Article 50 Transparency Duties and What Staff Must Disclose
Article 50 is the part of the AI Act that most employees will encounter directly, because it governs what people must be told when they meet AI-generated content or an AI system. Its obligations apply from 2 August 2026. Understanding it is part of literacy for anyone who writes to customers, publishes content, runs a chat channel or handles recorded interactions.
The Article splits its duties between providers and deployers, and staff need to know which side they are on. Providers of AI systems intended to interact directly with natural persons must design and build those systems so that the person is informed they are interacting with an AI system, unless that is obvious to a reasonably well-informed and observant person in the circumstances. If your organisation buys a chat product, the supplier carries that design duty, but you carry the consequences of configuring it to sound like a named human employee.
Providers of systems that generate synthetic audio, image, video or text, including general-purpose AI systems, must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The marking requirement is technical and sits with the provider, and it is where the December 2026 date matters. Generative systems that were already on the market before the transparency rules applied must meet the machine-readable marking requirement by 2 December 2026. For a deployer, the practical significance is that the absence of such marking on a supplier’s output after that date is a question worth asking the supplier.
The duties that fall on deployers are the ones staff perform. Deployers of emotion recognition systems or biometric categorisation systems must inform the natural persons exposed to them about the operation of the system, and must handle personal data in line with data protection law. This reaches further than security teams expect, since call-centre analytics that infer a caller’s emotional state fall inside it. Some emotion inference in the workplace and in education is prohibited outright under Article 5 rather than merely regulated, which is another reason this area needs a check before deployment rather than after.
Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake must disclose that the content has been artificially generated or manipulated. Where AI-generated or manipulated text is published with the purpose of informing the public on matters of public interest, deployers must disclose that as well, unless the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility. There is a carve-out for evidently artistic, creative, satirical or fictional work, where the disclosure must be made in a way that does not spoil the display or enjoyment of the work.
Disclosure must be given clearly and at the latest at the time of the first interaction or exposure, and it must be accessible to people with disabilities. A note in a privacy policy nobody opens does not meet the timing requirement, and an on-screen label that a screen reader cannot reach does not meet the accessibility requirement.
Translating this into staff behaviour produces a small number of rules that can be taught in minutes. If a customer asks whether they are talking to a person, say so honestly, and never allow a bot to claim to be a named human. If you publish a synthetic image, video or voice of a real person, label it. If you publish AI-generated text about a matter of public interest, either put a named human editor in the chain and record that, or label it. Do not use an AI voice or likeness of a colleague, a customer or a public figure without permission and a label. If you are unsure whether something counts as a deep fake, ask before publishing rather than after.
Article 50 also has a practical relationship with Article 4 that is easy to miss. The transparency duties can only be met by people who recognise the situations they cover, so a deployer whose staff have never been told what a deep fake disclosure is will breach Article 50 through ignorance rather than intent. Training on Article 50 is therefore both a transparency control and a piece of Article 4 evidence, which makes it one of the higher-value modules to build.
Two boundaries are worth stating plainly. The first is that Article 50 governs disclosure and not permission. Labelling a manipulated image does not make it lawful if it defames someone, breaches copyright or infringes a person’s likeness rights. The second is that internal use is not automatically outside the Article. Employees are natural persons, and a deployment aimed at staff can engage the same duties as one aimed at customers, which surprises organisations that treat internal tools as a lower tier of governance. Where an internal tool infers emotional states from staff communications, the correct first question is not how to disclose it but whether Article 5 permits it at all.