EU AI Transparency Guidelines Set August 2 Disclosure Deadline for Providers and Deployers

The European Commission published guidance on July 20, 2026, ahead of the August 2 start date for the EU AI Act’s transparency obligations. The rules are designed to help people recognise when they are interacting with AI and when content has been generated or altered by AI. The practical shift is not a universal “AI label” on everything: different duties apply to providers that build or supply systems and deployers that put them in front of people.

The Commission says the guidance covers interactive AI systems and the marking and labelling of AI-generated content. It clarifies which providers and deployers must comply under Article 50 of the AI Act, making the August 2 date a useful checkpoint for companies, publishers, public bodies and other organisations that use generative systems in public-facing workflows.

PanoramaDigest has covered the wider policy environment in its AI policy topic hub and its report on the Shanghai World AI Conference. This is a different kind of AI story: the question is not who has the most capable model, but how people are told what they are seeing, hearing or interacting with.

Providers and deployers have different jobs

The distinction between provider and deployer matters because responsibility follows the point of control. A provider designs or supplies an AI system. A deployer uses it in practice, such as placing a chatbot on a service website, publishing generated material, or operating a system that categorises or assesses people. The Commission’s guidance is intended to help organisations determine which obligations apply to their role.

For a provider, the central task is system design. The Commission says providers will have to make users aware when they are directly interacting with AI and add machine-readable marks to AI-generated or manipulated content. A visible notice may help a person, but machine-readable marking is aimed at technical systems that need to detect or process the signal at scale.

For a deployer, the question is exposure. The Commission specifically identifies deepfakes, public-interest content produced without human review or editorial control, and emotion-recognition or biometric-categorisation systems. That means a compliance process cannot stop at the model vendor. Organisations using an external tool still need to understand how their own deployment is presented to people.

Why machine-readable marking matters

A visible label and a machine-readable mark solve different problems. A label is meant to help the person looking at a post, image, video or interface. A machine-readable mark can help platforms, search systems, moderation tools and other software identify that content has an AI origin or has been manipulated.

The approach reflects a larger problem in synthetic media: content can travel far beyond the context in which it was created. A user may see a video after it has been reposted, cropped or translated. If disclosure depends only on a caption that disappears during redistribution, the signal is fragile. Machine-readable information is not a complete solution, but it gives downstream systems something more durable to inspect.

Deepfakes and public-interest content receive special attention

The Commission’s wording puts deepfakes and certain public-interest content in a higher-scrutiny category for deployers. This is understandable: a synthetic image of a fictional character and a fabricated video presented as evidence about an election, disaster or public-health event can have very different consequences.

The public-interest rule also includes an important qualification in the Commission’s summary: the obligation concerns AI-generated content on matters of public interest when there is no human review or editorial control. That does not turn “human in the loop” into a magic exemption. It does mean organisations will need to document how review works and whether a person actually exercised editorial judgment rather than merely clicking approve.

What organisations should do before August 2

These steps are an operational checklist, not legal advice. The Commission’s full guidelines, questions and answers, factsheet and Code of Practice are the controlling sources for organisations assessing their own obligations.

What the guidance does not establish

The July 20 publication is guidance, not a claim that every compliance question has been resolved for every business model. It does not mean that users can identify every synthetic image with certainty, and it does not replace sector-specific rules or an organisation’s own risk assessment. It also does not make a visible label proof that content is authentic when no label appears.

The larger policy bet is that disclosure becomes part of the technical infrastructure of AI services rather than an optional courtesy. If providers build reliable signals and deployers use them consistently, people may gain more context when they meet an AI system or encounter synthetic media. If labels are inconsistent, removable or too vague, the rules will be harder to apply and easier to misunderstand.

Watch related newsroom context: the Associated Press video European lawmakers vote on the Artificial Intelligence Act provides background on the law behind today’s guidance. If the player does not load, use the direct YouTube link.