If AI was here, do I have to tell you?
Transparent
tran(t)s-ˈper-ənt.
“Presenting no obstacle to the passage of light, so that what is behind can be distinctly seen,” early 15c., from Medieval Latin transparere “show light through,” from Latin trans, “across, beyond; through” + parere “come in sight, appear; submit, obey”.
Transparency is all about shining a light on something hidden or obscured. But much of what we see or hear online arrives without us knowing how it was made, or what’s been done to it along the way.
When it comes to artificial intelligence, yet another veil is added between the thing itself and the person trying to understand it. And it’s not just the technology that’s often opaque — but the companies behind it, and what people are actually using it for.
In 2024, the European Audiovisual Observatory commissioned me to write about transparency regarding digital doubles and cloned voices (FKA “deepfakes”). Using OpenAI’s now-infamous rollout of the Scarlett Johansson-esque voice model for ChatGPT-4o as a key example, the paper explored the risk that someone’s likeness, voice, or performance may be digitised and used without meaningful consent.
There’s that word again: and yes, lawyers rightly focus on consent when it comes to protecting those whose likeness, voice, or work is used by someone else. But although consent may be the legal mechanism, it is meaningless without relevant context and accurate information.
The same is true on the other side of the screen, because transparency is not only about protecting the person who is seen or heard in the content, but about protecting the audience and consumers who are watching, listening, and reacting to the content.
This is why the EU AI Act’s transparency rules are so important. Under Article 50, when AI is used to generate or modify content, this fact must be disclosed in a clear and distinguishable manner at the time of first interaction or exposure, most commonly by way of a label, caption, or watermark.
When we flip to Article 50, it seems both comprehensive and straightforward. Deepfakes have to be disclosed. Certain AI-generated or manipulated text has to be labelled. Some AI outputs must be marked and detectable.
But despite the detail, the letter of the law leaves many questions unanswered in practice — because creative work rarely happens in neat little legal categories.
Take, for example, the exception that says no label is needed if AI is “merely” used for “an assistive function for standard editing.” Does that cover automated subtitles? What about skin retouching, colour grading, upscaling, and other post-production effects?
Another special rule applies for deepfakes that form part of an “evidently artistic, creative, satirical, fictional or analogous work.” In those cases, disclosure can be made in a way that “does not hamper the display or enjoyment of the work.”
Broadly speaking, these exceptions make sense: a label should not ruin a film, kill a joke, spoil a reveal or flatten a piece of art. But they also raise the obvious follow-up concerns: where is the line between preserving the audience experience and leaving people unsure about what they are actually watching?
That’s why the new Code of Practice on Transparency of AI-Generated Content has been so eagerly awaited. And although it doesn’t answer every question, it moves the conversation from abstract principle to practical implementation for real-world workflows.
The Code gives practical detail on what complicance looks like, IRL. I covered my top five takeaways in this Instagram video, but the key points are also set out below.
1. This is not just an EU problem.
One of the first things to remember about the EU AI Act is that it can matter even if you are not sitting in the EU. If your goods, services or content are made available to people in the European Union, the rules may still be relevant. Put crudely: an app built in Miami but made available to subscribers in Madrid does not get to pretend Europe is someone else’s problem.
2. Labels need to be clear and timely.
If something is a deepfake, people should know when they first see or hear it. Not after the fact. Not buried in the small print. Not tucked away in a caption nobody reads. The point is that the audience understands what they are encountering at the moment it matters.
3. Format matters.
Video may need a visible label. Audio may need an audible disclosure. Long-form content may need reminders, not one tiny warning at the start. The label has to work in the medium, not just exist somewhere in theory.
4. Creative work gets some breathing room — but not a free pass.
Fiction, satire and artistic works may label in a way that does not ruin the experience. Good. A label should not kill a joke, spoil a reveal or flatten art into compliance sludge. But that is not a free pass to mislead people. But although the law won’t force you to slap ugly warning stickers onto everything, “creative work” or “fiction” is not a magic phrase that will allow you to mislead audiences.
5. You probably don’t need to hit the “panic” button
Labelling media is nothing new: we already tell audiences when content is advertising, dramatised, reconstructed, or sponsored. As such, AI labelling probably doesn’t need to become a full-blown compliance crisis — it just needs to be built into the workflow. Decide who checks AI use, who approves labels, where they appear, what records are kept, and how mistakes are corrected.
And finally, because it
really matters — transparency cannot
replace consent.
A label does not make an unauthorised use lawful. It does not fix a bad release form. It does not compensate a performer. It does not cure a misleading advert. It does not give anyone retrospective permission to clone someone’s voice, face, performance, or work.
