An AI campaign can be cleared to make and still not be yours to defend
Your agency delivers the campaign's signature image. It tests well, it looks expensive, and part of what you paid for was the promise that nobody else would have it. Six weeks after launch, a competitor runs something nearly identical. You call legal, and legal asks the question nobody asked in the production meeting: who is the author?
If the honest answer is "the model," you may have very little to enforce. The image still works. It still sells. But you didn't buy what you thought you were buying.
What bothers me is that the industry treats AI production as a clearance problem. Is the training data licensed? Is the customer data lawful? Did legal approve the tool? Those are real questions, and they are the wrong place to stop. Clearing the inputs tells you that you were allowed to make the thing. It tells you nothing about whether you own it.
Two documents published in Brussels on October 21 make that gap hard to ignore. EU member states agreed the Council's negotiating position on the Digital Omnibus, including changes to how the GDPR treats AI. The same day, the European Parliamentary Research Service published a briefing on whether AI-generated outputs can be protected by copyright. Neither is law. Read side by side, though, they describe two separate rulebooks that most procurement conversations still collapse into one.
Two rulebooks, one campaign
Take a campaign built from customer interviews and a licensed image library.
Data protection asks whether you may process the people in that material. Copyright asks who owns what comes out.
A lawful basis under the GDPR is not a copyright license, and a photographer's license gives you no lawful basis to process the identifiable people in the photo. Training a model and using its output are separate copyright questions again. Same project, different rights, different people who can come after you.
The Council's text doesn't blur this. It keeps the GDPR's current definition of personal data, and its recitals state that developing or operating AI "does not in itself constitute a legitimate interest." Whether legitimate interest covers an AI project still depends on the purpose behind it, case by case.
"We're doing AI" is not a justification. You still have to say what you're doing it for.
Nor did Brussels just open a locked door. In February, the European Data Protection Board and the European Data Protection Supervisor told the Commission a dedicated AI provision was unnecessary, because the EDPB had already said in 2024 that legitimate interest can support some AI processing under the GDPR as it stands. Anyone selling the October mandate as Europe's new green light for AI is selling something that partly existed already.
The part agencies should read twice is the exception for sensitive data. The Council would allow incidental and residual processing of special-category data during the development and technical operation of AI systems: data the controller never intended to process and tried to filter out. Once found, it has to be erased. If erasure is technically impossible or would take manifestly disproportionate effort, for instance because the model would have to be re-engineered, the controller still has to keep that data from feeding outputs or being disclosed.
Two limits matter for anyone producing campaigns. The exception does not cover sensitive data fed to a system during deployment. And it does not apply when the processing actually needs that data, with sensitive data tied to a person's voice or image named as the example. An agency that deliberately builds a campaign on customers' faces or voices cannot call that residual.
None of this replaces an Article 6 lawful basis, either. An exception to the ban on sensitive data was never a lawful basis on its own.
Six hundred prompts, no copyright
Now the other end of the campaign. The EPRS briefing describes an emerging consensus across jurisdictions. Purely AI-generated content with minimal or no human creativity does not qualify for copyright. Work where a person exercised substantial creativity may, and that gets decided case by case. The useful part is the case law the briefing collects, because it takes apart several comfortable assumptions at once.
The US Copyright Office refused to register an AI image even though the applicant had worked through more than 600 prompts. A Munich court ruled this year that the time invested, and whether you paid for a premium version of the tool, are irrelevant. The Higher Regional Court in Düsseldorf held that picking one image from several AI suggestions is not authorship.
That last ruling should worry every creative team whose "human input" consists of choosing a favorite and sending it to the client.
What does count is human creative control that shows up in the result: specific creative direction, real editing, substantial transformation, a creative arrangement of the material. A Frankfurt court also treated the human-written and AI-generated parts of one work separately. A single deliverable can be partly protected and partly free for anyone to take.
Then comes the twist. The briefing notes that an output with no copyright of its own can still reproduce or reflect protected works that went into the system. Put those two findings together and you get the worst commercial combination I can think of: an asset you cannot stop others from copying, which may still draw a claim from someone whose work it echoes. Liability without ownership.
Keep the receipts
If I were the buyer, I would want two records, and I would not accept either one as a substitute for the other.
The first is the data record. For the sensitive-data exception, the Council's text adds an explicit expectation that controllers document their assessment and keep processes to monitor and demonstrate that their safeguards work. The EDPB's draft guidelines on web scraping for generative AI, adopted in July for public consultation, push the same discipline to the very start. Define precise collection criteria, filter out categories you don't need, and skip sites that clearly object to scraping, all before anything is collected. If a campaign system was trained on scraped material, the questions begin there, not at launch.
The second is the authorship record. Versions that show what a designer actually changed. The composition decisions. The edits that moved the work away from what the model produced. A long prompt history is not that record, and an approval email certainly isn't. Signing off on a release does not make anyone an author.
A privacy assessment says nothing about who created the image. A revision history says nothing about whether you were allowed to use the customer data.
Keep both, keep them together, and make sure someone can walk a campaign from source material to final file.
Stop paying for exclusivity nobody can explain
None of this makes unprotected AI output worthless. Plenty of content is meant to live for a week. If you need forty variations for a short social burst, limited copyright protection may be a perfectly rational trade for speed and cost.
The problem is paying brand-asset prices for something the contract never examined. If an agency promises exclusivity, I want it to show which parts of the deliverable can carry that promise and what human creative work produced them. The agreement should separate any copyright actually being assigned from contractual restrictions on reuse. Those restrictions bind the parties who sign them. They create nothing you can enforce against a competitor who never signed anything.
The client side needs the same precision. Handing customer material to a supplier should trigger a real conversation about what processing is permitted, whether the supplier keeps the material, and whether it ends up training their models. A warranty about output rights doesn't answer that. A broad promise that the inputs were lawful doesn't make the output protectable, or free of someone else's claims.
Underneath the legal language sits a budget decision. If the campaign carries an identity you plan to defend for years, meaningful human creative control is a production cost, and it belongs in the plan. If the work is disposable, choose differently. Just make the choice visible before procurement compares two quotes that are not buying the same thing.
The Council's privacy changes still have a long road ahead. Parliament hasn't adopted its own position, and negotiations only start once it does.
Watch whether the limits on residual sensitive data, and the safeguards around them, survive. But don't wait for that text to answer a question it was never going to answer. No privacy reform will tell you who owns your campaign.
Settle the exclusivity question before production starts. If nobody can tell you which rights you're buying, that uncertainty belongs in the contract and in the price, not in a call with legal six weeks after launch.