Telling people they're talking to an AI: Article 50 in practice

Who must inform whom

Article 50 applies to providers and users of AI systems that interact with people or generate content ^[1]. If someone is talking to a chatbot, they must be told that it is an AI unless it is obvious from the circumstances ^[1]. The notice must be clear and distinguishable, but the article does not say what form it should take.

There is an exception. Law enforcement can use an AI system without informing people if telling them would hinder the detection, investigation or prosecution of a criminal offence ^[1]. That is the only exception the article names.

What counts as obvious is left to interpretation. A website with a heading that says "Chat with our AI assistant" might pass. A chatbot window that looks like a regular customer service chat might not.

Marking AI-generated content

Providers of generative AI must ensure that outputs are marked as artificially generated or manipulated, in a machine-readable format ^[1]. The article does not say which format, how detailed the marking must be, or what happens if the marking is stripped or lost along the way.

The obligation falls on the provider, the one who places the system on the market. If a user creates a deepfake with a model, the provider's marking duty applies to the technical output. The user has a separate duty under Article 50.4: to disclose that the content is AI-generated or manipulated when publishing it to inform the public, unless it is obvious or a person has reviewed it and takes editorial responsibility ^[1].

Until 2 December 2026, only systems placed on the market after 2 August 2026 must mark their outputs. From 2 December, the duty extends to generative AI systems already on the market before that date ^[2]. That gives providers of existing systems four months to add the markings.

What a good disclosure looks like

A good disclosure is honest and early. If a customer is about to type a question into a chat window, they should know they are talking to an AI before they hit send.

One tested form is a line at the top of the chat that says something like "You are chatting with an AI assistant. It can make mistakes". Some systems repeat the notice in their first reply. That is not required, but it does no harm.

For generated content, the marking must be machine-readable. A watermark in an image or metadata in a file meets that description. Plain text in a corner does not. The customer-facing label can be as simple as "AI-generated" or "Edited with AI".

The article is clear on the goal and vague on the method, which leaves room for providers to choose a solution that fits their workflow. It also leaves room for disagreement when someone questions whether a disclosure was clear enough.

What this means for you

If you run a chatbot, check that users see a notice before they interact with it. If the system generates text, images, audio or video for public use, add a machine-readable marking. If your system was released before 2 August 2026 and generates content, plan to have markings in place by 2 December.

Article 50 does not require sealing, hashing, blockchain or any particular technology ^[1]. It asks for information and markings. How you give that information and how you mark the content is up to you, within the requirement that markings must be machine-readable.

That said, there is a separate question Article 50 does not address: what happens when a customer disputes what your AI said or generated. The customer's screenshot or saved file becomes their evidence. If you do not have your own record of what the system produced, the conversation starts with you on the back foot.

Keeping your own record of what the AI said

When a customer disputes what your chatbot promised, their screenshot often becomes the evidence. If you have no matching record, or cannot show that yours is unchanged, the conversation starts with you explaining rather than checking.

The thing to seal here is each answer as it leaves the system: the text itself, the user identifier, the timestamp and the notice you showed under Article 50. Sealing happens in the same moment the answer is written to your log or sent to the customer, whichever you choose. What gets sealed is a fingerprint of that record, not the record itself, so the conversation stays in your systems.

Anyone you share the record with later can check it against the sealed fingerprint. If it matches, nothing has changed since the answer was given. Alter a single word and the check will fail.

This doesn't prove the answer was right. It doesn't settle whether your company is bound by what the chatbot said. What it does is show that the record you produce is the one written at the time, not one edited since.

Your system sends the fingerprint to WitniumChain through the REST API, the SDK or the MCP server and receives a witness ID. To verify, the other party calculates the fingerprint again and compares the two. The first 2,000 witnesses each month are free, and the documentation walks through integration in under an hour.

Say a regulator or an ombudsman asks what your AI told a customer on a given day six months ago. You produce the conversation, the witness ID and the Article 50 notice that was shown. They run the check. The fingerprints match. That closes the question of whether the text has been altered, and lets the substantive discussion begin with both sides looking at the same record.

How it would work

1. The chatbot sends an answer to the customer and writes it to your database or log.
2. At the same moment, your backend seals a fingerprint of that record: answer text, user ID, timestamp and the Article 50 disclosure.
3. Months later, when the answer is questioned, you produce the record and the witness ID. The questioner checks the fingerprint and sees the record is unchanged.

What to watch

On 2 December 2026, generative AI systems placed on the market before 2 August 2026 must start marking their outputs as artificially generated, in a machine-readable format ^[2]. If your system went live before August and you have not yet added markings, that is the deadline.

The AI Act's high-risk requirements do not apply to most chatbots. If your chatbot is used to screen job applicants, assess creditworthiness or make similar decisions about people, it may fall under Annex III. Those requirements, including logging under Article 12, apply from 2 December 2027 ^[3].

For now, Article 50 is the rule that applies to the widest range of AI systems, and it has applied since August. The obligation is straightforward: tell people when they are talking to an AI, and mark what your AI generates.