AI Content Labels: What the EU Watermark Rule Requires
AI content labels are now required in the EU. See what the December 2 watermark deadline means, how hidden marks work, and what the new AI icon tells you.

Table of contents
You are scrolling through a travel site and a sunset photo catches your eye. In the corner sits a small grey badge with two letters on it: AI. Tap it and a second panel tells you the beach was never photographed. A model drew it.
That badge is one of the new AI content labels that Europe now expects on generated images, video, audio and text. Most of the rules started applying in August. The last piece lands on December 2, when older AI tools already on the market lose their grace period for hidden watermarks. If you make content, buy ads, run a store or simply wonder whether a picture is real, this changes what you see and what companies owe you.
Here is what the rule asks for, how the invisible part works, and what it means even if you never set foot in the EU.
What the AI Content Labels Rule Actually Requires
The rule lives in Article 50 of the EU AI Act. It splits the work between two groups, and the split explains almost everything else.
- Providers build the AI system. Think of the company behind an image generator or a chatbot. They must mark every output in a machine-readable way, so software can tell it was generated or altered.
- Deployers use the system to publish something. That could be a brand running an ad, a news site, or a creator posting a video. They must show a visible label when the content is a deepfake, and when AI-written text is published to inform the public.
So the image generator stamps the file invisibly. The company that posts the image adds the badge you can see. Neither half works without the other, because a hidden mark tells you nothing on screen and a visible badge can simply be cropped off.
The text exemption most people miss
AI-written text gets a softer deal. According to Tech Policy Press, published text is exempt from the visible disclosure if a human reviewed it or holds editorial responsibility for it. The law firm Freshfields notes that the publisher only has to name who holds that responsibility and document how review works. It does not have to log every single check.
Why the December 2 Deadline Matters
The AI Act originally set August 2 as the start date for all of Article 50. Then Brussels passed a simplification package, the Digital Omnibus on AI, which the Council of the EU approved in late June. The law firm Lewis Silkin summarised the change for transparency like this.
- Generative AI systems placed on the EU market before August 2 have until December 2 to add machine-readable marks.
- Systems placed on the market on or after August 2 had to comply from day one.
- The visible labelling duties for deployers were not delayed. They have applied since August 2.
The European Commission had proposed a six-month grace period. Lawmakers trimmed it, and Freshfields describes the final version as a four-month transition. That is why December 2 is the date compliance teams have circled. It is the moment the popular, older tools (the ones most people actually use) fall fully under the rule.
How Invisible AI Watermarks Work
The law sets four tests for any marking system. In plain terms, it has to work, work across companies, survive editing, and not throw false alarms.
Here's the uncomfortable part: nothing on the market passes all four alone. Tech Policy Press reports that the Commission's own code of practice admits no single technique currently meets every requirement. Metadata is easy to read but disappears when you screenshot or re-save a file. A watermark buried in the pixels survives more abuse but is harder for outsiders to detect.
Two layers instead of one
The answer in the final code, published on June 10, is stacking. Freshfields lists what the code expects from providers.
- At least two layers of machine-readable marking wherever one technique cannot meet the four tests.
- Signed metadata that is digitally signed and time-stamped, so tampering shows.
- A watermark woven into the image, audio or video itself.
- Optional extras such as fingerprinting (a stored signature of the file) and internal logs of what the system generated.
Text is the hard case, because plain text carries no metadata. The code sets a threshold: free-form text longer than 200 tokens must be watermarked, while shorter snippets need not be. A token is a chunk of a word. Using the common rule of thumb that a token is about three quarters of an English word, 200 tokens is roughly 150 words.
A quick worked example. Say a large language model writes you a 600-word product description. That is about 800 tokens, four times the threshold, so the provider should watermark it. A 40-word social caption sits near 55 tokens and falls under the line.
Can anyone check a watermark?
Yes, eventually. Freshfields reports that detection must become interoperable by February 2, 2027, through options such as a public API or a shared detector run by several providers together. Detection tools should be free in principle. Smaller providers with under 1 million monthly users may charge a reasonable fee for heavy use, but regulators, researchers, journalists and fact-checkers always get free access.
What You Will See: The EU AI Icon
For the visible side, the code offers three ready-made icons. One marks fully AI-generated content. One marks AI-modified content. A third, basic icon opens a second layer of detail when you tap or hover.
Freshfields reports that user testing found something useful. Icons with a written word, such as "modified", beat bare symbols on both noticeability and clarity. Companies may use their own design instead, as long as it meets the same placement rules.
- The label must be visible at first exposure, not buried in a caption you have to expand.
