Every AI-generated asset carries an invisible history: which references seeded it, which prompt shaped it, which model rendered it, which parameters tuned it, who revised it, who approved it, and what derivatives followed. Provenance is that history made explicit and durable. Without it, teams argue about facts that should be lookup queries.
The seven elements of complete provenance
- Source assets: reference images, style guides, prior versions, licensed inputs.
- Prompts: full text, negative prompts, seeds, and prompt revisions per version.
- Generation models: model names, versions, and tool versions at render time.
- Parameters: sampler, steps, guidance, aspect, and workflow-specific settings.
- Versions: explicit states showing what changed between each revision.
- Approvals: who signed off, when, and under what constraints.
- Derivatives: every variant, crop, adaptation, and downstream use.
Why provenance must be captured, not reconstructed
The practical rule: the system that renders the asset should record the provenance, automatically, as a side effect of creation. Human-entered provenance covers decisions and approvals; machine-captured provenance covers everything else.
Provenance in three jobs
Who consumes provenance
| Consumer | Question answered | Elements used |
|---|---|---|
| Production artist | How do I reproduce or extend this? | Prompts, models, parameters, references |
| Producer / reviewer | Which version did we approve and why? | Versions, decisions, approvals |
| Legal / client | What are we allowed to use, and can we prove it? | Sources, licenses, approvals, derivatives |
Implementation checklist
Provenance checklist
- List the seven elements and mark which your stack captures today.
- Automate machine-capturable elements (model, parameters, seed, timestamp).
- Require human-entered elements at decision points (approvals with rationale).
- Link derivatives to parents at creation, not in cleanup sprints.
- Test the record: can a stranger reproduce the asset from provenance alone?
Production provenance vs. content credentials
External authenticity metadata such as C2PA content credentials can verify an asset’s claimed source. Internal production provenance records how a team made, revised, and approved it. They answer different questions; read the C2PA vs. production provenance comparison for details.
What to measure in a paid pilot
Track the share of assets with complete seven-element provenance, reproduction success using provenance alone, and audit-response time for rights questions. Scope a production pilot around those measures.
Start with the workflow assessment to find where your team loses context today, read the Production Memory guide, and browse the use-case library for related workflows.
Key takeaways
- Provenance = sources + prompts + models + parameters + versions + approvals + derivatives.
- Capture provenance at creation; reconstruction after the fact is unreliable.
- Provenance serves production (reproduce), governance (audit), and rights (attribute).
- Distinguish internal production provenance from external authenticity signals like C2PA.