EU AI Act Labelling Rules Explained: What Article 50 Requires from Creators and How SynthID and C2PA Fit In
From August 2, 2026, Article 50 of the EU AI Act requires providers to mark AI-generated content in a machine-readable way and deployers to disclose deepfakes. What the rules actually require, the penalties, how SynthID and C2PA help, and the practical gaps creators should know about.
On August 2, 2026, the transparency obligations in Article 50 of the EU AI Act came into force. They require providers of AI systems that generate synthetic audio, images, video or text to mark outputs in a machine-readable format so they can be detected as artificially generated, and require deployers who publish deepfakes to disclose that the content is artificial. Under Article 99, breaches of these obligations can lead to fines of up to €15 million or 3 percent of worldwide annual turnover, whichever is higher.
The rules sound straightforward. In practice, there is a meaningful gap between what the law asks for and what current technology can reliably deliver, and creators, developers and publishers of AI-generated work need to understand both.
What the EU Rules Actually Require
Article 50 of the AI Act applies from August 2, 2026 under Article 113. It contains several separate duties that fall on different parties:
- Providers of generative AI systems (Article 50(2)) must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Their technical solutions must be effective, interoperable, robust and reliable “as far as this is technically feasible,” taking into account the type of content, costs and the state of the art. This does not apply to assistive functions for standard editing or tools that do not substantially alter the input.
- Deployers of deepfakes (Article 50(4)) must disclose that image, audio or video content has been artificially generated or manipulated. For evidently artistic, creative, satirical or fictional work, the disclosure can be limited so it does not hamper the display or enjoyment of the work.
- Deployers of AI-generated text published to inform the public on matters of public interest must disclose that it is AI-generated, unless it has undergone human review or editorial control and someone holds editorial responsibility.
- Providers of chatbots and similar systems (Article 50(1)) must make sure people know they are interacting with an AI, unless that is obvious.
In short, the law combines machine-readable marking by the tools that generate content with human-facing disclosure by those who publish certain kinds of content. Machine-readable marks let platforms detect AI content even when no visible label is present.
The aim is broader than consumer protection: clear marking and disclosure are meant to make it harder for synthetic content to pass as authentic in news and public debate.
Enter SynthID: The Watermark That Survives
Google’s SynthID is currently the most prominent technology trying to make watermarks stick. Unlike metadata-based approaches that can be stripped by simply screenshotting an image or re-encoding a video, SynthID embeds its signal directly into the pixels of an image or the waveform of audio. The watermark is imperceptible to humans but detectable by a specialised decoder, and Google says it is designed to remain detectable after common modifications such as cropping, compression and resizing.
Here is how it works under the hood. SynthID modifies the pixel values of an image in a pattern that is statistically indistinguishable from random noise to the human eye but forms a recognizable signature to the decoder. For images, the watermark is embedded during the generation process itself — the AI model learns to produce images that contain the signal natively, rather than having a watermark applied as a post-processing step. For audio, it works similarly by shaping the waveform during synthesis. This baked-in approach is what makes SynthID harder to remove than a traditional overlaid watermark: you cannot strip it without fundamentally altering the image in ways that degrade quality.
The scale of the problem is large. Generative tools can produce realistic images, audio and video in seconds, so manual review cannot keep up, and automated detection depends on marks that tools embed at the point of creation.
Watermarking works best at scale when many generators participate. A mark from one vendor only helps detect that vendor’s content, so a shared or interoperable approach across major platforms matters as much as the technical sophistication of any single watermark. Article 50’s requirement that solutions be interoperable reflects this.
No watermark is unbreakable. Determined adversaries can try to strip or degrade a signal through adversarial perturbations or by regenerating content through another model. Robustness to everyday edits is still valuable, because most casual sharing involves compression, resizing and cropping rather than deliberate attacks.
The Two Standards Problem
There is a second approach to AI content labeling, and it works completely differently. C2PA, or the Coalition for Content Provenance and Authenticity, is an Adobe-led standard that attaches cryptographically signed metadata to media files. When you generate an image in Adobe Firefly or edit a photo in Photoshop with Content Credentials enabled, the file carries a tamper-evident record of its origin: where it came from, what AI tools were used, and what edits were made.
