Why AI Watermarking Was Introduced, and What It Means for Everyone Downstream

Providers of generative systems must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. This is the technical watermarking duty, discussed in how AI watermarking actually works.

Deployers, meaning anyone using these tools, have disclosure duties. If you generate or manipulate an image, audio or video that is a deep fake, you must disclose that it is artificially generated. The same applies to AI-generated text published to inform the public on matters of public interest, unless the content has undergone human editorial review with someone holding editorial responsibility.

That editorial exemption is significant for publishers. It means a newsroom that reviews and takes responsibility for AI-assisted text is not obliged to label it, while an automated content operation is.

Chatbots must also disclose that a user is interacting with a machine, unless it is obvious.

The honest criticism

Two objections deserve to be taken seriously rather than dismissed.

The mandate outpaces the technology. Text watermarking is genuinely weak, particularly for short passages, and paraphrasing degrades it. Watermarks can be stripped. Metadata is routinely removed by platforms. A legal requirement to make content detectable does not create a technology that reliably does so, and several commentators have noted that the rule arrived ahead of the capability.

It regulates the compliant. Large providers will comply. Open-weight models run locally, and systems operated by actors who do not care about EU fines, will not. The people most likely to produce an election deepfake are the least likely to watermark it.

Both criticisms are correct. Neither is an argument for doing nothing, but they do define what success can look like.

What it actually achieves

Not the elimination of deepfakes. What it does is shift the default.

If the overwhelming majority of synthetic content carries a signal, then the effort required to produce untraceable synthetic media rises. It stops being a two-minute exercise for anyone with a phone and becomes something requiring deliberate circumvention. That will not stop a determined state actor, and it will substantially reduce the casual volume, which is most of the volume.

The second achievement is the inversion described in our technical piece: over time, the useful question becomes whether content can prove it is genuine, rather than whether it can be shown to be fake. Newsrooms, courts and platforms can build on a positive signal in a way they never could on a negative one.

What it means for you

• If you publish: understand the deployer duties. If you use generative tools for public-interest content, either apply genuine editorial review and take responsibility, or label it.
• If you create: expect provenance metadata to become normal, and expect platforms to start surfacing it. Adobe's Content Credentials and similar systems are already there.
• If you read: remember the asymmetry. A watermark tells you something was AI-generated. The absence of one tells you nothing whatsoever, and treating unmarked content as verified is the mistake this entire system is most likely to cause.

The rules will not restore the situation where a photograph was self-evidently true. That is gone. What they might do is build the infrastructure for deciding what to trust in a world where it is not, which is a more modest ambition and the only one available.

The wider regulatory timetable is in our coverage of the transparency rules.