EU AI Act Article 50, in plain English

The short version

Article 50 of the EU AI Act (Regulation (EU) 2024/1689) is the EU's AI transparency rule. It has applied since August 2, 2026. A July 2026 amendment package, the Digital Omnibus, delayed other parts of the AI Act but left these start dates alone, with one grace period for older AI tools.

Article 50 sets five duties. Two fall on the companies that build AI systems: chatbots must be designed to tell people they are AI, and AI-generated media must carry hidden machine-readable marks. Three fall on businesses that use AI in their work. Two of those reach ordinary businesses: label deepfakes, and label AI-written text published to inform the public on matters of public interest unless a person has reviewed it and someone takes editorial responsibility. The third, telling people when emotion-recognition or biometric-sorting systems are running, mostly matters to retailers and venues using camera analytics.

For a typical US small business with no EU customers, Article 50 is background. For one that sells to EU customers, runs its own chatbot for them, or publishes realistic AI images or video aimed at EU audiences, it is a live rule.

Key dates

Content you made before August 2, 2026 does not have to be labelled after the fact, according to the Commission's Article 50 FAQ, though the Commission encourages it.

First question: are you a provider or a deployer?

The AI Act splits duties by role, and Article 50 is no exception.

• Provider: the company that builds or supplies an AI system and puts it on the market under its name. OpenAI, Google, a chatbot vendor, an image-generator app.
• Deployer: any business, organisation or professional that uses an AI system under its own authority. A shop that uses an image generator for ads, an accountant who embeds a vendor's chatbot, a publisher that drafts articles with AI.

Most small businesses are deployers. The line moves when you build or brand your own system: a business that builds its own bot, or sells its own AI tool, may be treated as the provider of it. Purely personal, non-professional use falls outside the Act.

Article 50(1): chatbots, voice assistants and AI agents

Who: providers. What: an AI system built to interact directly with people must be designed so those people know they are dealing with an AI. The notice has to be clear, easy to spot, given at the first interaction at the latest, and accessible.

There is one exception: where it is already obvious to a reasonably well-informed, observant and careful person that they are dealing with AI. The Commission's FAQ says that exception should be read narrowly, so do not count on it for a chat widget with a human name and a friendly photo. AI uses authorised by law to detect or prosecute crime are also exempt.

What it means for a small business: if you embed a vendor's chatbot, the duty sits with the vendor, and the Commission's FAQ lists no 50(1) duty for deployers. If you build or brand your own bot for EU customers, you may be the provider. Either way, an opening line like "I'm an AI assistant" costs nothing and gives users the notice Article 50(1) is aiming for, whoever turns out to hold the duty. See the chatbot greeting example.

Article 50(2): machine-readable marking

Who: providers of generative AI, including general-purpose AI. What: AI-generated or manipulated audio, images, video and text must carry a machine-readable mark, such as metadata or a watermark, so the content can be detected as AI-made. The marking must be effective, interoperable, robust and reliable as far as technically feasible. It does not apply to purely assistive editing, or where the AI does not substantially change what the user put in.

Systems placed on the EU market from August 2, 2026 must comply now. Systems that were already on the market before that date have until December 2, 2026, a grace period added by the Digital Omnibus.

What it means for a small business: your image, video or writing tool carries this duty, not you. Your part is not to strip the marks. Some image compressors and content delivery networks remove metadata by default; check your settings if you publish AI images. The same metadata now drives platform labels on YouTube, Meta, TikTok and LinkedIn.

Article 50(3): emotion recognition and biometric categorisation

Who: deployers. What: if you use a system that recognises emotions or sorts people into categories from biometric data, you must tell the people exposed to it that it is running, and handle the data under GDPR. Emotion recognition in workplaces and schools is banned outright under Article 5(1)(f), except for medical or safety reasons.

This matters for retailers or venues using camera analytics. Most small businesses will not touch it.

Article 50(4): label deepfakes

Who: deployers, which includes any business using AI to make marketing images, video or voice. What: if you use AI to create or alter image, audio or video content that counts as a deepfake, you must disclose that it is AI-generated or manipulated. A deepfake here means content that resembles real people, objects, places, entities or events and would falsely appear to be authentic.

The label has to be clear and distinguishable at the viewer's first exposure. The hidden watermark from your tool vendor does not satisfy this duty; it needs to be something a person can see or hear. For work that is evidently artistic, creative, satirical or fictional, a lighter touch is allowed: you only need to signal that AI content is present, in a way that does not spoil the work.

Examples that likely need a label: an AI-altered photo of your real shop, product or staff that would pass as a genuine photo; a realistic AI video of a real place or event; a cloned voice of a real person in an ad. Examples that likely do not: an obviously cartoon mascot; a stylised illustration no one would take for a photo. When in doubt, label.

