Collections numériques
AI Art, Copyright, and Attribution: A Practical Guide
Separate authorship, training data, platform terms, and licensing questions, and build a process record that protects your work.
AI art raises several different legal questions that are often collapsed into one. “Who owns this image?” is not the same as “was the model trained lawfully?” or “can I sell this?” or “can I stop someone else from copying it?” Each has a different answer, a different decision-maker, and a different level of uncertainty.
This guide is a practical map, not legal advice. It explains what to check, what to document, and where the unresolved parts are, so you can ask a lawyer a specific question instead of a vague one.
Four questions, not one
1. Is there a copyright in the output?
Copyright protects works of human authorship. In several jurisdictions, including the United States, material generated purely by a machine from a prompt has been treated as lacking sufficient human authorship for protection. A prompt alone has generally not been enough. Where a human contributes substantial creative expression — composition, arrangement, editing, selection, painting over, compositing — the resulting work may be protected, but the scope is often thin and fact-specific.
Practical consequence: do not assume you can stop others from reusing a raw generation. If exclusivity matters commercially, add human-authored contribution and document it.
2. Was the model trained on material that could not be used?
This is the dispute you read about in the news. Training datasets have included copyrighted images, often scraped at scale. Whether that constitutes infringement, fair use, or something else is being litigated and legislated differently across countries. There is no settled global answer, and the outcome may depend on the model, the jurisdiction, and the specific use.
Practical consequence: you cannot verify a model’s training set from the output. Ask vendors what they disclose, prefer models with published data policies, and treat “the model is safe” as an unverified claim, including when you make it yourself.
3. What do the platform terms allow?
Your rights are shaped by a contract before they are shaped by copyright law. A tool’s terms may grant you ownership of outputs, a licence to use them, or nothing at all. They may restrict commercial use, require attribution, prohibit certain subjects, or permit the provider to use your inputs and outputs to train future models.
Practical consequence: read three sections — ownership of outputs, licence to your inputs, and commercial restrictions. Save a dated PDF of the terms when you publish. Terms change.
4. Are there third-party rights in what you generated?
Even a lawful generation can reproduce protected material: a logo, a character, a living artist’s distinctive style as misappropriated likeness, a private person’s face, a trademarks. Style itself is generally not protected, but specific characters, brands, and identifiable people carry separate rights.
Practical consequence: avoid prompting for named characters, brands, and real people unless you have a reason and a permission. For editorial or commercial work, assume you need clearance.
Reading platform terms without a lawyer
Look for these clauses and write down the answers:
- Output ownership. “You own the outputs” versus “we grant you a licence.”
- Input rights. Does the provider claim a licence to your uploads? Can it train on them?
- Commercial use. Permitted, restricted by tier, or silent.
- Attribution. Required, requested, or not mentioned.
- Content restrictions. Named categories that could void your account.
- Termination and retention. What happens to your files if you leave.
If a clause is ambiguous, email support and keep the reply. A written answer from the vendor is not a legal opinion, but it is evidence of what you were told.
What documentation actually helps
A process record is the single most useful habit. It does not settle authorship by itself, but it makes your contribution legible. Keep, per project:
- Tool and version. Model name, version string, date, and account tier.
- Inputs. Your own photographs and sketches, with dates, plus any licensed source material and its licence.
- Prompt history. Not for aesthetics — for evidence of iterative direction.
- Selection and rejection. Contact sheets of what you discarded, and why.
- Human edits. Layers, masks, overpaints, crops, colour work. Save editable files where possible.
- Correspondence. Permissions, licences, vendor answers, collaborator agreements.
- Publication record. Where it appeared, under what licence, with what credit.
Store this alongside the work, not in a different app. In a dispute, the person with the folder wins more often than the person with the better argument.
Attribution in practice
Attribution norms for AI art are still forming, so be more generous than the minimum:
- Credit the tool and version, as you would credit a darkroom process or a printing house.
- Credit collaborators in plain language: “images generated by X, edited by Y.”
- Do not credit the model as an artist; it cannot hold rights or responsibility.
- If you used someone’s photograph as an input, credit and licence it as you would any source.
- Label synthetic images where viewers could reasonably mistake them for records of reality.
For projects that treat attribution as part of the artwork, Rhizome and the Ars Electronica archive are good case-study sources. AIArtists.org tracks the evolving debate.
Browsing, downloading, and commercial use
Three permissions get mixed up constantly. Keep them separate:
| Action | What you need |
|---|---|
| Viewing an image online | Nothing beyond the site’s terms |
| Downloading and reposting | A licence or an exception such as fair quotation |
| Selling prints, products, or client work | A commercial licence, plus clearance for depicted rights |
A museum collection page usually permits viewing and often allows download for research, but not commercial reuse. The Met Collection publishes rights metadata per object; read the object record, not the site footer. Wikimedia Commons hosts freely licensed media, but each file carries its own licence and attribution requirements — check the file page, not the search result.
For AI-generated material, add one more layer: the tool’s commercial terms. A model that permits personal use may forbid selling outputs, or may allow it only on a paid tier.
A checklist before you publish
- I know which tool and version produced each output.
- I checked the platform’s terms on ownership and commercial use, and saved a copy.
- I have a licence for every input I did not make.
- No real person, brand, or protected character appears without clearance.
- Human contribution is documented and, where possible, visible in the file history.
- Credit lines name the tool, collaborators, and sources.
- Synthetic images are labelled where the context implies documentation.
If any box is empty, the safe move is to hold the work back or publish it non-commercially while you resolve it. That is a business decision, not a legal conclusion.
Where the uncertainty lives
The unsettled parts are authorship thresholds, training-data legality, and the scope of “style” protection. None of these will be resolved by a blog post, and they may resolve differently in different countries. What you control is the record: what you made, what you used, what you were permitted to do, and what you told people.
For the wider picture of how institutions collect and display this work, see the collections category. For a beginner’s process guide, start with AI Art for Beginners.