Europe’s new AI labels are here. They do not mean the content is true.

Article 50 of the EU AI Act covers several different moments that often get flattened into one headline. The first is conversation. A system designed to interact directly with a person has to disclose that it is AI unless that would already be obvious to a reasonably observant user in the situation.

The second is synthetic output. Providers of systems that generate audio, images, video or text must make those outputs detectable in a machine-readable format, as far as technically feasible. Think of this as a signal software can inspect even when a person cannot see it.

The third is visible disclosure by the organization using the system. Realistic AI-generated or manipulated images, audio and video—deepfakes in the law’s language—must be labelled. People must also be told when emotion-recognition or biometric-categorisation tools are operating around them.

There is a narrower rule for text published to inform the public about matters of public interest. It needs disclosure when AI generated or manipulated it, unless the material went through human review or editorial control and a person or organization accepts editorial responsibility. Standard editing that does not substantially change the input or its meaning is also treated differently from generation. In other words, the rule is not meant to slap an AI badge on every cleaned-up sentence.

A visible label is for the person who is about to believe, answer or share something. It should arrive at the first interaction or exposure, not three clicks into a policy page. The EU has published a set of icons that companies can use so the warning can look more consistent across services. The icons are optional. The disclosure duty is not.

A machine-readable mark is for the chain behind the screen: platforms, verification tools and other systems trying to detect whether a file was generated or altered. It may travel in metadata or another technical signal that an ordinary viewer never notices.

You need both because each one fails in a different way. A tiny visible badge can be cropped out or ignored. A hidden mark may survive the clean file but disappear after a screenshot, export or re-encoding. Even when it survives, it only helps if the next service checks it and tells the viewer what it found.

The law splits the work between providers—the companies making AI systems—and deployers, the organizations putting those systems in front of people. That matters for a normal user. The model maker can add a technical mark, but the shop, news site or app using the model still owns the moment when a person needs a plain warning.

A generated image can be clearly labelled and still make a false claim. It can also be labelled and show an accurate reconstruction based on good evidence. The badge describes production. It does not grade the result.

The same goes for chatbots. A bot can announce itself perfectly, then give you the wrong cancellation date, invent a medical detail or confidently misunderstand the reason you called. Disclosure removes one deception: pretending the machine is a person. It does not remove the need to check the answer.

The reverse is just as important. No label does not prove that a photograph, voice or paragraph is human-made. A service may be outside the rule, exempt, late, noncompliant, or unable to preserve the mark through the way the file was shared. Old-fashioned editing and staged human deception still exist too.

Treat the label as context, not a verdict. It changes the next question from ‘is this secretly AI?’ to ‘what source, person or record supports this?’ That is useful progress. It is not the end of the job.

Theo Marlow would resist the easy sentence that labels make synthetic content trustworthy. The law requires origin signals and disclosure. It does not require every labelled claim to be accurate, consensual or fair. For a consequential post, his next move is the same as before: find the original source, date and accountable publisher. The label is one piece of evidence about the file, not evidence for the story the file tells.

Ivy Chen starts inside the small team that now has to comply. She would not hand this to a designer as an icon-cleanup project. First list every place where a customer might meet AI: chat widget, support phone line, avatar, generated product image, public explainer and automated social clip. Then name who owns the notice and who checks that it still appears after the content leaves the original tool.

Theo is protecting readers from overtrusting the badge. Ivy is protecting teams from shipping a badge that only exists in the happy path. One keeps the promise honest. The other makes sure the promise reaches a person.

For low-stakes material, the label may be all the context you need. A synthetic illustration in a clearly fictional post does not need a forensic investigation. Notice it and move on.

For anything that could cost money, change a record, damage someone’s reputation or send people into the street, slow down. Find the earliest copy. Look for a named source, a capture date and an organization willing to stand behind the claim. If the material depicts a real event, compare it with reporting or records from somewhere independent.

When a chatbot is involved, ask for the relevant policy, order record or source rather than a second confident explanation. If it is customer support, confirm what actually changed after the conversation. A polite answer is not a refund, a reservation or a corrected account.

Do not read an absent mark as a green light. If the post is surprising enough to make you share it immediately, that is exactly when another minute is cheap. ‘I could not verify this’ remains a useful message.

Start with an inventory, not a policy rewrite. Where does AI talk directly to people? Where does it generate or substantially alter audio, images, video or text? Which outputs concern public-interest information? Which ones already have a person reviewing and accepting responsibility before publication?

Put the human-facing notice at the first meaningful contact. ‘AI assistant’ beside the chat title is clearer than a disclosure buried in terms. For realistic synthetic media, keep the visible label close enough to survive ordinary viewing and sharing. Preserve the machine-readable mark instead of stripping metadata during export when you can.

Test the route out. Download the file, post it to the place customers actually see it and check the result from another account or device. Call the support line. Open the social clip. Forward the image. The compliance story should survive the same boring path as the content.

Give one person the list and a date to recheck it. The Verge reports that new systems face the rules now, while services launched before August 2 have until December 2 to comply. Do not wait until December to discover that three vendors each thought another company owned the disclosure.

Finally, write the sentence support staff can use when someone asks what the label means: ‘This tells you AI generated or changed the content. It does not confirm that every claim is accurate.’ That sentence is less impressive than a trust badge. It is also true.

The best outcome here is not a new symbol people learn to admire. It is the end of a small, exhausting guessing game. You should not have to interrogate a cheery customer-service voice to learn whether a person is on the line. A convincing video should not get to borrow reality without saying it was made.

Europe’s rule gives that disclosure a legal job and backs it with possible fines of up to €15 million or 3 percent of global annual turnover for companies, with proportionality for smaller firms. Enforcement will decide whether the label becomes ordinary or merely another box in a launch checklist.

For readers, the useful habit is simple: believe the label about the process, then check the claim on its own terms. For teams, make the disclosure plain enough that nobody needs a lawyer or a settings tour to find it.

A label cannot tell you what is true. It can stop wasting your time on one question before the real checking begins.