What has applied since 2 August 2026
Article 50 of the AI Act has applied since 2 August 2026. It governs when the involvement of AI has to be disclosed. It was not postponed and it was not softened: the Digital Omnibus moved the high-risk obligations to 2 December 2027, but it left the transparency duties where they were.
On 20 July 2026 the European Commission published guidelines on implementing Article 50. Across 51 pages they set out how authorities will read the provision. They are currently the most reliable source on what is expected of you.
Start by listing the AI tools actually in use in your company. None of the questions below can be answered without that list.
Provider or deployer: this distinction decides your duties
The AI Act separates providers, who develop or place AI systems on the market, from deployers, who use them. A company that buys ChatGPT, Copilot or an AI-assisted recruitment tool and uses it at work is a deployer.
Two of the five paragraphs in Article 50 address providers. Paragraph 1 requires a system that interacts with people to make that fact apparent. Paragraph 2 requires synthetically generated content to be marked in a machine-readable format. For systems already on the market before 2 August 2026, that marking duty has a transitional period until 2 December 2026 (Article 111(4)).
Neither paragraph is your job. Both are worth checking when you select a vendor: a tool that does not meet its own obligations becomes your problem the moment you build on it.
The three duties that fall on deployers
Emotion recognition and biometric categorisation
Article 50(3) requires you to inform people when they are exposed to an emotion recognition system or a biometric categorisation system. This applies whether the system runs in real time or after the fact.
In the workplace and in education, emotion recognition is largely prohibited under Article 5 in any case. Where it remains permitted, the duty to inform applies on top.
Deepfakes
Under Article 50(4) you must disclose when you publish image, audio or video content that has been generated or manipulated with AI and appreciably resembles real people, places or events.
AI-generated text on matters of public interest
The second subparagraph of Article 50(4) is the one that surprises most companies. It applies when you publish text that has been generated or manipulated with AI and is intended to inform the public on matters of public interest.
The guidelines narrow this. It covers text accessible to an indeterminate, fairly large number of readers, that communicates knowledge, opinions or facts, on topics that matter to society at large. The Commission expressly includes the protection of fundamental rights, public health, consumer safety, and economic and scientific developments of public interest.
The guidelines put several things outside the scope, among them internal communications on a corporate network, private correspondence, advertising and product descriptions that make no claims about matters such as health or consumer safety, and AI-assisted advice from a consultant to an individual client.
Go through your public-facing text and separate marketing from content genuinely intended to inform. Only the second group is in scope here.
The exception many companies miss
For AI-generated text, the disclosure duty does not apply where two conditions are met together: the text has undergone human review or editorial control, and a natural or legal person holds editorial responsibility for the publication.
The guidelines set a clear standard. Human review means the deliberate examination of the substance of the content by a person with relevant knowledge and professional judgement about the subject matter, and fact-checking the accuracy of the content is a minimum requirement. Expressly not sufficient: superficial or purely formal checks, spell-checking alone, the mere existence of an editorial policy, automated review processes, or cursory sign-off without substantive engagement.
Two points that are easily lost in practice:
The responsible person or function must be publicly findable, for instance through the legal information on your website. And if AI is used again after editorial sign-off to change, supplement or reformulate the substance of the text, the exception falls away.
If you use AI when writing, decide who performs the substantive review, record that it happened, and publish who carries the responsibility.
What non-compliance costs
Breaches of the Article 50 transparency obligations can be fined up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. For small and medium-sized enterprises, including start-ups, the lower of the two figures applies.
The frequently quoted ceiling of EUR 35 million or 7% does not apply to the transparency duties. It applies only to the practices prohibited under Article 5.
Map your AI applications to the right duties before you discuss risk. The figure depends on which duty is engaged.
What to do now
- Record which AI tools are used in the company, and for what.
- Flag the applications that fall under paragraph 3 or paragraph 4.
- For published text, decide who reviews the substance and who holds editorial responsibility, and make that publicly findable.
- Document both. The burden of demonstrating compliance sits with you.