Watermark Stress Test
AI Governance · Explainer

EU AI Act Article 50: what "machine-readable marking" actually requires

If your product generates images, text, audio, or video with AI, Article 50 expects those outputs to carry a machine-readable mark that says a machine made this. Here is what the rule asks for, when it bites, and why a hidden watermark alone is a fragile way to meet it.

TL;DR
  • Article 50(2) requires providers of generative AI to mark synthetic outputs in a machine-readable format, detectable as artificially generated or manipulated.
  • The obligation applies from 2 August 2026. A provisional May 2026 agreement (the "AI Omnibus") would give systems already on the market until 2 December 2026 to meet the machine-readable part. Treat both as live and re-verify.
  • Scope is broad. It is not limited to high-risk AI. It reaches any business using generative AI to produce content.
  • The legal bar is that the technique be "effective, interoperable, robust, and reliable." A watermark hidden inside the pixels fades silently under ordinary edits, so watermark-only marking is hard to defend against that bar.
  • Cryptographically signed provenance (C2PA / Content Credentials) is the machine-readable, tamper-evident layer that answers the reliability test, with watermarking as one corroborating signal.

Article 50 is the transparency article of the EU AI Act. It is short, it is not restricted to the high-risk tier everyone spent 2024 worrying about, and its marking obligation lands on a much wider set of products than most teams assume. If you ship a feature that generates synthetic media, this is the clause that now expects a machine, not just a person, to be able to tell.

What Article 50 asks for

The part that matters for content marking is Article 50(2). It puts the obligation on providers of generative AI systems, including general-purpose systems: the outputs they produce (synthetic audio, image, video, or text) must be marked in a machine-readable format and detectable as artificially generated or manipulated. A separate obligation, in Article 50(4), sits on deployers who publish deepfakes or AI-generated text on matters of public interest, and is about human-facing disclosure. The two are often conflated. This piece is about the first one: the machine-readable mark on the output itself. The practical guide from the EU Artificial Intelligence Act project and the European Commission's own transparency FAQ are the primary references.

The dates that matter

Three dates are worth writing down, and one caveat is worth writing in bold.

Caveat, in bold because it is the important one: the Omnibus was provisional at the time of writing and regulatory dates in this file can move. Verify the current status against the official sources before you plan a release around any of them. Nothing here is legal advice.

The bar you actually have to clear

Article 50(2) does not just say "mark it." It says the technical solution must be "effective, interoperable, robust, and reliable, taking into account the state of the art." Those four words are the whole test, and they are where a marking strategy either holds up or quietly does not:

The last two are the trap for anyone planning to lean on a hidden watermark. An in-band watermark, the family Google DeepMind's SynthID belongs to, hides a recoverable signal inside the pixels or tokens. It survives a screenshot. It does not reliably survive someone editing the file: a frequency-notch filter, a denoise pass, a re-encode, or a paraphrase can attenuate the signal while the content still looks and reads fine. Worse, the failure is silent. A stripped mark announces nothing, so a marking method can pass on an untouched file and then fail invisibly after one routine re-save. That is a hard thing to call "reliable."

You do not have to take that on faith. Our interactive lab injects a synthetic watermark and lets you watch its detection score collapse under exactly these edits, and the companion white paper, Why AI watermarks break, works through the mechanism.

Your machine-readable options

There are two families of machine-readable marking, and they fail in opposite ways.

ApproachWhat it isWhere it struggles
In-band watermarkHides a recoverable signal inside the media (SynthID and relatives)Fades silently under ordinary edits; detection is a probability, and absence proves nothing
Signed provenance (C2PA)Attaches cryptographically signed metadata bound to the asset (Content Credentials)A recipient has to check it, and the credential can be stripped, but stripping or tampering is detectable rather than silent

The key difference for the reliability bar: a watermark fails quietly, and a signed credential fails loudly. When tampering breaks a C2PA signature, "invalid" and "no credential" become explicit, checkable states rather than a probability that faded. That auditability is exactly what a regulator, or your own compliance team, needs to stand behind a claim.

What a defensible marking posture looks like

The pragmatic answer is not to pick one. It is to layer, so that no single technique carries the whole load:

Document the lawful basis and retention for anything you log along the way. Again, not legal advice: this is an engineering-readiness view, not a substitute for counsel at the point you formalize a compliance claim.

A short readiness checklist

Frequently asked questions

When does Article 50 apply?

Article 50 transparency obligations apply from 2 August 2026. A provisional May 2026 agreement (the AI Omnibus) would give systems already on the market until 2 December 2026 to meet the machine-readable marking part of Article 50(2). Both are moving targets. Verify current status against official sources.

Does Article 50 only apply to high-risk AI?

No. The transparency obligations are not restricted to the high-risk tier. They reach the situations the Article covers, which in practice includes any business using generative AI to produce synthetic images, text, audio, or video.

Does watermarking satisfy the machine-readable marking requirement?

It can contribute, but a watermark-only strategy is hard to defend against the requirement that the technique be effective, interoperable, robust, and reliable. In-band watermarks fade silently under ordinary edits, so signed provenance (C2PA / Content Credentials) is the tamper-evident layer that answers the reliability test, with watermarking as one corroborating signal.

Is this legal advice?

No. This is an engineering-readiness explainer, not legal advice, and the regulatory dates here can change. Confirm the current position with the official EU sources and your own counsel before relying on it.