C2PA and SynthID are not competitors so much as answers to different questions. C2PA answers “where did this file come from?” It works great when the metadata chain remains intact — when a file is shared on platforms that support C2PA and viewed in software that can read the credentials. It breaks when someone takes a screenshot, which discards all metadata.
SynthID answers “is this content AI-generated?” It is designed to survive common transformations that strip metadata, but on its own it carries far less provenance detail than a full C2PA record. A combined approach, where invisible watermarks support detection and C2PA handles attribution, is the direction much of the industry is heading.
What the Rules Actually Miss
The EU AI Act’s labeling requirements face two problems that no watermarking technology can solve.
First, the rules only apply to content generated by compliant systems. A bad actor who wants to create a deepfake of a politician does not use Google’s Imagen with SynthID enabled. They use an open-source model running on a local GPU, or a tool hosted in a jurisdiction that does not enforce the EU rules. The labeling requirement creates a compliance burden for legitimate creators while doing nothing to stop intentional deception. The people who are most likely to misuse AI-generated content for disinformation are also the least likely to use tools that voluntarily embed watermarks.
Second, detection is only as good as the platforms that enforce it. SynthID requires a decoder to check for the watermark. If social media platforms do not run that decoder on every uploaded image — and they currently do not — a watermarked image and an unwatermarked one are indistinguishable to the average user scrolling through a feed. The EU rules mandate that content be labeled at the point of creation, but they do not mandate that platforms verify those labels or check for watermarks at the point of distribution. That gap between creation and distribution is where most disinformation does its damage.
There is a third issue that gets less attention: watermarking works against the grain of how creative professionals actually use AI tools. An artist who uses AI to generate a base image, then spends hours painting over it in Photoshop, has created something that is neither purely AI-generated nor purely human-made. SynthID can detect the AI component even after heavy editing. But should that hybrid work be labeled as AI-generated? The law does not have a clear answer, and the ambiguity puts creators in a difficult position.
What Creators and Developers Should Do Now
If you publish AI-generated images, video, or audio — especially if you have users or audiences in the EU — here is the practical starting point.
Use tools that support SynthID or C2PA. Google’s image models embed SynthID, Adobe Firefly attaches Content Credentials, and other major providers are adopting C2PA or their own marking. If your workflow involves open-source models like Stable Diffusion or Flux, you will need to add labeling yourself, either through post-processing tools or by selecting a distribution platform that applies watermarks automatically. Some hosting services and CDNs are beginning to offer automatic C2PA signing as a feature — worth checking if your platform supports it.
Label visibly even when the law does not require it. The EU rules are the floor, not the ceiling. Visible labels — a “Generated with AI” badge or a note in the description — build trust with audiences and reduce the risk of your work being misused or misinterpreted. They also future-proof you against regulations that are likely to tighten over time. Many stock and publishing platforms already ask contributors to flag AI-generated submissions, so visible labelling is becoming a normal expectation.
Audit your pipeline for gaps. If you use multiple AI tools in a workflow — for example one tool for generation, another for editing and a third for upscaling, check whether each step preserves or strips the watermarks and Content Credentials added earlier. A mark embedded at generation is of little use if a later step removes it. This kind of pipeline audit is tedious but it is the only way to know whether your compliance measures actually survive through to the final output.
Assume the technical solutions are incomplete. Watermarks can be stripped. Metadata can be discarded. Even SynthID, as robust as it is, is not a guarantee. Treat labeling as one layer of a broader approach to transparency — alongside clear disclosures, audience education, and responsible publishing practices. The best defense against AI disinformation is not any single technology. It is a combination of tools, norms, and a public that knows to ask where content came from.
The Bottom Line
The EU’s labeling rules are the most significant AI transparency regulation to date, and they will force real changes in how AI-generated content is created and distributed. Google’s SynthID and the C2PA standard provide workable technical tools for compliance. But the gap between regulatory aspiration and technical reality remains wide, and closing it will require platforms, tool developers, and creators to do more than the minimum the law requires.
For now, the practical advice is boring but correct: label your AI-generated content visibly, use tools that embed watermarks when you can, and do not assume that a watermark makes a deepfake problem go away. The rules are a start. They are not a solution.
Sources and further reading
Our analysis: Article 50 of the AI Act sets the transparency duties, and the Commission’s code of practice on marking and labeling fills in the details. Creators should track both, as well as each platform’s labeling tools.