The AI Act says mark AI content, not how to prove you did

What Article 50 actually says

Article 50 of the AI Act is about transparency ^[1]. Providers of generative AI systems must make sure outputs are marked in a machine-readable format, showing that the content was generated by AI. The rule applies from 2 August 2026, and after the Digital Omnibus it covers systems already on the market as well, from 2 December 2026 ^[2].

The machine-readable marking is for machines to read. Think of it as a barcode on a tin of beans: the shop's scanner sees it, you barely notice it. The marking tells platforms, fact-checkers and automated tools that a piece of content came from AI.

Deployers have their own duties under Article 50(4) ^[3]. If you use AI to create or manipulate images, audio or video that could be mistaken for real—a deepfake—you must tell people it's artificial. If you publish AI-generated text to inform the public and no human has reviewed it, you must disclose that too. But the machine-readable marking itself is the provider's job, not yours.

The bit the Act leaves out

Article 50 says mark the output. It doesn't say how you prove, six months later, that the marking was there when the content left your system.

Say a platform finds unmarked content and traces it back to your service. You say it was marked when you sent it. The platform says it wasn't. Who keeps the record?

Or say a rights holder claims your system generated an image without marking it, and you need to show what your system actually produced on that day, complete with its marking. Where's that record?

The Act asks for the marking. Whether you can prove you did it is a separate question, and one that tends to come up later, when the original file has moved through three platforms and two formats.

What providers are doing

Some are embedding technical markings like C2PA metadata or invisible watermarks. Both are good ideas. Both can still be stripped, either deliberately or by a platform that doesn't preserve them.

Logs help. A log that says "generated image XYZ at 14:32 on 3 March, marked per Article 50" is useful, until someone asks whether the log itself has been edited. When the people who keep the logs could also change them, the log is easier to challenge.

That's where the evidence question lands. The law says what to do. Proving you did it is your problem.

How a digital witness fits here

A digital witness seals a record of what was generated, when and by which system, together with confirmation that it was marked. The content itself stays with you. What gets sealed is a fingerprint: a short string that changes completely if a single character in the record changes.

Say your generative system produces an image. At the moment of generation, you seal a record: the image fingerprint, the timestamp, the model version, the marking applied and the request ID. That record is posted to WitniumChain, which adds it to a chain where any later change would show. You get a witness ID back.

Months later, a platform questions whether the image was marked. You share the sealed record and the witness ID. They calculate the fingerprint of the record you've given them and check it against the sealed one. If it matches, the record hasn't changed. If it doesn't, something has.

The witness shows that the record of what you generated existed at that time, unchanged since. It doesn't replace the marking itself, and it doesn't prove that the marking survived every platform the content passed through afterwards. What it does is give you your own checkable record of what left your system, marked as required.

Your system sends a fingerprint of the record, not the record itself, to WitniumChain through the REST API, the SDK or the MCP server, and gets a witness ID back. To check, anyone you share the record with calculates the fingerprint again and compares it with the sealed one. The first 2,000 witnesses a month are free ^[4].

How it would work

1. At generation, seal a record: image fingerprint, timestamp, model version, marking applied.
2. Store the witness ID with the record.
3. When questioned, share the record and the witness ID. Anyone can check the fingerprint matches.

What this means for you

If you provide a generative AI system, ask yourself:

• When we mark outputs per Article 50, where's the record that we did?
• If a platform or a supervisor questions an unmarked file six months from now, could we show what our system generated and that it was marked?
• Who keeps that record, and can they edit it?

The marking duty is clear. The evidence of having done it is up to you. A witness per generation, sealed at the moment the content is created, means the record is yours and nobody has to trust that you haven't quietly changed it since.

What to watch

2 August 2026: Article 50 applies to new generative AI systems placed on the market ^[1].

2 December 2026: Existing generative AI systems must also mark their outputs in a machine-readable format ^[2].

The Commission's AI Office is working on technical guidance for marking methods. Watch their updates on the AI Act Service Desk.