In forceAI transparency9 min read

EU AI Transparency and Labelling Requirements

Article 50 creates different transparency duties for providers and deployers. Depending on the system and content, organisations may need to inform people that they are interacting with AI, enable machine-readable detection of synthetic outputs or disclose deepfakes and certain AI-generated public-interest text.

Applies from: 2 August 2026
For: AI providers, deployers, marketing teams, publishers and digital product owners
Last legally reviewed: 30 September 2026

Educational guidance: This overview explains EU rules in general terms. Applicability depends on the organisation, system, service and national law.

There is no single universal AI label

Article 50 separates duties by role and use case. A provider that designs an AI system to interact directly with people has different responsibilities from a deployer publishing a deepfake or AI-generated public-interest text.

The first step is to identify who provides the system, who deploys it and what kind of output or interaction is involved. A blanket label placed on every use of software may be both unhelpful and legally imprecise.

Direct interaction with an AI system

Providers of AI systems intended to interact directly with natural persons must generally design them so people are informed that they are interacting with AI, unless that is obvious to a reasonably well-informed, observant and circumspect person in the circumstances.

The Commission says the information should be given at the latest at the start of the first interaction and should be clear, distinguishable and accessible. Product teams should test where the disclosure appears, whether it survives different devices and languages, and whether users can understand it before relying on the interaction.

Synthetic content and deepfakes

Providers of systems generating synthetic audio, image, video or text must address machine-readable marking and detectability requirements described by Article 50. Deployers using systems to generate or manipulate deepfake content must disclose the artificial generation or manipulation at the latest upon first exposure, subject to the Regulation's specific conditions and exceptions.

A 2026 transition rule gives providers of synthetic-content systems placed on the market before 2 August 2026 until 2 December 2026 to comply with Article 50(2). This limited transition does not postpone the other Article 50 duties that apply to their own actors and use cases.

The disclosure process should travel with the content through publication workflows. Removing metadata, cropping a label or exporting through another tool can break the evidence chain.

Public-interest text and editorial responsibility

Deployers publishing AI-generated or manipulated text to inform the public on matters of public interest should assess the Article 50 disclosure rule. The Regulation includes a qualification where content has undergone human review or editorial control and a person holds editorial responsibility.

That is not a reason to treat a superficial approval click as editorial review. Record the reviewer, source checks, substantive changes, publication decision and responsible publisher.

Keep evidence for each provider and deployer control

Providers should retain the system and release version, technical marking specification, detectability testing, known limitations, exception analysis and monitoring evidence. Deployers should retain the notice or label, its placement and timing, the content and publication version, accessibility checks and the responsible owner.

Where public-interest text relies on human review or editorial control, record who reviewed the substance, what sources and claims were checked, what changed and who accepted editorial responsibility. A spellcheck or superficial approval is not substantive human review.

Action list

Practical checklist

Classify the organisation as provider or deployer for the relevant system.
Identify direct AI interactions and place timely disclosures where required.
Check provider controls for machine-readable synthetic-content marking.
Create a disclosure workflow for deepfakes and relevant public-interest text.
Preserve labels and metadata through the publishing chain.
Record meaningful human review and editorial responsibility.
Test accessibility, clarity and language coverage of disclosures.

Questions

Common questions

Must every AI-assisted document carry an AI label?

No. Article 50 contains role- and use-specific duties. The system, output, publication context and any relevant exception or qualification must be assessed.

When must a chatbot say it is AI?

Providers of systems intended to interact directly with people must generally ensure people are informed, unless the AI nature is obvious to a reasonably well-informed, observant and circumspect person in the circumstances.

Does human review remove every disclosure duty?

No. Article 50 contains a specific qualification for certain public-interest text with human review or editorial control and editorial responsibility. It is not a universal exemption for all AI content.

Primary material

Official sources

Last legally reviewed: 30 September 2026. Recheck the consolidated legislation and relevant national implementation before relying on this guide for a specific decision.