Since the start of August, anyone in the EU who puts out an AI-generated image that could pass for a photograph of a real place has to say so. My studio makes visuals all day: fascia mock-ups, a night view of an illuminated sign, mood boards for a fit-out. So the question was never whether the rule reaches a five-person practice. The question is which of our images it actually catches, and what a sensible working routine looks like once you know.
What changed, in plain terms
There are two duties, and they sit with two different people.
The first is on whoever makes the tool. Generative tools have to mark their output in a machine-readable way, so a piece of software downstream can tell that a file was produced or altered by a model. Tools already on the market when the rule landed were given a short run-in and have until early December. Anything launched since owes it now.
The second duty is on us, the people using the tool. If you generate or manipulate an image, audio or video that resembles a real place, object, person or event, and the result would pass as authentic, you have to disclose that it was artificially made. That one came with no run-in at all. It applies today.
The penalty tier for this kind of breach runs to fifteen million euro or three percent of worldwide turnover, with the lower figure used for small companies. Nobody is going to fine a small studio fifteen million euro over an unlabelled mood board, and that is not the reason to care. The reason to care is that an unlabelled generative image inside a tender pack is now a defect in the pack, and the client's legal team will treat it as one.
The line that matters: render or photograph
A CAD visualisation is not the problem. Everyone in the room knows a render is a render, and nobody has ever mistaken one for a site photograph. The trigger is whether the image would falsely appear authentic. In practice that splits our work cleanly:
- A modelled fascia on a modelled shopfront, lit by a render engine: obviously synthetic, no label needed.
- A photograph of the client's actual shopfront with a modelled sign composited on: a montage, and it should be captioned as one, but the underlying photograph is real.
- The same shopfront where the sign, the reflections and half the pavement came out of an image model: that is exactly the case the rule was written for. Label it.
The rule of thumb we use is blunt. If the client could screenshot it, send it to their landlord, and have the landlord believe it is a photograph of the finished job, it gets labelled.
We build client visuals as modelled scenes rather than generated ones for exactly this reason, and it is also why they survive being questioned later (visualisation and prototypes).
The retouching question
Nobody wants to caption a lens correction, and the rules leave room for that: tools doing standard editing assistance, or which do not substantially change what you fed them, sit outside the marking duty. Straightening verticals, lifting shadows, cloning out a bin bag. That is housekeeping. Where it stops being housekeeping:
- Sky replacement on a site photograph used to show a finished installation.
- Generative fill inventing a stretch of facade the camera never saw.
- Removing scaffolding, a competitor's sign or a parked van from a survey photograph.
The last one is where I worry least about a regulator and most about our own work. A survey photograph with a van generatively removed is no longer a survey photograph. Someone six months later will be measuring off it, and the pixels behind that van are a guess.
Where it bites in signage and fit-out
- Pitch decks and mood boards: highest volume, lowest risk. One caption line under the image and you are done.
- Planning and landlord consent packs: highest risk. Our rule is that no generative imagery goes into a submission at all. If a planner needs to see the sign in context, it is a modelled visual placed over a dated, untouched site photograph.
- Marketing and case studies: photograph the finished job. If the photograph has been cleaned up by a model, say so, or do not publish it.
- Handover and maintenance files: never. A drawing set and a set of as-built photographs are evidence, and evidence does not get improved.
Text is a quieter question for us. The disclosure duty on generated text is aimed at material published to inform the public on matters of public interest, and it falls away where a person takes editorial responsibility for what goes out. A method statement drafted with a model, then read, corrected and signed by an engineer, is that engineer's document.
Ireland and Poland: who is actually watching
Ireland took the distributed route. A central AI office acts as the coordinating body and the single point of contact, sitting over the sectoral regulators that already exist, fifteen of them, with the data protection, media, consumer and health and safety bodies among them. For a signage studio the realistic path to trouble is a consumer or advertising complaint, not a knock from an AI inspector.
Poland built one new body for the job. The national act was signed in July, the chair is due in October and the commission itself starts work in November. So right now, in September, the duty is fully live for a Polish deployer while the office meant to supervise it is still being staffed. Do not read that as a grace period. The obligation applies directly, and clients ask about it long before regulators do. Two tender questionnaires crossed my desk this summer asking whether any submitted material was AI-generated. That is the mechanism that will change behaviour, not the fines.
What we changed in the studio
Five things, none of them expensive:
- A caption convention. Every generative image carries a line under it: AI-generated concept image, not a photograph of the finished installation. Same wording in both languages.
- File naming. Generative output gets a prefix in the filename, so it cannot quietly migrate from a mood board into a drawing register.
- Keep the metadata. Export routines that flatten and re-save will strip the machine-readable marking the tool wrote into the file. Ours no longer do.
- Two folders per project, concept and evidence. Nothing generative ever enters the evidence folder.
- One line in the proposal saying what we used AI for on this job. Clients read it, and it has yet to cost us work.
What to do with this
If you produce visuals for clients, this is an afternoon of work, not a project. List which of your outputs are actually generative, agree a caption convention and put it in the template, and draw a hard line between images that sell an idea and images that stand as record. The compliance part is small and mostly mechanical. The professional part, that a photograph in a handover file is still a photograph, is the part worth keeping regardless of what any regulator does next.
