Provenance
Provenance is the history of an artifact: where it came from, how it changed, who controlled it, and what rights or constraints travel with it. For AI, provenance applies to datasets, labels, model weights, prompts, embeddings, retrieved documents, evaluation sets, and generated outputs.
Without provenance, teams inherit hidden assumptions. A dataset may contain outdated policy, copied personal information, low-quality labels, or material that should be removed when a license changes.
Repair the Chain of Custody
A support-chat dataset is being reused to train a sales assistant. Toggle which provenance links this dataset actually has, and watch the chain of custody hold together — or break.
Cannot tell whether cleaning, filtering, or labeling introduced errors or bias.
No way to know how much of the data is low-quality, duplicated, or mislabeled.
A removal request cannot be honored because the data's location isn't tracked.
Chain integrity
2 / 5 links intact
A provenance record should identify the source, collection date, collection purpose, permission or license, transformation steps, quality checks, and retention or deletion process. For derived assets, it should also say which upstream artifacts were used.
Provenance becomes especially important when data is reused. A chat transcript collected for support quality may not be appropriate for training a sales assistant. The chain of custody helps teams decide what reuse is allowed.
Derived Data Still Has Lineage
Embeddings, summaries, labels, and synthetic examples can still reveal or depend on the original source. Treat derived artifacts as part of the provenance chain.
A model was trained on 'public web data' but the team cannot name sources, licenses, or removal paths. What risk does that create?