What are moral rights, and how are they different from copyright?

Copyright is the economic right: reproduce, distribute, adapt, license, sell. Moral rights are personal rights that stay with the human author regardless of who owns the economic side. The two that matter commercially are attribution — the right to be named as author, and not to have work falsely attributed — and integrity, the right to object to distortion, mutilation, or other modification prejudicial to the author's honour or reputation. Buying a copyright buys the first stack. In many jurisdictions it does not touch the second.

Can we just get a moral rights waiver in the contract?

In the UK and Canada, yes — waiver is standard practice and generally effective. In the United States moral rights are narrow, applying mainly to certain works of visual art under VARA, and they can be waived in a signed writing though not assigned. In France, Germany, and a number of other civil law jurisdictions the rights are inalienable and a blanket advance waiver is unenforceable or of very limited effect. So the answer depends on the author's jurisdiction, not on where your company sits, and a global template with a single waiver clause will be partially void across your contributor base without telling you which parts.

Does running an AI upscaler or style filter over a licensed photo create a moral rights problem?

It can, and the test is not whether the edit was technically impressive. Integrity claims turn on whether the modification is prejudicial to the author's honour or reputation, judged in context. A resolution uplift that preserves the image is a weak claim. Generative outpainting that invents scenery the photographer never shot, a style transfer that changes the work's character, or an edit that places a documentary photograph into an advertising context the author would find objectionable are meaningfully stronger. The generative cases are stronger than the corrective ones because they add authored content the author did not make and would still be credited for.

Our stock licence permits modification. Isn't that the end of it?

Not necessarily. A stock licence is granted by the agency, which holds the economic rights it received from the contributor. Where the contributor's moral rights could not be assigned, the agency never had them to license onward. Many agency terms recognise this and separately restrict uses likely to be defamatory, or that place a model or subject in an unflattering context — read those clauses, because they are the part of the licence that carries the moral rights logic. Several agencies have also added explicit restrictions on using licensed assets as AI training or generation inputs, which is a distinct restriction that a general modification permission does not override.

Who is the author for attribution purposes when AI did most of the work?

The unresolved half of the question is what protection a heavily AI-generated output attracts at all, which we cover separately. The half that is settled enough to act on is the reverse direction: where identifiable human authorship went in, attribution obligations attach to that human contribution and do not evaporate because a machine touched it afterwards. The practical risk is not that a model becomes an author. It is that a pipeline strips a credit line that was contractually or statutorily required, and does so at scale.

Moral Rights and AI: The Permission Your Copyright Licence Didn't Buy

The Right Nobody Diligenced

Content diligence has a standard shape. Who made it, what did we get, is the assignment or licence broad enough for the use, and is there an indemnity. Teams have become good at this, and our guide to content ownership clauses covers the economic side in detail.

Moral rights fall outside that shape because they do not travel with the asset. They stay with the human being who made it. The result is a category of claim that a clean chain of title does not prevent and a warranty from your counterparty may not cover: the photographer who assigned everything can still object to what your model did to the picture, and the freelance writer whose copy you own outright can still object to being named as the author of a version they did not write.

This was a low-frequency risk when modification meant a human designer making a deliberate edit that someone reviewed. AI changed the volume. A pipeline that restyles ten thousand licensed assets, or auto-translates an archive, or rewrites bylined articles into a house voice, performs ten thousand modifications nobody looked at — each of which is a discrete act against a discrete author's work.

What the Map Actually Looks Like

France and Germany. The strongest position. Moral rights are personal, perpetual, and inalienable; an author cannot validly sign away the right of integrity in advance in the way a UK contract contemplates. Contracts can regulate how the rights are exercised and can obtain consents to specific known modifications, which is a meaningfully different instrument from a blanket waiver.
United Kingdom and Canada. Waiver works and is standard. Note the UK quirk that the attribution right must be asserted to be enforceable — which cuts your way, but do not rely on an author having failed a formality if their contract asserted it.
United States. Narrow. VARA covers attribution and integrity for defined categories of visual art — paintings, drawings, prints, sculptures, certain photographs produced for exhibition — in single copies or limited editions of 200 or fewer, signed and numbered. It does not reach commercial photography, film, advertising material, or work made for hire. The practical US exposure for most companies is contractual credit obligations and false attribution claims under other theories, not VARA itself.
Everywhere via Berne. Article 6bis sets the baseline that member states implement in very different strengths, which is why the author's home jurisdiction is the fact that decides your exposure rather than your own.

The operational consequence is that the relevant field in your asset metadata is not just "licence type." It is the author's jurisdiction — and almost no content management system records it.

