Come 2 August 2026, the European Union is officially calling time on AI pretending to be human. The EU AI Act will force creators to embed digital watermarks into machine-made output, while publishers must display clear warnings on public chatbots and synthetic content. The European Commission is even offering a Code of Practice, which is essentially a golden ticket to immunity for companies that promise to behave.
The mandate covers any company offering services in the EU, regardless of its country of incorporation. Global platforms across the UK, the US and beyond fall directly into scope, using the familiar territorial reach of GDPR. The real debate is about technical feasibility and whether it can achieve its core purpose.
What The Watermarking Mandate Means In Practice
Instead of enforcing one specific technology, the rule establishes functional goals. Marks must prove reliable, durable and interoperable across systems using multiple technical layers. To satisfy regulators, the ideal setup uses a two-pronged approach: cryptographic metadata embedded via the C2PA standard, paired with pixel or text watermarks applied during generation.
C2PA metadata works like a signed provenance certificate attached to the file. A compatible reader can verify that the content was AI-generated and trace it back to the system that produced it. The hurdle is unavoidable: C2PA metadata is trivially stripped when a file is re-saved, uploaded to a social platform that removes metadata or screenshotted. It’s excellent for provenance verification in controlled environments and useless once the content has been through a standard content pipeline.
Pixel-level watermarks, like Google’s SynthID approach, embed signals into the image data in ways designed to survive compression, resizing and colour adjustment. These are more durable than metadata but can be degraded by adversarial processing, heavy re-encoding or the oldest workaround of all: photographing a screen with a different camera. Text watermarking for longer outputs is also required. The guidance acknowledges it as technically harder and less durable than image watermarking.
The third layer is server-side: providers are expected to maintain hashing and logging of AI outputs, creating an audit trail even when embedded signals have been stripped or defeated. This is the fallback that enforcement will rely on when the first two layers fail.
More from Artificial Intelligence
- What’s The Difference Between Open Weight And Open Source AI?
- NVIDIA’s AI Tool Can Spot Deepfake Videos – Is The Company Starting Too Many Projects At Once?
- Frontier AI Or Budget Models – Which Strategy Makes Sense For Your Business?
- Nvidia Has Its Own AI Model – How Will This Affect Competition In The Industry?
- AI Is Supposedly “Replacing” Writers. So Why Is Anthropic Hiring Editors For $295,000 A Year?
- Should AI Literacy Be Taught Alongside Math And English?
- Should AI Have An Appreciation Day When Other Industries Don’t?
- The Rise Of AI-Powered Product Governance: From Institutional Knowledge To Scalable Decision-Making
Who Must Act And How To Prepare
Every affected business faces a clear two-track challenge. Providers, meaning companies that create or offer AI systems, must implement the technical marking measures. Platforms and publishers using artificial media face strict labelling rules for deepfakes and public-interest text. The regulation exempts deployers from this requirement only if a human editor actively reviews and verifies the material before publication.
For businesses in scope, the compliance structure the Code of Practice expects covers four areas. C2PA manifests on image and video outputs. Pixel or text watermarks where technically feasible. Server-side hashing and logs for all EU-facing outputs. Clear chatbot disclosure text on any interactive AI surface. Signing the Code offers a public safe harbour and is expected to reduce early scrutiny from national enforcement authorities.
Enforcement falls to individual national authorities, who will judge compliance on a holistic basis instead of inspecting every technical output for an unremovable mark. The question regulators will ask is whether a provider made a good-faith, best-available implementation across multiple layers. Small providers using open-source tooling face a more difficult evidentiary challenge: if watermarks are absent or stripped, they need server-side logs to demonstrate compliance, and many won’t have built that system.
Can Digital Watermarks Really Stop Deepfakes?
The truth is that it increases friction and boosts detection in everyday cases, but still doesn’t make deepfakes impossible.
The underlying weakness lies in access to open-source models. Anyone running local inference on an open-source image or video model has no obligation to implement watermarking and no technical mechanism forcing them to. The EU’s territorial reach covers service providers operating in the EU market, but has no reach over a locally run model on a private machine.
Getting around this offline is simpler still. Photograph a screen showing AI-generated content with a different device. Re-record AI-generated audio through a microphone. The resulting file carries no embedded metadata, no pixel watermark and no provenance trail. These methods don’t require advanced skills. Anyone who wants to bypass detection can easily use them.
The regulation successfully creates a deterrent and detection system for most AI content on commercial platforms. It offers a vital layer of accountability which allows clear tracking when output comes from a compliant provider. The regulation is better understood as infrastructure for accountability than as a technical barrier to misuse.
The cat-and-mouse game between watermark protection and removal tactics will continue no matter what the law states. The Code explicitly acknowledges this, building in continuous improvement requirements instead of treating current tools as a solved problem.
