Rights and contracts
Creative asset provenance exclusions: separating owned, client, stock and AI work
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Separate creative assets by origin and governing document, one component at a time: agency-owned originals, client deliverables, retained materials, freelancer work, stock, client-supplied material and AI outputs. Attach the contract or license to each row, treat silence as no permission, and let a composite take the outcome of its most restrictive part. Agency process records often survive when finished creative cannot.
Provenance trail
Creative asset provenance
Key takeaways
- Classify components, not folders: one layered file can hold four rights classes.
- Every row needs its governing document; a missing document holds the row.
- Stock licenses commonly prohibit AI use, so stock layers exclude the composite unless they can be removed.
- AI-generated and AI-edited material gets its own label and a separate decision.
- Brief versions, revision rounds and approvals are often the most licensable part of a creative archive.
How do you separate owned creative work, client work, stock assets and AI outputs?
Classify every file and every process record by where it came from and which signed document governs it, then exclude anything whose governing document is missing, silent on AI use or restrictive. Creative asset provenance exclusions work at the component level: a finished banner can hold an agency-drawn layout, a client logo, a stock photograph and an AI-extended background, and each part follows a different rule.
The practical result is usually a narrower but cleaner set than teams expect. Many finished deliverables drop out because one layer is stock or client-supplied, while the agency's own process records, such as brief versions, revision rounds and approval decisions, often survive once client identities are masked and the client contract allows it. The worksheet below makes that sort explicit and leaves a dated trail a reviewer can check.
Why a client folder in the DAM is not a rights record
Folders record who a job was for, not who owns each piece inside it. A folder named after a client mixes deliverables the client owns, templates the agency kept, stock it licensed for that one campaign, and files a freelancer delivered under a separate agreement. Treating the folder as one rights class is the most common way an exclusion list goes wrong.
Composition is normal in creative files, and provenance standards model it that way. The C2PA specification describes ingredients with a relationship field that must be parentOf, componentOf or inputTo, so a single output can record a parent file it was edited from, components placed into it and inputs fed to a generative model. A layered design file has the same structure whether or not anyone signed a manifest.
Exports make it worse. Flattened JPEGs and PDFs often lose the layer names and embedded metadata that showed a component was stock or generated, so the classification has to be done on the working files, or on the job record that lists what went into them, before the export set is built.
Which document decides each class of creative asset?
Each class is decided by a document, and the document decides the default. If you cannot attach the document to the row, the row is held, not guessed. The table lists the seven classes most creative teams meet and the default treatment when the records may be licensed for AI use.
| Class | Typical examples | Governing document | Default for AI licensing | Evidence to attach |
|---|---|---|---|---|
| Agency-owned original | Self-promotion work, internal templates, unused pitch concepts kept by the agency | Employment terms, client contract clauses on pitch work | Candidate, after confidentiality check | Creator, date, contract clause that leaves it with the agency |
| Client deliverable | Final ads, brand systems, packaging delivered under a statement of work | Master services agreement and statement of work | Exclude unless the client grants AI use in writing | Ownership clause, confidentiality clause, any written client permission |
| Agency retained materials | Tools, presets, frameworks licensed to the client but kept by the agency | Retained or pre-existing materials clause | Candidate if the clause keeps ownership and imposes no use limit | Clause text and a list of the retained items |
| Freelancer or contractor work | Illustration, photography, retouching bought from outside creators | Contractor agreement, assignment or license | Hold until the agreement shows ownership or AI use rights | Signed agreement and invoice that ties it to the file |
| Third-party stock | Licensed photos, footage, fonts, music beds placed in layouts | Stock provider license terms in force at download | Exclude the component, and the composite unless it can be removed | License ID, download date, license version |
| Client-supplied material | Logos, product shots, customer photos sent by the client | Client contract and the client's own sources | Exclude | Intake record showing it came from the client |
| AI-generated or AI-edited output | Generated backgrounds, inpainted fixes, generated copy variants | Generator terms, internal AI policy | Hold, label and decide separately | Tool, date, prompt log reference, digital source type |
Freelancers and commissioned work: ownership depends on the paperwork
Freelance work is the row most likely to be wrong, because ownership does not follow payment. The U.S. Copyright Office explains in Circular 30 that a work made for hire is either made by an employee within the scope of employment or specially ordered or commissioned, and the commissioned route applies only when the work falls in one of nine statutory categories and both parties sign a written agreement saying it is a work made for hire.
