EU AI Act: What AI Content You Have to Label from 2 August 2026

If your team uses AI to make images, write copy or run a chatbot, the EU AI Act now asks you to label some of that. From 2 August 2026, certain AI content has to be disclosed as AI. The good news: it is a shorter list than the panic suggests, and most of what a marketing team produces every day does not need a label at all.

This is the practical version: what you have to label, what you don't, where the line runs on the two cases that actually trip people up, and from when.

Note: this is a practical explainer, not legal advice for a specific case. For a binding assessment, please consult the primary sources or qualified counsel.

From when does this apply?#

2 August 2026. That is the date the Article 50 transparency rules start to apply.

You may have heard the AI Act was delayed. Part of it was. The Digital Omnibus (in force since 27 July 2026) pushed the high-risk rules back to 2027 and 2028, the ones for systems used in areas like recruiting or credit scoring. It did not touch the transparency rules that apply to everyday content. Those still land on 2 August 2026. The only piece with extra time is the invisible, machine-readable watermarking that AI vendors build into their tools, which has until 2 December 2026 for older systems. That is their job, not yours.

Does this even apply to you?#

Almost certainly, if you use AI at work. The Act calls a company that uses existing tools like ChatGPT, Midjourney or Copilot a deployer, as opposed to the providers who build those tools. As a deployer you do not need technical files or conformity assessments, but you do have to disclose AI in the cases below. Purely private use is exempt, though a "personal" post on a work-related topic is in practice business communication.

What you have to label#

Three things, and only these, matter for most marketing and communications:

• Chatbots. If a bot talks to people, they have to be able to tell it is AI and not a person. A line at the start of the chat does it, such as "You are chatting with an AI-powered assistant".
• Deepfakes. Realistic AI-generated or AI-manipulated image, audio or video that imitates real people, places or events has to be disclosed as artificial.
• Fully automated AI text on public matters. AI text published to inform the public on political or societal topics, with no human reviewing it, has to be marked as AI-generated.

What you don't have to label#

This is the part most people get wrong, usually by over-labelling. None of the following needs an AI Act label:

• Your normal product and marketing copy. Sales pages, ads, newsletters, captions.
• AI drafts a person edits and owns. If someone reviews, edits and stands behind the text, it is exempt, even on public topics.
• Clearly illustrative or stylised AI images. Cartoons, obvious 3D renders, fantasy or surreal scenes. These are AI content, but not deepfakes.
• Minor AI edits. Light retouching or small fixes that do not change the meaning of an image.

The line that trips people up: when is an AI image a "deepfake"?#

This is the question marketing teams actually get stuck on, because so much visual content is now AI-assisted.

The test is simple: would a viewer reasonably take it for a real photo or video of a real person, place or event? If yes, and it isn't, it is a deepfake and needs a label. If it is obviously an illustration, a stylised render or physically impossible, it is not a deepfake and needs none.

Two examples make the line clear. A photorealistic "team photo" of people who never posed, or "footage" of an event that never happened, is a deepfake: label it. A cartoon mascot, a clearly stylised hero image or a surreal composite is not: no label. A short caption is enough where a label is needed, such as "AI-generated image, not a real photo".

The line for copy: what counts as "reviewed"?#

For text, the escape hatch is editorial control. If a person deliberately reviews and edits the content and takes responsibility for it, the Article 50 labelling duty does not apply, even for public-interest topics. A quick, purely formal glance does not count as review.

In practice this means almost all marketing content is fine: it is written or edited by a human who signs off on it. The label really only kicks in for content published on autopilot, for example a bot that posts political or societal opinions on its own. For that kind of output, the honest question is whether it should be fully automated at all.

For completeness: these transparency rules sit on top of duties that already apply and never paused, including the AI literacy duty (Article 4, in force since February 2025) and, of course, GDPR.

What to do before 2 August#

You do not need panic compliance. You need a short, clear process:

• Inventory. List where AI images, AI text, video and chatbots appear across your marketing and communications.
• A labelling line. Agree the wording you will use, for example "Image: AI-generated" for realistic visuals and a chatbot notice for any bot.
• A review step. Decide who signs off on AI content before it goes live. That sign-off is also what keeps your AI copy out of the labelling duty.

A voluntary line like "created with AI support, editorially reviewed" is never required, but it can build trust with a sceptical audience.

The bottom line#

The EU AI Act does not make you label every use of AI. It asks for disclosure in three everyday cases, chatbots, realistic deepfakes and fully automated public-interest text, and leaves the rest to sensible internal rules. Get those rules written once and the 2 August date is a non-event.

That is where we come in at Ninja Partners: turning the AI Act's duties into a lean AI content guideline your team actually uses, rather than a document that sits in a drawer. If you want a hand mapping where you stand, start with an AI Readiness Audit.

Sources#

• European Commission, Guidelines on transparency obligations for providers and deployers of AI systems - the official Article 50 guidelines, adopted 20 July 2026
• European Commission, Transparency obligations under Article 50 of the AI Act - Q&A
• EU AI Act, Regulation (EU) 2024/1689 (EUR-Lex, official text), in particular Articles 50 and 113
• European Commission, AI Omnibus enters into force - on the Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026)