Article 50(4): AI-written text on matters of public interest

Who: deployers who publish text. What: if AI-generated or AI-manipulated text is published to inform the public on matters of public interest, you must disclose that it was AI-generated. The exemption is the important part: no label is needed if the text went through human review or editorial control and a person or company holds editorial responsibility for publishing it.

Routine marketing copy about your own products is generally not "public interest" text. A newsletter on local zoning changes, a blog post explaining new tax rules, or a health explainer is closer to the line. The practical fix is the same either way: have a named person review and stand behind what you publish, and you are outside the labeling duty. The Commission's guidelines go into more detail on what counts as public interest and genuine review; read them before relying on fine distinctions.

Fines

Breaking an Article 50 duty can bring a fine of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher (Article 99(4)(g)). For small and medium-sized businesses and start-ups the cap is whichever is lower (Article 99(6)). National market surveillance authorities in each EU country enforce it.

When a US small business is affected

The AI Act reaches providers wherever they are based if their system is placed on the EU market or used in the EU. How far it reaches non-EU deployers depends on where the AI's output is used, and that territorial question is beyond what we can settle here. Use these scenarios as a first sort, then get specific advice if EU sales matter to you.

Probably not your concern

• You sell only in the US, your site and ads are not aimed at EU audiences, and you have no EU customers.
• You use AI tools internally (drafting, summarising, research) and a person reviews everything before it goes out.

Worth a closer look

• You run a chatbot on a site that sells to EU customers. If it is a vendor's bot, the provider duty is theirs; if you built or branded it, it may be yours. Add the disclosure line regardless.
• You publish realistic AI images, video or audio in ads or social posts aimed at EU audiences. That is the deployer deepfake duty.
• You publish AI-drafted articles on public-interest topics for EU readers without human editorial review.
• You build and sell your own AI tool, app or generator used in the EU. You may be a provider under 50(1) and 50(2).

Content you published before August 2, 2026 does not need to be relabelled.

Help from Brussels: the Code of Practice and the guidelines

The Code of Practice on Transparency of AI-generated Content, published in final form June 10, 2026, is a voluntary rulebook in two parts: one for providers on marking and detecting AI content, one for deployers on labelling deepfakes and AI text, including official EU icons for AI-generated content. About 190 organisations had signed by the end of July 2026, and the Commission has assessed it as adequate. Signing is optional, and following it is not conclusive proof of compliance. Businesses that do not sign must show they meet Article 50 some other way.

The Commission's Article 50 guidelines, published July 20, 2026, are not binding. They explain who counts as a provider or deployer, the narrow reading of "obvious" for chatbots, what counts as a deepfake, what counts as public-interest text and genuine human review, and timing.

Outside the EU: the UK, Canada and China

United Kingdom

The UK has no law requiring AI content to be labelled or chatbots to identify themselves. In a March 2026 ministerial statement the government said it would set up a taskforce on labelling best practice, with an interim report expected in autumn 2026; we had not seen it as of October 15, 2026. General consumer law against misleading practices still applies.

For ads, the Committee of Advertising Practice published advice on AI disclosure on May 29, 2025. There is no blanket rule. CAP suggests asking two questions: would the audience be misled without a disclosure, and if so, does a disclosure clarify the ad or contradict it? A disclosure cannot fix an ad that is misleading anyway, such as an AI image that exaggerates what a cosmetic does. Routine AI retouching similar to normal post-production generally needs no disclosure.

Canada

• Ontario job postings. Since January 1, 2026, Ontario employers with 25 or more employees must state in each publicly advertised job posting whether they use AI to screen, assess or select applicants, including when a recruiter does the screening for them. Postings must be kept for three years after they come down. Ontario's guide.
• Quebec. Under Law 25 (section 12.1 of Quebec's private-sector privacy act), a business that makes a decision about a person based exclusively on automated processing must tell them, and on request explain the information and main factors used. In force since September 22, 2023. We could not re-check the official text on October 15, 2026, so confirm it at LégisQuébec.
• Federal. The Artificial Intelligence and Data Act died with Bill C-27 in January 2025 and was never enacted. Bill C-36, introduced June 15, 2026, would require plain-language accounts of automated decision systems; it has not passed.

China

If you market on Douyin, WeChat, Weibo or other Chinese platforms, China's Measures for Labeling AI-Generated Synthetic Content have applied since September 1, 2025. AI services must add visible labels and hidden metadata, platforms must check for them, and users who post AI-generated content must declare it. Removing or hiding labels is prohibited. A mandatory national standard, GB 45438-2025, sets the technical format.