Many design and photography jobs fall outside those nine categories, so the agency's rights in commissioned work usually come from an assignment clause or a license instead. That changes what the worksheet records: the clause itself, whether it covers all media and uses, and whether a license is limited to the client campaign. A license scoped to one campaign does not reach AI training, even if it is exclusive.
Ownership is also only the first gate. An agency can own a concept outright and still be bound by a confidentiality clause in the client contract that bars disclosing anything made for that client. The worksheet keeps ownership and permission as separate columns so one cannot quietly stand in for the other.
Stock components: the license usually rules out AI use
Assume a stock component cannot go into an AI dataset unless the license says it can. Getty Images' content license agreement, as published on its site when checked on October 18, 2026, prohibits using licensed content and its metadata for machine learning or artificial intelligence purposes unless a Getty Images invoice, sales order confirmation or license agreement explicitly allows it. Adobe's Stock product terms dated January 2026 bar using Stock assets to create, train, test or otherwise improve machine learning or AI systems.
Read the version in force when the asset was downloaded, not today's page, and keep the license ID with the row. If the stock element sits on its own layer, the agency may be able to remove it and keep its own layout, but only if what remains is still useful and no longer reproduces the stock image. If the stock element is baked into a flattened file, exclude the composite.
Embedded signals help but do not decide. The IPTC Photo Metadata Standard 2023.1 added a Data Mining property, with values including allowed and prohibited for AI or ML training. A prohibition value is a reason to exclude. A missing or unspecified value is not permission, so it sends the row back to the license.
AI outputs: label the source before anyone decides
AI-generated and AI-edited files get their own label before anyone decides whether to include them. The IPTC Digital Source Type vocabulary distinguishes media created using generative AI (trainedAlgorithmicMedia) from media edited using generative AI, such as inpainting or outpainting (compositeWithTrainedAlgorithmicMedia), and C2PA manifests use the same IPTC terms in the digitalSourceType field of a created action.
The label matters for rights as well as quality. The Copyright Office's registration guidance says works whose traditional elements of authorship were produced by a machine lack human authorship, and asks applicants to disclose and exclude AI-generated material that is more than de minimis. An agency may therefore hold weaker rights in a generated background than in the hand-drawn layout around it.
Record the tool, the date, a pointer to the prompt or session log, and which layers were generated. A buyer may want generated material excluded, kept separate or flagged, and that choice can only be made if the flag exists. Do not infer generation from appearance; if the job record is silent, mark the source as unknown.
The rights classification and exclusion worksheet
The worksheet has one row per component, plus one row per process record such as a brief version or approval note. Each field has a fixed set of values so two reviewers reach the same answer from the same evidence. Fill it from the working files and the job record, then let the decision rules below set the outcome.
At a glance
The rights classification and exclusion worksheet
- Rule 1: a composite takes the outcome of its most restrictive component, unless that component is removed and the remainder re-classified.
- Rule 2: silence is not permission; a silent license or contract is recorded as not granted.
- Rule 3: a process record about a client job is kept only when the client contract allows that disclosure and client identities and creative are masked.
- Rule 4: unknown origin is its own value; it holds the row and is counted, never folded into owned.
- Rule 5: re-run the affected rows when a contract, release or license version changes, and keep the earlier decision with its date.
| Field | What to record | Allowed values | Rule it feeds |
|---|---|---|---|
| Item ID and type | File, layer or process record, with its job number | finished asset, component, brief, revision note, approval | Process records are classified separately from the assets they describe |
| Origin class | One of the seven classes in the table above | owned, client, retained, freelancer, stock, client-supplied, AI | Sets the default before any document is read |
| Governing document | Contract, license or policy reference with clause number | document ID, or missing | Missing sends the row to hold |
| Ownership | Who owns the item under that document | agency, client, creator, licensor, unclear | Unclear sends the row to hold |
| AI use permission | What the document says about machine learning or AI use | granted, prohibited, silent | Prohibited excludes; silent is treated as not granted |
| Embedded signals | IPTC Data Mining value and digital source type, if present | value as found, or none | A prohibition excludes; none changes nothing |
| Components | Item IDs of every layer or placed asset inside it | list of item IDs | The most restrictive component sets the composite's outcome |
| Restrictions | Confidentiality, territory, term, talent or property release limits | none, or the limit and its end date | Any active limit that covers AI use or disclosure excludes |
| Decision | Outcome and the rule that produced it | include, include process record only, exclude, hold | Hold rows never ship |
| Reviewer and date | Who decided, when, and the evidence link | name, date, link | Rows without a reviewer are not final |
One campaign folder, five rows (illustrative)
Here is how the worksheet resolves a single campaign folder. The job, files and counts are illustrative, not drawn from a real agency. The folder held a final key visual, its working file, three revision rounds of comments, and a client approval note.
