From chatbots and AI-generated content to deepfakes and machine-readable provenance, Article 50 of the EU AI Act has brought a new set of transparency obligations into effect. The immediate challenge is compliance. The bigger question is whether Europe is beginning to establish something more fundamental: a trust layer for an information environment increasingly shaped by artificial intelligence.

For much of the discussion surrounding the EU AI Act, the emphasis has been on what organisations will eventually have to do. On 2 August 2026, another part of that conversation moved firmly into the present.

Article 50 of the AI Act is now applicable, introducing transparency obligations covering AI systems that interact directly with people, AI-generated and manipulated content, deepfakes, emotion recognition and biometric categorisation systems. The rules also address AI-generated text used to inform the public on matters of public interest.

It is important to be precise about what has happened. The AI Act itself entered into force in August 2024. Article 50 did not suddenly become law on 2 August this year. Rather, this is the date from which its transparency obligations began to apply. From this point, affected providers and deployers must be able to demonstrate how those obligations are being met.

That may sound like another regulatory milestone. Its implications could be considerably broader.

Knowing when AI is present

At the simplest level, Article 50 introduces a principle that most people will find reasonable: if someone is interacting directly with an AI system, they should generally know that they are doing so.

Providers of systems such as chatbots, AI agents and avatars must design them so that users are informed they are interacting with AI, unless this would already be obvious to a reasonably well-informed and observant person in the circumstances.

The significance lies in the principle behind the requirement. AI is becoming less visible as an individual technology and more deeply embedded within customer service, enterprise software, digital platforms and automated workflows. The question is therefore shifting from whether an organisation uses AI to whether people can recognise when AI is mediating an interaction. Article 50 places transparency into that relationship.

For organisations deploying AI commercially, this means that transparency cannot simply be considered a communications decision made after a system has been introduced. In relevant applications it increasingly becomes part of how the system itself is designed and presented.

The provenance problem

The more technically interesting element of Article 50 concerns generated content. Providers of AI systems that generate synthetic audio, images, video or text are required to ensure that their output can be marked in a machine-readable format and detected as artificially generated or manipulated. The technical measures are expected to be effective, interoperable, robust and reliable, while taking account of the type of content, implementation costs and the state of the art.

This takes the discussion beyond the familiar idea of putting an “AI-generated” notice underneath a photograph. Human-readable disclosure remains important in particular circumstances, but machine-readable marking introduces another layer. It creates the possibility that platforms, applications and verification systems can establish something about the origin of content without relying solely on a person noticing a label.

That is essentially a provenance problem. As generative AI improves, judging authenticity simply by looking at content becomes progressively less reliable. Images no longer need six fingers to give themselves away. Voices can be recreated. Video can be altered. Text can be generated at enormous scale while appearing entirely conventional.

The answer cannot therefore depend indefinitely upon humans spotting the synthetic material. Some form of technical signal has to travel with the content.

Deepfakes make the issue visible

Deepfakes are perhaps the most obvious example of why this matters. Under the AI Act, deployers using AI to generate or manipulate image, audio or video content that constitutes a deepfake must disclose that the content has been artificially generated or manipulated.

There are exceptions and specific provisions for artistic, creative, satirical and fictional material, where disclosure should not unnecessarily interfere with the work itself. But the underlying expectation is increasingly clear: content capable of presenting something artificial as authentic cannot simply be released into the information environment without consideration of its provenance.

The Commission defines a deepfake around more than the use of generative technology. The content must resemble an existing or plausibly existing person, object, place, entity or event and be capable of falsely appearing authentic or truthful.

That distinction matters because AI manipulation is rapidly becoming part of ordinary production. Removing background noise from audio, altering an image or using AI within film production does not automatically turn content into a deepfake.

The issue is not simply whether AI touched the content. It is whether AI changes what the audience reasonably believes the content represents.

Publishing also enters the equation

Article 50 also has an important implication for publishers, corporate communications teams, public bodies and organisations producing material on issues of public interest.

Where AI generates or manipulates text that is published to inform the public on matters of public interest, deployers may have a disclosure obligation. The Commission’s guidance interprets this broadly enough to include areas such as politics, public administration, security, health, environmental protection, consumer safety and economic, financial, scientific or cultural developments that could become subjects of public debate. There is, however, a significant distinction.

The disclosure requirement does not apply in the same way where AI-generated content has undergone substantive human review or editorial control and a natural or legal person holds editorial responsibility for its publication.

The Commission is also quite clear about what it considers meaningful review. Human review involves deliberate examination of the substance by someone with relevant knowledge and professional judgement. Editorial control requires the authority to approve, alter or reject the substance of the text, including checking information and sources. Superficial grammar or spelling corrections are not enough. That may prove to be one of the more consequential aspects of Article 50.

It effectively recognises a distinction between AI as an autonomous publishing mechanism and AI operating inside an accountable human editorial process. For media organisations, companies and institutions increasingly using AI within content production, the question therefore becomes less simplistic than “Was AI used?”. It becomes: Who took responsibility for what was ultimately published?

Provider and deployer are not the same thing

Another potential source of confusion is that Article 50 does not place every responsibility on the company building the AI model. The Act distinguishes between providers and deployers.

Providers develop AI systems, or have them developed, and place them on the market or put them into service under their own name or trademark. Their obligations include designing relevant systems to disclose AI interaction and introducing machine-readable marking for generated content.

Deployers are the organisations or professionals actually using AI systems under their authority. Their responsibilities can include informing people exposed to emotion recognition or biometric categorisation systems and disclosing deepfakes or relevant AI-generated public-interest material.

This matters because the practical responsibility for AI transparency therefore extends considerably further than the major foundation-model companies. An enterprise cannot necessarily assume that buying an AI service transfers every transparency obligation to the supplier.

From labels to infrastructure

There is a danger of seeing Article 50 mainly as a labelling exercise. That would miss the more interesting development.

If AI-generated content can carry reliable machine-readable signals concerning its artificial origin, those signals can potentially become part of a wider technical ecosystem for establishing provenance. Platforms can detect them. Verification services can interpret them. Organisations can incorporate them into workflows. Other systems can make decisions based upon them.

That does not establish whether the content itself is true. Authentic human content can still be false, while AI-generated content can be entirely accurate. Provenance answers a different question: what is this, and where did it come from?

In an information environment increasingly populated by humans, AI agents and synthetic media, that question is becoming foundational. The challenge will be interoperability. A provenance system that only operates inside one model, platform or vendor ecosystem solves relatively little. Signals need to survive distribution, editing and transformation while remaining detectable across technical boundaries.

Article 50 explicitly points towards effective, interoperable, robust and reliable marking. Whether industry can deliver that consistently will be one of the areas worth watching as implementation develops.

Trust has to become machine-readable

The AI Act is sometimes presented primarily as an attempt to restrict artificial intelligence. Article 50 demonstrates another side of the regulation. AI is going to participate increasingly in the creation of information, communication with customers, software workflows and decisions made throughout the digital economy. Preventing that development is neither realistic nor necessarily desirable.

What becomes essential is the ability to understand when AI is present and to establish accountability around what it produces. For humans, that requires transparency. For digital systems operating at machine speed, it increasingly requires provenance that machines themselves can recognise.

Article 50 will not solve the wider problem of synthetic content, misinformation or deepfakes. Nor will a machine-readable marker suddenly establish truth in an information environment where manipulation existed long before generative AI. But the direction is important.

Europe is beginning to move the debate from asking whether AI should be trusted towards building mechanisms that allow trust to be assessed. And in an AI-driven information economy, that may prove far more important than the label itself.