Article 50 applies from 2 August: what it actually says about AI transparency

The marking is a barcode, not a certificate

Think of the machine-readable marking in Article 50(2) as a barcode on a tin of beans. A scanner reads it, people barely notice it, and it says what is in the tin, not whether the beans are any good. The marking tells you that an AI made the image or the audio. It does not tell you when, by whom or on what instruction, and the law does not say how to protect it from being stripped.

That is the gap between transparency and evidence. One tells you something was made by AI. The other lets you show, later, that a specific thing was made at a specific time and has not been changed since.

What Article 50 requires

Article 50 of the AI Act ^[1] is titled "Transparency obligations for providers and deployers of certain AI systems". It sets out three main duties, all of which apply from 2 August 2026.

First, providers of AI systems that interact with people must design them so that people are informed that they are talking to an AI, unless it is obvious from the circumstances ^[1]. If you run a chatbot, that applies to you.

Second, deployers of an emotion recognition system or a biometric categorisation system must tell the people exposed to it that the system is in use ^[1].

Third, providers of AI systems that generate images, audio or video must ensure the outputs are marked in a machine-readable format and detectable as artificially generated ^[1]. The same goes for text meant to inform the public, if no human has reviewed it or takes responsibility for publishing it.

And the rule reaches backwards. Systems that were already on the market before 2 August 2026 must mark their outputs by 2 December 2026 ^[2].

The penalties

Failure to meet the transparency obligations in Article 50 can cost up to €15 million or 3% of total worldwide annual turnover from the preceding financial year, whichever is higher ^[1]. For small and medium-sized enterprises, including start-ups, the lower of the two amounts applies.

That makes it one of the more expensive rules in the AI Act.

What the rule does not say

Article 50 says providers must mark AI-generated content. It does not say how robust the marking must be, how to protect it from being removed, or what happens when the marking is lost along the way.

It does not require a record of what was generated, when or by whom. It does not require a blockchain, a hash, a seal or a timestamp. Those might be sensible steps for other reasons, but Article 50 does not ask for them.

That is worth knowing, because the market is full of claims that the AI Act requires tamper-proof records. It does not. The transparency duty is about the marking, and the marking is about telling people what they are looking at.

The evidence question

The transparency question is: did the user know they were talking to an AI, or looking at AI-generated content?

The evidence question is: can you show what your AI generated, when it generated it, and that the record has not been edited since?

Article 50 answers the first. It does not touch the second. That matters, because disputes often turn on what was said or shown, and a marking that can be stripped is not much help a year later when someone asks where an image came from.

What a digital witness does

Here, the thing to seal is a record of what was generated: the output itself, or a fingerprint of it, together with the time, the system version and the user reference. Sealing takes a fingerprint, not the content, so the material never leaves your systems.

When someone questions what was generated, you share the record and they check the fingerprint. If it matches, the record has not changed since it was sealed. Change a single pixel or a single word and it will not match.

Anyone you share the record with can check it, without asking you or anyone else for permission. That is the point: the check does not depend on trust.

It is done through WitniumChain, using the REST API, the SDK or the MCP server. The content stays with you. The first 2,000 witnesses a month are free, and the documentation shows you how to get started.

How it would work

1. Your system generates an image, and at the same moment it seals a record: a fingerprint of the image, the timestamp, the model version and the user reference.
2. You get a witness ID back and store it with the image.
3. When the image is questioned months later, you share the record. The recipient calculates the fingerprint again and checks it against the sealed one. If it matches, nothing has changed.

What it does not do

A digital witness does not replace the marking duty in Article 50. The marking goes on the content; the witness seals a record of what was made.

It does not prove that the content is lawful, accurate or safe. It shows that a record existed at a given time and has not been altered. Whether the thing recorded is any good is a different question.

And it does not prevent someone from stripping the marking. It gives you your own record, so that when the marking is gone and someone asks where the image came from, you can show them.

What this means for you

If you provide or deploy an AI system covered by Article 50, ask yourself:

• Do our users know when they are interacting with an AI, and is that information clear from the start?
• Is our generated content marked in a machine-readable format, and have we checked that the marking survives the formats and platforms where the content will be shared?
• If someone disputes what our AI generated six months ago, do we have a record we can show, and can we prove it has not been edited?
• Have we documented how we meet the transparency requirements, and have we tested the marking?

The first two are about compliance. The last two are about evidence.

What to watch

The AI Office is expected to publish guidance on the technical standards for machine-readable marking. Until then, providers are working with what the rule says and what existing standards, such as C2PA, offer.

The enforcement date is 2 August 2026 for new systems and 2 December 2026 for existing ones. Supervisory authorities have said transparency will be an early focus, so assume the rules will be tested soon after they apply.

And if you work with generated content, watch for the first disputes where someone claims an image or a piece of text was AI-generated and the provider says it was not, or the other way around. Those cases will show what counts as evidence.