any of them. Scoping provenance to the training set alone leaves the auditor a trail that stops halfway, exactly where a real problem often lives.
  • "If the vendor says the data is clean, we record it as clean." No. You record what you verified, not what you were told, which is the whole lesson of Topic 2.5 (see Topic 2.5). A provenance entry that says "basis: confirmed" on a vendor's say-so is a fabrication waiting to be exposed. The honest entry records what the vendor disclosed, what they refused to disclose, and your own verification status. A vendor claim is an input to the entry, never the entry itself.
  • "A good file makes us look like we do not trust our own data." The opposite is true, and it is the same misconception that made people fear the failure account in Topic 1.6. A specific, honest provenance file raises the confidence of the people whose confidence matters, the regulator, the board, the successor, because it proves you know exactly what your data is and where it came from. The organization that cannot produce a provenance file is the one to distrust, because it is governing data it cannot describe.
  • "The rare or old dataset is not worth an entry." The dataset you are tempted to skip, the small legacy table, the one-off licensed batch from years ago, is often the one with the weakest provenance and the highest surprise risk, precisely because it is old and half-forgotten. Skipping it to keep the file short is how the untraceable input hides. Every dataset that feeds the system earns an entry; the ones you least want to write are usually the ones the auditor most wants to see.
  • "Provenance is a compliance chore with no operational value." Provenance is what lets you act fast when a data problem breaks. When a public dataset is found to be poisoned, when a vendor is found unlawful, when a person demands deletion, the provenance file is what tells you, in minutes, whether and where you are affected and what to do. An organization without it faces every data incident as an unanswerable panic. The file is not only for the auditor who might come; it is for the incident that will.
  • "One jurisdiction's rules are enough because that is where we are based." Data does not respect your headquarters. A dataset collected in one country, processed in a second, and used in a system deployed to a third can face three different rulebooks, and a legal-basis field that answers only your home law leaves the other two open. The honest entry names which jurisdiction's basis it is relying on and flags where a basis is established in one place but not another. The cross-border shipping decision later in the program rests directly on this (see Topic 6.6); a file that assumes home law travels is the file that turns a launch abroad into a multi-regulator problem.
  • "The systems inventory and the provenance file are the same thing." They answer different questions. The AI systems inventory from Topic 0.2 lists what AI you have and where it runs (see Topic 0.2); the provenance file describes the data behind one of those systems in depth. The inventory tells you which systems need a provenance file; the file is the data map for each. Confusing them leads to a shallow list that names systems but never traces their data, which fails the auditor's walk at the first output.

Questions people ask

What is data provenance file?
The structured, plain-language document produced in this topic that records, for every dataset feeding an AI system, its source, legal basis, original collection purpose, transformations and lineage, retention and destruction, and named owner, plus a statement of the system it feeds and an honest list of datasets whose provenance is incomplete. It is the followable data map an auditor could walk, and it is consumed by the conformity file (Module 5) and the evidence annex (Module 10).
What is provenance?
The origin and history of a thing: where it came from and what happened to it along the way. Borrowed from the art world, where a work's provenance is the documented chain of ownership that proves it is genuine. Applied to data, it is the documented chain from source to model input. More on Provenance
What is dataset (for the file)?
Any distinct body of data that feeds the system, including training, validation, and test sets, live production inputs, retrieved reference data, licensed corpora, purchased batches, and synthetic sets. If it touches the model or the decision, it earns a provenance entry.
What is legal basis?
The specific lawful ground on which an organization holds and uses a dataset. For personal data under the General Data Protection Regulation (GDPR), one of the grounds in Article 6 (such as consent, contract, legal obligation, vital interests, public task, or legitimate interests), plus, for special categories like health data, an additional condition in Article 9. A blank legal-basis field is the finding most likely to become an enforcement action.
What is original collection purpose?
What a dataset was gathered for in the first place, which is often not the purpose for which it is later wanted. Under the GDPR's purpose-limitation principle (Article 5(1)(b)), personal data collected for one purpose may not simply be reused for an incompatible new one without a fresh basis. The provenance file's job is to make any collection-versus-use gap visible so it can be resolved.

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →