The shortest useful definition
The unit of an AI production log is not the chat session and not the project folder. It is the deliverable or shot. Every final image, clip, voice track, or composite receives a record that identifies the exact output and preserves the decisions needed to review or hand it over.
That distinction matters because production evidence is distributed. A generator may keep the prompt, a compositor keeps the edit history, a producer keeps the approval email, and a delivery folder contains the final export. The log joins those facts without pretending that any one system witnessed the entire creative process.
The minimum fields
- Asset identifier and a hash or other stable reference to the delivered file.
- Generator, model name, model version where available, and generation date.
- Prompt, negative prompt, seed, aspect ratio, duration, and other material settings that the tool exposes.
- Input images, footage, audio, control signals, adapters, and their source or rights status.
- Selected output, rejected variants when material, and the reason the final take was chosen.
- Human edits after generation, including retouching, compositing, grading, sound work, and cropping.
- Disclosure or labeling decision for the place where the work will appear.
- Reviewer, approval state, approval date, and the exact version that was approved.
What a production log can prove
| Claim | Supportable? | Why |
|---|---|---|
| These inputs were declared for this asset | Yes | The record preserves the declaration and its asset link. |
| This record existed by a stated time | Yes, with a trustworthy timestamp | The timestamp supports existence at or before that moment. |
| The record has not changed | Yes, with integrity protection | Hashes or signed manifests expose later modification. |
| This reference caused these pixels | No | A production record observes declared inputs, not model causality. |
| The work is automatically legally compliant | No | Compliance depends on context, disclosure, rights, and human decisions. |
How it differs from C2PA and generator history
C2PA is a technical standard for signed provenance assertions attached to content. It can carry origin and edit information in a tamper-evident manifest. A production log is broader operational evidence: it includes approvals, rights decisions, rejected takes, and records held outside the media file. The two layers complement one another.
Generator history is useful raw material but remains tied to an account, product interface, and retention policy. A production log keeps the fields the team needs in a neutral record connected to the delivered version.
When to create the record
- Create the asset record when the first production-worthy generation is made.
- Attach prompts, settings, and references before leaving the generator session.
- Update the record when a take is selected or materially edited.
- Bind approval to the exact export rather than to a filename that may be overwritten.
- Close the record with the delivery date, disclosure decision, and delivered-file identifier.
Frequently asked questions
Is an AI production log required by the EU AI Act?
The EU AI Act does not prescribe a document called an AI production log. Article 50 creates transparency and disclosure duties for certain generated or manipulated content. A production log is an operational way to preserve the facts a team needs when deciding and demonstrating how those duties were handled.
Is a spreadsheet enough?
A spreadsheet can hold the necessary fields, but ordinary spreadsheets are weak at file identity, attachments, access history, and tamper evidence. The useful test is whether the record stays linked to the exact delivered asset and preserves who changed or approved what.
Does the log need every failed generation?
Not always. Preserve the selected output, its material inputs and settings, and any rejected variant that affected a rights, safety, or approval decision. Recording thousands of irrelevant drafts can make the important evidence harder to find.