Where AI Pipelines Actually Touch This

Generative fill and outpainting. Extending a photograph adds content the photographer did not shoot, inside a frame still credited to them. This is the cleanest integrity fact pattern in the whole category — see also the licensing angle in our generative fill and stock licence guide.
Style transfer on commissioned art. Changing an illustrator's line, palette, or rendering while keeping their name attached goes to the character of the work rather than its resolution.
Automated translation and localisation. A poor machine translation published under the author's byline is a textbook attribution-plus-integrity complaint, and a common one because translation is where volume automation started.
House-voice rewriting of bylined copy. Publishers running an AI editor across an archive frequently keep the original byline. If the rewrite changes meaning, tone, or claims, the byline is now attributing to a journalist words they did not write and may not stand behind.
Metadata stripping. The quiet one. Image processing pipelines routinely drop EXIF and IPTC fields, including the creator and credit fields, as a side effect of resizing. Where a licence conditions use on credit, that is a licence breach as well as an attribution failure — and it is caused by a default setting, not a decision.
Training on the archive. Distinct question, different analysis, and worth keeping separate rather than folding into the moral rights discussion. Our fine-tuning licence risk guide covers the input side.

The Contract Language That Helps

A blanket "author waives all moral rights" is the wrong instrument for the jurisdictions where this actually bites, because it is exactly the thing those systems refuse to enforce. What works better is specific, informed consent to described modifications.

Describe the modifications. Name the categories the work will undergo — cropping, colour correction, upscaling, generative extension, style adaptation, translation, summarisation, rewriting — rather than reciting "modification" as an abstraction.
Say AI out loud. Consent to modification given in 2019 did not contemplate generative editing. Where you rely on an old contract for a new pipeline, refresh the consent rather than arguing about scope later.
Handle the credit both ways. Agree how the author is credited on a modified version, and agree the author's right to have the credit removed when a version departs far enough from their work that they no longer want their name on it. That second option is often the cheapest resolution to an integrity complaint and costs you nothing to offer.
Waive where waiver works, consent where it does not. A layered clause — waiver to the fullest extent permitted, plus specific consents, plus an agreement not to assert — is more durable than a single sentence drafted for one legal system.
Keep the credit in the file. A contractual obligation to preserve IPTC creator and credit fields through processing, mirrored by an actual technical control, which is where most credit obligations are broken.

A Proportionate Control Set

This does not warrant a programme. It warrants four things: record the author's jurisdiction alongside the licence in your asset system; classify your AI operations as corrective or generative and gate the generative ones on assets with identifiable human authors; audit one image end-to-end through the pipeline and check whether the creator metadata survived; and refresh consents in the contributor template at the next renewal cycle. The first and third can be done in an afternoon and between them catch most of the real exposure.

Frequently Asked Questions

How likely is an author actually to bring a moral rights claim?

Individually, not very — damages are usually modest and the process is slow. That is the wrong way to size this risk. The realistic exposures are reputational and contractual: a public complaint from a photographer or a journalist about what your AI did to their work travels further than any judgment would, and the same conduct is frequently a straightforward breach of a licence condition requiring credit or restricting alteration. The licence breach is the claim that gets brought, and it is enforced by the agency rather than the individual.

Does an employee-created or work-made-for-hire asset avoid this?

In the United States, largely yes — VARA excludes work made for hire, so employee-created commercial material is mostly outside it. In Europe the answer is weaker than people assume, because the employment relationship transfers economic rights while the personal rights remain with the individual. In practice employee claims are rare because employment provides other resolution routes, but the analysis is different from the US one and 'they were an employee' is not a complete answer in a French or German context.

We are outputting fully AI-generated assets with no human input. Is any of this relevant?

Not for moral rights, since there is no human author for them to attach to — though whether you own anything at all is then the live question, and human authorship is the requirement to watch. It stays relevant at the edges: a fully generated asset that closely reproduces an identifiable human work, and any pipeline that mixes generated and licensed elements in the same deliverable, where the licensed elements bring their author's rights with them.

Do content credentials or provenance metadata help?

They help with the attribution half and are the most practical tool available. Preserving creator and credit fields, or carrying signed provenance assertions through processing, directly addresses the failure mode where a credit is dropped by a resizer. They do nothing for the integrity half — recording faithfully that you restyled someone's illustration documents the modification rather than authorising it. Provenance is evidence, not permission.

Our agency delivers client work that clients then run through their own AI tools. Where does the risk sit?

With whoever performs the modification, which means your client — but you are the one holding the contributor relationship, and you are the one who promised the contributor something. The clean structure is to obtain the modification consent broadly enough to cover downstream client processing, disclose to the client which assets carry restrictions, and pass the restriction through in your deliverable terms. Agencies that grant broad client rights on the front end and take narrow contributor rights on the back end are absorbing a gap they have not priced.

Audit One Asset, End to End

Pick a single licensed image that goes through your normal publishing pipeline. Check the creator and credit fields at ingest, after processing, and as served. Then do the same for one bylined article that passes through an AI editing step and compare the published version against what the author filed.

Two files will tell you more than a policy review. Either the credits survived and the modifications are corrective, in which case you can stop — or one of them did not, and you have found the control worth building before the volume makes it a pattern rather than an incident.