In this illustrative case, three of the five items end in exclude or hold, and only the agency's own process records move forward. That is a typical shape, not a disappointing one: the decision history is often the part an AI developer cannot get elsewhere.
| Item | Class found | What the document says | Decision | Reason |
|---|---|---|---|---|
| Final key visual (flattened) | Client deliverable with stock component | Client owns deliverables; stock license prohibits AI use | Exclude | Most restrictive component, cannot be separated |
| Working file, agency layout layer | Agency retained material | Retained materials clause keeps templates with the agency | Hold | Clause silent on use outside the client relationship |
| Extended background layer | AI-edited output | No prompt log kept for this job | Hold | Generation tool and date unknown |
| Three revision rounds of comments | Agency process record | Contract allows anonymized internal records | Include process record only | Client names and creative masked |
| Client approval note | Agency process record | Same clause | Include process record only | Approval timing and decision kept, quotes removed |
Revisions, expiring releases and the unknown column
Decisions change when the paperwork changes, so the worksheet records dates, not only outcomes. A talent or property release that limits use to a territory or term, a stock license replaced by a newer version, or a client who later grants written permission for anonymized use can each move rows in either direction.
Revisions can change rights too. A concept the client rejected may stay with the agency under some contracts and pass to the client under others, so the decision for an unused round depends on that clause, not on whether the work ran. Approval records show which version the client accepted, which is often the only way to know which file is the deliverable.
Show uncertainty as a number rather than hiding it. Report how many rows are hold or unknown, by class, next to the include count. If the held share is large, the honest next step is to recover documents, not to relabel rows, and a rights reviewer should look at any row where the decision turns on reading a clause.
Carrying the worksheet into a licensing review
A completed worksheet is the evidence a licensing review asks for: what is offered, what was removed and why. An agency or studio licensing its process records through SourceX can hand over the worksheet as the starting point for the rights review, contracting, preparation and delivery SourceX runs. Delivery waits for the company's own sign-off on scope, terms and the agreement, and qualifying begins with a short assessment that needs no file uploads.
Frequently asked questions
What should a rights classification and exclusion worksheet include?
One row per component or process record, with its origin class, governing document, owner, AI use permission, embedded metadata signals, components, restrictions, decision and reviewer. Each field should have fixed values so two reviewers reach the same outcome. The decision should cite the rule that produced it.
If the agency owns a design, can it license it for AI training?
Not automatically. Ownership is one gate; confidentiality clauses in the client contract, stock components inside the design and talent or property releases are others. Each needs to be checked and recorded separately, and counsel should review any clause that is unclear.
Can a layout stay in the set if the stock photo is removed?
Sometimes. If the stock element is on its own layer and the remaining layout is the agency's work and still useful, the remainder can be re-classified as a new row. A flattened file where the stock image cannot be separated should be excluded.
Does an IPTC Data Mining value of allowed mean the image can be used?
It is a useful signal from whoever embedded it, but it does not replace the license. Check that the value was set by the rights holder and matches the license terms. A prohibited value is a reason to exclude, and a missing value says nothing.
How should AI-generated files in an archive be handled?
Label them first, using the tool, date, prompt log reference and a digital source type such as created or edited using generative AI. Keep them in their own class and decide them separately. If the job record does not show whether a file was generated, mark it unknown and hold it.
Sources
- A work made for hire is made by an employee within the scope of employment, or is specially commissioned in one of nine categories under a written agreement signed by both parties. Source
- The Copyright Office will not register works whose traditional elements of authorship were produced by a machine, and asks applicants to disclose and exclude AI-generated material that is more than de minimis. Source
- C2PA ingredient assertions declare a parentOf, inputTo or componentOf relationship, and created actions can carry an IPTC digital source type such as trainedAlgorithmicMedia. Source
- IPTC Photo Metadata Standard 2023.1 added a Data Mining property with a standard list of values, including prohibited for AI or ML training. Source
- The IPTC Digital Source Type vocabulary separates media created using generative AI from media edited using generative AI, such as inpainting or outpainting. Source
- Getty Images' content license agreement prohibits using licensed content for machine learning or AI purposes unless expressly authorized. Source
- Adobe Stock product terms bar using Stock assets to create, train, test or otherwise improve machine learning algorithms or AI systems. Source
See if your company qualifies
A short company assessment. No data uploads are needed.