When people should know it is AI

People should know it is AI, and someone has to tell them, in four cases: a direct chat or voice reply, a deepfake, a public-interest text, and emotion recognition or biometric categorisation. The tool name does not decide.

In a direct exchange with a system built to respond, they should be told unless it is already obvious they are dealing with AI. The company that developed the system, or had it developed, and placed it on the market or put it into service under its own name designs that notice. A company that only uses a vendor chatbot is not automatically that company.

When they are looking at a deepfake, the company using it has to say the image, audio, or video was generated or manipulated, in a way a person can perceive. A deepfake, in the short reading used here, is generated or manipulated image, audio, or video that resembles something existing and would falsely appear real. A machine-readable mark in the file is not that disclosure.

When they are reading text published to inform the public on a matter of public interest, the company publishing it has to say the text was generated or manipulated, unless a person with the relevant knowledge, and the authority to approve, change, or reject the substance, reviewed it before it went out. A spell-check, or a grammar pass, is not that review. The exception covers this text duty only. It does not cancel the chatbot notice, the deepfake disclosure, or the provider's marking duty. An internal meeting note is not that publication. An ordinary email to one customer is not either. Do not invent a public label for either. That does not mean no duty can arise inside the company. Public health and consumer safety are Commission examples. They are not a classification of your post.

If the company uses a system that recognises emotion, or that categorises biometric data, the people exposed to it should be told. The company using the system under its own authority informs them. Staff using a vendor tool under that authority are not each a separate deployer.

Not every sentence an assistant touched needs a label. Reading this piece, or filling the optional list below, does not meet Article 50, does not decide that your company is the provider or the deployer, and does not replace a specialist. The dates, the article numbers, and the limits this piece does not walk are in the reference below.

If you want the five marks

The answer is above. This list is optional. You do not need it to know when people should be told.

Take five AI uses already in the work. You do not need an account, a file, or an email. For each one, write one mark:

• internal: only people in the company see it
• customer: a named customer or caller sees it
• public: it is published for people you do not know

If you do not know who sees it, write "unknown".

For the customer and public rows only, add two notes. What the person actually meets: a chat or voice reply, an image or video, a published text, or something else. If it is a published text, name the person who can change or reject it before it goes out. If nobody has that authority, write "none".

That list is a mark you write down. It is not a legal classification.

An internal meeting note stays internal. An ordinary email to one customer stays customer. A chatbot on your site, a generated face in an ad, and a public post are different rows. A quick glance is not the person who can change or reject a published text. Do not put a warning on every sentence an assistant touched and call the question closed. Do not read a vendor badge as a classification.

Reference: the four cases and the clauses

You do not need this section for the answer above. It keeps the statutory limits next to those sentences, so the opening is not read as a label on every sentence, or as a duty met. The four cases above are Article 50(1), the direct exchange; Article 50(4), the deepfake; Article 50(4), the public-interest text; and Article 50(3), emotion recognition or biometric categorisation. Article 50(2) is the machine-readable mark. It is not a fifth notice to a person. Full wording: consolidated Regulation (EU) 2024/1689, 27 July 2026.

These notices have been required since 2 August 2026, and not before that date. Article 113 sets the date. Article 50 has no earlier one. It is not a duty to label every sentence an assistant touched.

The provider of a system meant to interact directly with a person has to design it so the person is informed they are dealing with an AI system, unless that is obvious. Article 3(3): a provider develops the system, or has it developed, and places it on the market or puts it into service under its own name. A company that only uses a vendor chatbot is not automatically that provider.

The provider of a system that generates synthetic audio, image, video, or text has to mark the output in a machine-readable way, so it can be detected as artificially generated or manipulated. That duty does not apply to the extent the system only assists standard editing, or does not substantially alter the input or its meaning. That mark is not the optional list above. Article 111(4), inserted by Regulation (EU) 2026/1744, gives providers of such systems already on the market before 2 August 2026 until 2 December 2026 for this marking step. That later date does not delay the rest of Article 50. The mark is not the notice a person has to be able to perceive, and it is not a list you fill in.

A deployer who uses an emotion recognition system, or a biometric categorisation system, has to inform the people exposed to it. Article 3(4): a deployer uses a system under its authority, other than purely personal use. Staff using a vendor tool under the company's authority are not each a separate deployer. The Commission Q&A, last updated 24 July 2026, says the company remains the deployer.

A deployer of a deep fake has to disclose that the content was artificially generated or manipulated. Article 3(60): generated or manipulated image, audio, or video that resembles something existing and would falsely appear real. The Act limits that disclosure for some artistic uses and some criminal-investigation uses. This piece does not walk those limits. Whether one of them applies is a question for a specialist. The same Q&A says the provider's machine-readable mark is not the disclosure a person has to be able to perceive.

A deployer who publishes AI-generated or manipulated text to inform the public on a matter of public interest has to disclose that the text was artificially generated or manipulated, unless human review or editorial control and a natural or legal person holds editorial responsibility. The Q&A says a spell-check, or a grammar pass, is not that review. The person needs relevant knowledge and the authority to approve, alter, or reject the substance. That exception covers this text duty only. It does not cancel the chatbot notice, the deep fake disclosure, or the provider's marking duty. An internal meeting note, and an ordinary email to one customer, are not that publication. Public health and consumer safety are Commission examples, not a classification of your post.

Where this stops

Reading this piece, or filling the optional marks, does not meet Article 50. It does not decide that a named company is the provider or the deployer. It does not replace a qualified specialist.

This is practical orientation, not legal advice. For a consequential use, involve qualified legal and technical specialists and check the current official text. This list is not a compliance programme, a certificate, or proof that any obligation has been met.

• Consolidated Regulation (EU) 2024/1689, 27 July 2026
• Regulation (EU) 2026/1744
• AI Act Service Desk, Article 50
• Commission Q&A on Article 50, 24 July 2026

Earlier pieces in this series:

• EU AI Act: list where your company already uses AI
• The EU AI Act Is Not One Rule for Every AI Use
• Using an AI Tool Still Means Responsibility | AI Act