JwumaWhat data provenance actually means, and how to verify it
· 3 min read · provenance, procurement
Provenance means being able to show where each item in a dataset came from, who produced it and under what consent. Real provenance is artefacts, not assurances: preserved capture metadata, recorded consent, per-item contributor attribution and the rubric version that produced each label.
The short version
- Provenance is per-item, not per-dataset.
- Preserved EXIF is the primary provenance signal for captured imagery.
- Consent records should be attached to items, not held as a blanket statement.
- Every label should carry the rubric version that produced it.
- A supplier who cannot produce per-item provenance is asserting, not evidencing.
Scoping something like this?
We would rather talk about the problem than quote against a volume figure.
See capabilitiesWhy is dataset-level provenance not enough?
Because the question you will eventually be asked is about a specific item, not the dataset. When a rights holder or a regulator queries one image, a statement that the collection was ethically sourced answers nothing.
Per-item provenance means each row can be traced to who produced it, when, under what consent and against which rubric version. That is the only form that survives a challenge.
What artefacts actually constitute provenance?
For captured imagery: intact EXIF showing capture time and device, which is what distinguishes original capture from a scraped file. For speech: the recording conditions and the consent record for the speaker. For labels: the annotator, the timestamp and the frozen rubric version applied.
Ask to see these for a sample of specific rows during evaluation. A supplier who can produce them for the rows you pick has real provenance; a supplier who produces a policy document does not.
- Intact capture metadata on imagery
- Per-item consent records where people appear
- Contributor attribution per item
- Rubric version stamped on every label
- Timestamps that survive delivery
Why does capture metadata get destroyed so often?
Because most collection pipelines route images through galleries, messaging apps or resizing steps that strip or rewrite EXIF. By the time the file reaches the buyer the provenance is gone and cannot be reconstructed.
This is why in-app capture matters operationally rather than as a feature. If the contributor uploads from their gallery, the metadata is already lost regardless of what the contract says.
What should the contract require?
That provenance artefacts are delivered with the data rather than retained by the supplier, and that they survive any format conversion in the delivery pipeline. Provenance held in a supplier's internal system is provenance you lose when the relationship ends.
Also require the right to audit a sample of rows against source records, and use it at least once early rather than saving it for a dispute.
People also ask
What is data provenance in AI training data?
The ability to show, for each item, where it came from, who produced it, under what consent and against which labelling standard. It is a per-item property rather than a dataset-level claim.
How can I check a supplier's provenance claims?
Pick specific rows during evaluation and ask for their provenance artefacts: capture metadata, consent record, contributor attribution and rubric version. A policy document is not an answer.
Why does EXIF metadata matter for image datasets?
It records when and on what device an image was captured, which is the primary evidence distinguishing original field capture from scraped material. Most pipelines strip it unless capture happens in-app.