New Rules for AI Pictures
New Rules for AI Pictures
New European rules on AI-generated pictures came into force on 2 August 2026. If you make Easy Read, they may apply to you. Here is what changed, what it means in practice, and what we are doing about it.
You may have seen the headlines about AI labels becoming compulsory in Europe. The reporting was a bit breathless, so we wanted to write something calmer and more useful for the people who actually use our pictures.
This is a plain guide to what the rules ask for, what the real risks are, and where Photosymbols stands.
What has changed
The European Union's AI Act now requires AI-generated content to be identifiable as AI-generated. The relevant part is called Article 50, and it took effect on 2 August 2026. The European Commission has also published a set of official icons that organisations can use to label AI content.
It asks for two different things, and they are easy to confuse:
- Something inside the file. A technical marker embedded in the picture itself, which software can read but people cannot see. Its job is to let platforms and checking tools recognise AI content automatically.
- Something people can see. A clear label shown to anyone looking at the picture, at the moment they first see it.
One does not replace the other. A visible label does nothing for the technical requirement, and a hidden marker does nothing for a reader trying to understand what they are looking at.
Does this apply to you if you are in the UK?
Possibly, yes. The rules are written to reach beyond the EU's borders. They apply where AI content ends up being seen in the European Union, whoever made it.
For most UK organisations that means asking a simple question: could the Easy Read you are publishing be read by someone in an EU country? If it is on an open website, the honest answer is usually yes.
There is a second point that catches people out. When you take a picture and publish it in your own leaflet, poster or web page, you are the publisher. The duty to show a visible label sits with the organisation putting the content in front of readers, not only with whoever made the picture.
What happens if you get it wrong
The headline number is large. Fines can reach 15 million euros or 3% of worldwide turnover.
We would rather be straight with you than use that figure to frighten you into anything. It is a ceiling, aimed at very large technology companies, and the rules specifically say smaller organisations should be treated proportionately. If you are a council, a health trust, a housing association or a charity, a fine is not the most likely consequence of getting this wrong.
The likelier costs are quieter, and in our view they matter more:
- A complaint you have to answer, from someone who spotted a generated person and felt misled.
- A procurement or funding form asking how you comply, and no good answer to give.
- A piece of Easy Read that loses its authority, because once a reader doubts one picture they start to doubt the words too.
That last one is the real risk for Easy Read. Trust is the whole product.
The bigger issue: pictures made without disabled people
The law is concerned with labelling. We think there is a more important question sitting underneath it, and it is one the law does not ask at all.
Who was involved in making the picture?
Generative AI is very good at producing an image of a person who does not exist. It is much less good at producing an image of a disabled person that disabled people recognise as themselves. The mistakes are often invisible to a non-disabled reviewer.
Here is a real example from our own work. We had a library photograph of a young woman with Down's syndrome sitting on her pink bike, in a hi-vis vest and glasses, grinning at the camera. We asked our AI tool to do one small thing: make the bike blue.
The bike turned blue. The model also quietly replaced the woman's identity. Her Down's syndrome features were erased. Her braces were gone. Her hair was longer and lighter, her shirt had changed, and she had been shifted and resized within the frame. We had asked about a bicycle, and it handed us a different person.
We then tested it properly, because one bad result is a story and five is a finding. The model we were using at the time failed the same way in five runs out of five, and in every run it also shifted and resized the subject within the frame.
Before. The photograph, as we took it. |
After. We asked for a blue bike. Look at her face. |
At a glance the second picture looks fine. Here are the two faces side by side, close up:
Left, the person we photographed. Right, what the AI handed back. She is a real person. The person on the right is not her, and is not anybody.
A small detail can carry a whole identity. Nobody asked the AI model to change her. It happened because the system has no idea what it is looking at, and no reason to treat a person's face as different from a bicycle. We caught it because we know her. Someone who had never met her would have had no way to know.
That is the danger of pictures made without the people they claim to represent. Not that they look obviously wrong, but that they look plausibly right to everyone except the community being pictured. A generated "disabled person" has consented to nothing, has been paid nothing, and cannot tell you that you have got them wrong.
We changed the tool. Our AI editing now runs on a model that holds the person still while it changes what you actually asked about, and we test it against exactly this failure. But the lesson is bigger than one setting: if you are editing pictures of disabled people with AI, assume it will change them unless something is stopping it.
Disabled people have spent decades arguing for accurate representation. It would be a poor outcome if the technology that finally made pictures cheap also made real people optional.