- It must stay on screen long enough to read.
- Audio-only content, such as a cloned voice in a podcast ad, needs a spoken disclaimer.
- Satire, fiction and art get a lighter, more flexible disclosure regime.
This matters because the scam playbook leans on fake faces and voices. Labels will not stop a criminal who ignores the law, but they make honest content easier to spot. Our guide to the tells that still give AI scams away covers what to check when a label is missing.
Who Signed Up, and What Ignoring It Costs
The code is voluntary, but signing it gives companies a recognised way to show they comply. The Commission's signatory page, last updated November 18, lists 266 unique organisations. That includes 103 providers and 219 deployers.
Named providers include Anthropic, Google, Meta, Microsoft, Mistral and OpenAI. Deployers include Getty Images, Lenovo and Lufthansa. Google said in July that it was signing and that its SynthID watermark now has five outside partners: Apple, ElevenLabs, Kakao, Nvidia and OpenAI. Google also voiced a worry that overlapping labels and legal notices could confuse people rather than help them.
The fine math
Article 99 of the AI Act sets the penalty for breaking the transparency duties at up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. For smaller firms and startups it flips to whichever is lower.
- A large firm with 2 billion euros in turnover: 3% is 60 million euros, so the ceiling is 60 million.
- A mid-size firm with 100 million euros: 3% is 3 million euros, so the higher figure, 15 million, applies.
- A startup with 10 million euros: 3% is 300,000 euros, and as the lower figure, that becomes its ceiling.
Those are maximums, not automatic bills. Still, they explain why so many companies signed before the rules bit. They also fit a wider pattern of AI rules gaining real penalties, which we covered in how frontier AI safety laws got teeth.
What This Means If You Live Outside Europe
Big AI providers rarely build one product for Europe and another for everyone else. When a company adds signed metadata and a watermark to meet the EU rule, the cheapest path is usually to add it everywhere. So the image you generate in Ohio or Mumbai may well carry the same hidden marks.
Visible labels are less certain. They depend on the deployer, and a US brand advertising only at home has no EU duty to show one. Expect a patchy mix for a while.
Practical steps for you
- If you publish AI images or video for an EU audience, label deepfakes at first view and keep a short note of who reviews your AI-assisted articles.
- If you rely on a tool launched before August, ask the vendor whether it adds machine-readable marks by December 2. If it does not, your output may carry no provenance at all.
- If you want to check a suspicious image, look for a provenance or content credentials panel in your browser or photo app, and treat a missing label as a reason for doubt, not proof.
- If you write with AI help, remember the human review exemption only covers you when a real person takes editorial responsibility.
Frequently Asked Questions
Do AI generated images have to be labeled in the EU?
Yes, in two ways. The AI provider must add a machine-readable mark to every image its system makes. If the image is a deepfake of real people, places or events, whoever publishes it must also show a visible label at first exposure.
What is the EU AI Act deadline for watermarking?
Generative AI tools launched on or after August 2 had to mark outputs immediately. Tools already on the EU market before August 2 have until December 2 to add machine-readable marking.
Does AI written text need a disclaimer?
Only in some cases. Text published to inform the public needs disclosure unless a human reviewed it or holds editorial responsibility for it. Separately, providers are expected to watermark free-form text longer than 200 tokens.
Can AI watermarks be removed?
Some can. Metadata often vanishes when a file is screenshotted or re-saved, and no single technique meets every legal test. That is why the EU code asks providers to combine at least two methods, so one layer can survive when the other is stripped.
What is the fine for not labeling AI content in the EU?
Up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. For small and medium firms the lower of the two applies.
The Label Is Small, the Shift Is Not
AI content labels look like a tiny badge in a corner. Behind it sits a two-part system: invisible marks from the companies that build the models, and visible notices from the people who publish with them. December 2 closes the last gap for the older tools most people use.
The technology is not finished. Watermarks can be stripped, detectors are still being linked up, and the shared detection deadline is not until February 2027. But for the first time, a major market has made "was this made by AI?" a question companies must help you answer. Start looking for the badge, and start asking why it is missing when it should be there.
Written by
Nora Whitfield
Technology & Science
Covers technology and science for Quick Trend Insights, with a focus on what new AI tools, consumer hardware and space missions actually change for the people following them.
Related Articles
View all
Killer Robots: Why UN Talks Fell Short of a Treaty
Killer robots talks at the UN ended without binding rules. Here is how the CCW works, who blocks a treaty, and what meaningful human control would require.

Prompt Injection: Can AI Learn to Ignore Hidden Orders?
Prompt injection tricks AI agents with hidden orders. IBM's new models cut attack success to as low as 0.02%. See how it works and what still protects you.

AI Math Proofs: What Google's Lean Results Prove
AI math proofs just hit 1.8 million machine-checked lines. See what Google's Lean results really prove, what they don't, and where humans still matter.
