Third-party model and dataset provenance

What it is

Provenance is the history of an artefact: who produced it, from what, under what licence, and what has been done to it since. Models and datasets are easy to obtain from public repositories and resellers, and their descriptions are often thin or unverified. An organisation that builds on an artefact of unknown origin inherits whatever is in it: unlawful or biased training data, restrictive licence terms, poor quality, or deliberate tampering. This entry concerns not knowing. Deliberate compromise is covered in LLM supply chain vulnerabilities.

How it happens

• A developer downloads a model from a public hub because it performs well.
• A dataset is bought from a broker who cannot say how it was collected.
• A supplier's product contains a third party's model that is not disclosed.
• Versions are updated in place, so the artefact in use is not the one first assessed.

Why it matters

Questions about lawfulness of training data, intellectual property and fitness for purpose cannot be answered for an artefact with no history. A model is harder to inspect than code, so the record of its origin carries more weight.

Warning signs

• Nobody can name the original producer of a model or dataset in use.
• No licence terms are on file.
• Artefacts are pulled from public sources at build time.
• The model card or datasheet is missing or blank.