Why this is a list of questions, not answers
Most of the governance section can describe direction from published guidance: what regulators say about claims, what data protection asks of a prompt, what platforms document about labelling. Copyright in AI output has no equivalent that this site can cite as settling the matter, so any confident statement here would be an opinion dressed as a rule. That is why the page is written entirely as questions.
That is not a reason to ignore the subject. Marketing produces a great deal of creative work, publishes it widely and builds brands on it, so uncertainty about rights in that work is a business risk, whatever the eventual answers are. The practical response is to ask the right questions early, to the people who can answer them for your organisation, and to keep records that will matter whichever way the answers go.
The AI for creative guide raises rights and licensing for generated images in general terms; the creative production playbook covers the production process generated assets go through. This page covers the open questions across all output types, from copy and images to audio, video and code, and the records a team keeps in the meantime. The academy checkpoint on AI incidents, claims and agency rules tests the neighbouring judgements a team makes in practice, such as what may be shown of generated reviews and synthetic people and what an agency agreement should settle about AI use.
Questions about what goes in
What a team puts into a tool, and what the tool was built from, raise the first set of questions. Put them to legal counsel, with the specifics of your uses:
- Training material. Does it matter, for our use of the output, how the model was trained and on what? Does the answer change if the training data included copyrighted work used without a licence?
- Our uploads. When we upload a brand asset, a photograph or a document as a reference, do we hold the rights needed to share it with the vendor for that purpose? Do agency, photographer or stock licences we hold allow it?
- Third-party material in prompts. What is the position when a prompt includes text or images we did not create, such as a competitor's advert, a magazine article or a customer's post?
- Style references. Does prompting "in the style of" a named artist, illustrator, photographer or brand raise questions we should avoid, regardless of what the tool allows?
- Customer content. Where customers have submitted UGCContent about a brand made by customers or creators rather than by the brand itself. under our terms, do those terms cover feeding it into an AI tool, as opposed to republishing it?
- Our own material. Could material we upload be used to train the vendor's models, and what would that mean for our rights in it? This overlaps with the data questions in the disclosure and data privacy guide, which treats customer data and First-Party DataData a business collects directly from its own customers, visitors and users. under data protection rather than copyright.
Questions about what comes out
The second set concerns the output itself. Again, these are questions to ask, not points this page settles:
- Ownership. Can the output be owned at all in the markets we operate in? If so, by whom: us, the person who wrote the prompt, the vendor, nobody?
- Human contribution. Does the amount and kind of human creative contribution change the answer, and if so, what kind of contribution counts and how should we record it?
- Resemblance. What is our position if an output closely resembles an existing work, a recognisable character, a trade mark or a real person's likeness? How would we find out before publication?
- Exclusivity. If the output cannot be owned, or another user can generate something nearly identical, what does that mean for brand assets we want to protect, such as logos, mascots and signature imagery?
- Licensing on. Can we grant clients, partners or retailers rights in work that includes generated elements? What should our own contracts with agencies and freelancers say about AI use?
- Mixed works. Where a piece combines our own photography, licensed stock and generated elements, how should rights in each part be recorded?
Questions to put to vendors
Vendors cannot give you legal advice either, but they can tell you what their terms say and what they do. Ask in writing, and keep the answers with the register entry for the tool:
| Question | Why it matters |
|---|---|
| What do your terms say about who holds rights in output? | The terms may assign, license or say nothing about output |
| Do the terms differ between consumer and business plans? | Rights and promises often vary by plan |
| Do you offer any indemnity or commitment about output, and on what conditions? | Where a commitment exists, its conditions decide whether it applies to your use |
| What can you tell us about the material the model was trained on? | The answer informs counsel's view of the input questions |
| Do you use our uploads or prompts to train models, and can that be excluded by contract? | Affects both rights in our material and data protection |
| Does the tool include any filter or check for output that resembles existing work? | Tells you how much checking falls to your team |
| Do you attach provenance information or labels to generated output? | Affects your records and platform labelling |
| How will we be told if any of these terms change? | Terms change; your approval was for the version you read |
Record the answers as statements of what the vendor said, not as conclusions about your rights. The conclusions are for counsel.
