Article 50: when do you have to *tell people* you used AI?

The question nobody answers clearly

You run a newsroom, a brand, an agency. You try a generative model to draft a press release or sketch out a brief. At some point someone asks: do we have to tell people?

The AI Act has an answer, in Article 50. It's just rarely explained in plain terms.

Here's the short version. If you publish AI-generated text to inform the public on matters of public interest, you must disclose it—unless a person has reviewed or edited the text and takes editorial responsibility for it ^[1]. If you create or manipulate images, audio or video that could be mistaken for the real thing, you must say it's artificial ^[1]. Both duties apply from 2 August 2026 ^[2].

The bit about editorial responsibility is the one people trip over. Think of it like bylines in a newspaper. A journalist can use spell-check, prompts, search engines, whatever. As long as someone puts their name on it and stands behind the facts, the piece is theirs. The same goes here: if a human has reviewed the AI output and is accountable for it, you've met the rule ^[1].

What Article 50 actually says

Article 50 covers transparency for AI that interacts with people or generates content. It has four parts ^[1].

First, providers of AI systems must make sure people know when they're talking to an AI, unless it's obvious from context. That's the chatbot notice.

Second, providers must mark AI-generated content in a machine-readable way. That marking sits with whoever built the model.

Third, if you deploy an emotion-recognition system or a biometric categorisation system, you have to tell the people affected.

Fourth—and this is where most questions land—if you use AI to create or manipulate images, audio, video or text, you have duties as a deployer. Deepfakes must be disclosed as artificial. AI-generated text published to inform the public on matters of public interest must also be disclosed, unless a human has reviewed it and assumes editorial responsibility ^[1].

The carve-out matters. Say a campaign team drafts a press release with a model, an editor checks the facts and signs off, and the organisation publishes it as its own statement. That's fine. The human took responsibility. You don't need a disclaimer at the bottom saying "partly AI-generated".

But if the same text goes out with nobody checking it, or if it's published in a way that hides the fact a machine wrote it, that's where the rule bites.

Deepfakes and the artistic exception

A deepfake is content that resembles real people, objects, places or events and would wrongly appear authentic ^[1]. If you make one, you must say so.

There's an exception for clearly artistic, creative, satirical or fictional works. In those cases you just need to disclose, in an appropriate manner, that such content exists—without spoiling the work itself ^[1]. A credit at the end of a film would do it. A pop-up halfway through would not.

The test is whether a reasonable person would be misled. A mock-up for an internal presentation isn't a deepfake. A manipulated video that looks like a real press conference is.

The evidence question

So you disclose. You put a note on the page, a line in the credits, a box at the top of the article. Six months later someone questions it. A regulator asks how you know the disclosure was there. A rights holder wants to see what you actually published. Your own lawyer wonders whether the version that went live was the one the editor signed off.

That's the gap. The rule says to disclose. It doesn't say how you prove you did.

How a digital witness fits here

This is where a witness makes sense, and where it doesn't replace anything else.

What you seal is the published version, with its disclosure, at the moment it goes live. If it's a press release, you seal the final text. If it's a video, you seal the file with the credit. If it's a campaign landing page, you seal a snapshot of the HTML.

You seal it by sending a fingerprint—not the content—to WitniumChain through the REST API, the SDK or the MCP server, and you get a witness ID back. The content stays with you. Later, when someone questions what the page said or whether the disclosure was there, anyone you share the file with can calculate the fingerprint again and check it against the sealed one. If it matches, nothing has changed since you sealed it.

What the witness shows is that the version you published, with whatever notice or credit it had, existed at that time and hasn't been altered. What it doesn't show is whether the disclosure was adequate, whether the work counts as clearly satirical, or whether a human actually reviewed the text. Those are judgement calls, and a seal doesn't make them.

But it does give you a record that isn't just your word. If a supervisor asks what ran on 15 March, you can show them, and they can check it themselves.

The first 2,000 witnesses a month are free. Documentation is at witniumchain.com/docs.

What this means for you

If you publish content, here are the questions to ask now.

Does any AI-generated text you publish inform the public on matters of public interest? If yes, and if nobody reviews it and takes responsibility, you need to disclose that.

Do you create or use images, audio or video that could be mistaken for real? If yes, and if the work isn't clearly artistic or fictional, you need to say it's artificial.

Can you show, six months later, what you actually published and whether the disclosure was there? That's not a legal requirement under Article 50, but it's a sensible one. A supervisor or a court won't take your word for it.

If the answer to any of those is unclear, sit down with whoever handles your content, your campaigns and your legal questions, and decide now. The rule applies from 2 August 2026 ^[2]. That's not far off, and changing your process later is harder than getting it right the first time.

What to watch

From 2 August 2026, the disclosure duties in Article 50(4) apply ^[2]. Providers' duties—the chatbot notice and the machine-readable marking—have the same date.

The European Commission's AI Office is expected to publish guidance on what counts as adequate disclosure for different types of content. Watch for that.

National supervisors will enforce the rule. In Sweden, that's the Integritetsskyddsmyndigheten (IMY) for most questions touching personal data, and potentially other authorities depending on the sector ^[3]. Penalties go up to €15 million or 3% of global annual turnover, whichever is higher, and for small and medium enterprises the lower amount applies ^[2].

If you work in advertising, note that some jurisdictions already have self-regulatory codes on AI disclosure. Article 50 sets a floor, not a ceiling.