Related daily report
July 31, 2026: EU transparency duties move from preparation to operation
The latest report identified August 2 as an immediate compliance deadline for affected chatbot providers, generative-AI providers and professional deployers of certain AI systems.
Open the permanent July 31 reportThe direct answer
People must be informed in several situations where AI could otherwise be mistaken for a person or authentic content
Article 50 of the EU AI Act starts applying on August 2, 2026. It creates transparency duties for providers and professional deployers of certain AI systems. The rules are designed to reduce deception and confusion by making the artificial origin of an interaction or piece of content easier to recognise.
When AI is acting like a person, generating realistic media or producing certain public-interest material, the people exposed to it should receive an appropriate notice or label.
This does not create one universal sticker that must appear on everything touched by AI. Different obligations apply to different actors and situations. Some information must be visible to a person. Other information must be embedded in a machine-readable form so software can help detect that content was generated or manipulated by AI.
The transparency rules focus on four main cases
Direct AI interaction
Providers of systems intended to interact directly with people must generally ensure users are informed that they are interacting with AI, unless that fact is already obvious in the circumstances.
Example:A customer-service chatbot should not quietly present itself as a human employee.
AI-generated or manipulated content
Providers of generative systems must generally enable synthetic audio, image, video and text to be identified through machine-readable marking, subject to scope limits and exceptions.
Example:An AI image system may embed provenance information that compatible tools can inspect.
Deepfakes and public-interest text
Professional deployers must disclose deepfakes. They must also disclose certain AI-generated or manipulated text published to inform the public on matters of public interest when it has not received human review or editorial control.
Example:A realistic synthetic video of a public official should be clearly identified as artificially generated or manipulated.
Emotion and biometric systems
Deployers of emotion-recognition or biometric-categorisation systems must inform people who are exposed to those systems, subject to the Act’s specific conditions and exceptions.
Example:A workplace using an in-scope system to infer emotional states cannot treat the system as invisible background software.
Machine-readable marking and visible labelling are not the same thing
A machine-readable mark may be embedded in metadata or another technical signal. A visible disclosure may appear as text, an icon, a spoken notice or another clear communication appropriate to the medium. One does not automatically replace the other where both duties apply.
The European Commission has also published optional icons and a voluntary Code of Practice to help organisations implement marking and labelling consistently. Following the code can support compliance, but it does not remove an organisation’s responsibility to satisfy the law.
What ordinary users may notice
- Clearer chatbot notices: a service may state that replies are generated by AI before or during the conversation.
- More labels on synthetic media: videos, images or audio may carry a visible “AI-generated” or “AI-altered” disclosure.
- Platform detection signals: social networks and publishing tools may read embedded provenance information and add their own notices.
- Public-interest disclosures: certain unreviewed AI-generated text intended to inform the public may be labelled.
- Notices around sensitive systems: people may receive information when exposed to in-scope emotion-recognition or biometric-categorisation tools.
A label tells you how content was produced. It does not tell you whether the content is true, fair, safe or useful.
Users should still check the original source, date, context and evidence. Authentic human content can be false, and labelled AI content can be accurate. Transparency is a starting point for judgment, not a replacement for judgment.
Businesses need to identify whether they are a provider, a deployer or both
A company that develops or releases an AI system under its own name may be a provider. A company that uses an AI system under its authority for professional purposes may be a deployer. The same organisation can occupy both roles in different parts of a service.
Inventory the AI systems
List chatbots, content generators, synthetic-media tools, emotion systems and biometric categorisation systems used in products or operations.
Map each output
Record whether a system creates text, audio, images or video and whether the output is customer-facing, public or used only internally.
Separate technical and visible controls
Confirm which outputs require machine-readable provenance, which require a visible disclosure and which may require both.
Keep evidence
Document the notice wording, marking method, exceptions relied upon, testing performed and the person responsible for review.
This article is a general explainer, not legal advice. Organisations operating in or serving the European Union should review the official guidelines and obtain qualified advice for their own systems and use cases.
There are exceptions, and one limited transition period
The exact scope depends on the system and context. For example, the direct-interaction notice is not required where it is obvious to a reasonably informed and observant person that they are interacting with AI. Article 50 also contains specific law-enforcement and creative-work provisions.
The Commission explains that AI systems placed on the market before August 2, 2026 receive a limited transition period until December 2, 2026 for the machine-readable marking and detection obligation under Article 50(2). That limited period does not postpone every Article 50 duty.
Content generated before August 2 does not generally have to be labelled retroactively, although the Commission encourages voluntary labelling where practical.
What the rules do not mean
Ordinary assistive editing and changes that do not substantially alter meaning may fall outside particular marking duties.
The public-interest text disclosure concerns defined professional publication circumstances, including whether human review or editorial control occurred.
It identifies AI involvement; it does not certify factual accuracy or quality.
Organisations outside the EU may still be affected when they place relevant systems on the EU market or put them into service there.
Use this today
A practical transparency checklist
- 1State when users are interacting with AI.
Use a notice that is clear, timely and appropriate to the interface.
- 2Check whether generated content carries a reliable technical mark.
Test whether the mark survives the normal publishing and export process.
- 3Label deepfakes and covered public-interest content visibly.
Do not hide the disclosure where an ordinary user is unlikely to see it.
- 4Record human review.
Document who reviewed public-interest content and what editorial checks were performed.
- 5Tell people about in-scope emotion or biometric categorisation.
Review whether the system is lawful and whether other AI Act or privacy rules also apply.
- 6Keep an audit trail.
Store policies, test results, versions, labels, exceptions and responsible owners.
Verified sources
Official European Union material
- European Commission guidelines on Article 50 transparency obligations
- Commission overview of AI-generated-content transparency rules
- Quick facts: transparency rules for AI systems
- The EU AI Act on EUR-Lex, including Article 50 and Article 113
Verification note: This explainer was checked against official European Commission and EUR-Lex material on August 1, 2026. Implementation details may be clarified further by regulators and courts.
The bottom line
AI transparency is becoming an operational requirement, not a voluntary courtesy
From August 2, affected organisations need more than a general statement that they “use AI.” They need notices, machine-readable marking, visible disclosures and records matched to the specific system and use case. For ordinary users, the change should make it easier to recognise when an interaction or piece of content has an artificial origin—but every label still requires human judgment about truth, context and trust.