Habits that lower exposure while the law settles
None of the habits below resolves a legal question. They are general working practices many teams adopt so that, whichever way the answers go, they have made deliberate choices and can show what they did.
- Tier creative uses by where the output goes. Internal concepts, mood boards and storyboards sit differently from published hero imagery, packaging and brand marks. Reserve generated output for the uses your counsel is comfortable with, and keep the most valuable brand assets human-made until advised otherwise.
- Avoid naming living artists, photographers, brands and real people in prompts. It reduces the chance of output that resembles their work or likeness.
- Upload only material you have clear rights to share. Check licences before using stock, agency work or customer content as references.
- Use tools on plans whose terms you have read. Prefer plans where the terms on output and training are clear and documented.
- Add meaningful human work and record it. Editing, compositing, rewriting and art direction are part of good creative practice anyway; recording them keeps the option of showing them later.
- Check output for resemblance before publication. A reverse image search and a reviewer's eye on familiar characters, logos and styles catch the obvious cases.
- Route generated assets through the normal rights review. The creative approval SOP and the production playbook already ask about rights; generated assets answer the same questions, plus the provenance record below.
Recording provenance
Provenance is the record of how an asset was made. It is the one thing that is useful under every possible answer to the open questions, because it lets counsel assess an asset rather than guess. Keep it with the asset in the digital asset library, not in someone's chat history.
| Field | What to record |
|---|---|
| Asset | File name, version and where it is used |
| Tool and plan | The tool, the plan or account type, and the date generated |
| Prompt and settings | The prompt text, reference uploads and any settings that shaped the output |
| Inputs and their rights | Each uploaded reference and the licence or permission behind it |
| Human contribution | Who edited, composited, rewrote or art-directed it, and what they changed |
| Checks | Resemblance check, rights review and who signed off |
| Labels | Any platform or disclosure label applied, and why |
| Vendor terms version | The date of the terms in force when the asset was generated |
The record also gives a governance measure. The share of published generated assets with a complete provenance record is a simple KPIA metric tied directly to an objective. for whether the habit has taken hold.
Tollcarrow Toys audits its generated assets
Tollcarrow Toys, an invented online toy retailer, audits its published creative and finds 64 assets with generated elements across product pages, social posts and 2 seasonal campaigns. Only 9 have any record of the tool or prompt used. 5 assets were generated from prompts that named a well-known illustrator, and 1 mascot sketch had become the face of a newsletter series. The marketing lead sends the open questions on this page to the company's legal adviser, sends the vendor questions to the 3 tool providers in use, and pauses new generated work in packaging and the mascot until counsel replies. The 5 illustrator-style assets are replaced. From then on every generated asset gets a provenance record in the asset library before approval; after a quarter, 41 of 43 new generated assets have complete records, and the 2 gaps are found at the creative approval step and fixed before publication.
Common mistakes
- Treating a confident blog post as the answer. Copyright in AI output is unsettled; any single source claiming otherwise is giving an opinion.
- Assuming the tool's permission settles the question. A tool letting you generate something says nothing about whether you hold rights in it.
- Asking the vendor for legal advice. Ask vendors what their terms say and what they do; ask counsel what it means.
- Prompting with names. Named artists, brands and real people in prompts invite output that resembles their work or likeness.
- Building a brand mark on generated output without advice. The assets you most need to protect are the ones to keep under the closest review.
- Keeping no provenance. Without a record of tool, prompt, inputs and human contribution, nobody can assess an asset later.
- Uploading material you do not have rights to share. Stock, agency and customer content come with licences that may not cover AI tools.
- Freezing on uncertainty or ignoring it. Neither helps; ask the questions, decide by use and keep the records.