What Photosymbols has always done
We have never mixed AI pictures into the library quietly. From the moment we began offering them, AI-generated images have been marked as AI-generated and kept in their own clearly identified part of the library. If you have been searching our library, you have always been able to see which is which.
That was a decision we made because it was right, not because a regulation required it. We are glad the law has now caught up.
Behind that, our foundation has not changed:
- Real photography since 2003. We still travel, listen, photograph and pay real actors for their work.
- Five paid Expert Advisers with lived experience, who review AI-assisted pictures every week.
- Paid co-production. People with learning disabilities help shape what we make, and what our technology becomes.
What we have built
When we first wrote about these rules, most of this was a plan. It is now built. Here is what it is, in plain terms.
1. An invisible marker inside the picture file. AI pictures leave our library carrying content credentials that record how the image was made, written in C2PA – the open standard that camera makers, Adobe and the large platforms are adopting. Alongside that, a second marker is embedded into the pixels themselves, so the information survives being resized, re-saved, screenshotted or reposted. That is the moment when ordinary file information is normally stripped away, and it is the part of the rules most organisations will find hardest to do for themselves. If your pictures come from us, it is handled.
2. A visible mark you can see at a glance. The European Commission's mark for AI content is a heavy "AI" set beside a short qualifier in spaced capitals. We built our marks to follow that pattern rather than inventing a private visual language, so anyone who has seen an AI label elsewhere in Europe will recognise ours. "AI+" is our plain-language shorthand for the EU's "AI modified" category.
3. A "Real People" mark, which the law does not require. This is the one we care about most. The rules tell you when AI was involved. They say nothing about when a real person was involved. We think that is the more useful fact for Easy Read, so we show it. It has no EU equivalent, so we drew a pictogram in the same family as the other two, and it is the only mark that carries colour.
4. A plain line on AI-written text. The rules cover words as well as pictures. Where AI helped write something – an EasyMaker pack, a page in our news feed, an Easy Writer draft – a short sentence now says so in ordinary English, for example "Made with EasyMaker. AI helped write this."
What the marks look like
Three marks, one family. Each is a white pill with an ink mark, so it stays readable on top of a photograph:
| Real People. A genuine photograph of real people, with no AI generation and no AI editing. Real actors, paid for their work. | |
| AI | AI Generated. Made entirely by AI. The picture was generated, not photographed. Nobody in it is a real person. |
| AI+ | AI Modified. A real photograph that has been edited or enhanced using AI, so parts of the image are AI-generated but the people are real. |
A picture can carry two marks. A photograph of a real actor standing in a scene we extended with AI shows Real People and AI Modified together: the first credits the person, the second is honest about the rest of the frame.
AI+
Some pictures carry no mark at all – a photograph with no people in it and no AI involvement has nothing to declare.
Where you will see them
On a picture's own page the marks sit in the bottom-left corner of the photograph as small round discs, always visible:
AI AI+
Hover over one, or reach it with the keyboard, and it explains itself in a sentence. Underneath the picture, a panel headed "How this picture was made" shows the marks written out in full. In search results the discs appear on a picture as you move over it, so a page of results does not become a wall of badges.
Where our team has something worth saying about how a particular picture came about, you will find it written underneath in their own words, by hand. That part is not automated, and it is not generated.
What we suggest you do
None of this needs to be complicated. Four practical steps:
- Know which of your pictures are AI-generated. If you cannot tell, that is the first problem to solve, and it is a good reason to use a library that tells you.
- Label AI pictures where your readers can see it. A short, plain line near the picture or in the document credits is enough. It does not need legal language.
- Be careful with pictures of people. Generated images of people carry the most risk, both legally and in terms of trust. For Easy Read about real services and real rights, real people are the safer and better choice.
- Ask your suppliers what they do. Any picture library selling you AI images should be able to tell you plainly which images are AI, what marking they carry, and who was involved in making them. If they cannot, that tells you something.
A note on EasyMaker and PictureMaker
If you generate your own pictures inside EasyMaker or the new 'PictureMaker' app in the photo library, those images carry the same invisible marking automatically when you download them. EasyMaker packs whose words were written with AI help now carry a short line saying so. When you publish a pack anywhere else, remember that you are the publisher, so keep a visible note in your own document too.
This post is general information about how we are responding to new transparency rules. It is not legal advice. If you are unsure whether the EU AI Act applies to your organisation, please seek your own advice.