EU’s AI labels come with exceptions

"New EU rules require AI-generated content to be labelled." That sentence ran in dozens of European newsrooms last weekend, and it is only partly right. The EU AI Act's transparency rules took effect on 2 August (EU AI Act), but they do not slap a warning on every spell-checked email or AI-polished photo. The obligations depend on who is using the system and what kind of content reaches the public.

Consider three situations. A company builds a chatbot that talks to customers. A newsroom uses an AI tool to help draft an article about an election. A political campaign posts a synthetic video of a rival candidate. Each triggers different rules, aimed at different actors, and only some of those rules produce a label visible to the public.

Builders label the pipes, users label the product

The Act splits duties between two roles: the provider, whoever builds the AI system or puts it on the EU market, and the deployer, the organisation or professional that uses it (EU AI Act). One organisation can hold both. A bank that builds its own customer chatbot is both provider and deployer, and owes both sets of duties.

Think of it like food safety: the manufacturer prints the nutritional panel, the restaurant flags allergens on its menu. Both obligations exist; neither replaces the other (Stibbe).

Providers carry design-stage duties. A chatbot or voice assistant must tell users they are dealing with AI at the start of the conversation (EU AI Act). Systems that generate synthetic images, audio or video carry a different obligation: embedding a hidden tag in the file — closer to a barcode than a warning label — so regulators and detection tools can spot AI-generated content (Commission guidance).

Deployers owe the disclosures audiences actually see. A campaign posting a deepfake video must tell viewers it is synthetic. Separately, a newsroom or agency publishing AI-generated text about elections, public health or government spending must disclose that fact (Stibbe). The two sets of duties run in parallel: meeting one does not satisfy the other.

If you are using ChatGPT to brainstorm dinner ideas, none of this applies. The rules change workflows for newsrooms, campaigns, companies and public agencies that publish AI-generated content to audiences.

Your grammar checker is fine

Everyday AI tools get a clear pass. A grammar checker, spell-checker or translation tool does not trigger marking duties as long as it does not substantially alter the meaning of what you wrote (Commission FAQ). Personal, non-professional AI use falls outside the rules entirely (EU AI Act).

The editorial carve-out matters for newsrooms. AI-generated text on public affairs can skip the visible label if a human editor genuinely reviewed it and a named person or organisation takes editorial responsibility (Commission FAQ, Canal CEO). Italian and Spanish legal commentary both stress that this responsibility must be documented and real, not a rubber stamp (Corriere della Sera). Deepfakes in artistic or satirical works still need disclosure, but in a form that does not wreck the viewing experience.

Live now, with one grace period

From 2 August, chatbots must identify themselves, deepfakes need disclosure, and AI text on public-interest topics needs a label unless a real editor takes responsibility. The single grace period runs until 2 December 2026 and covers only one thing: the hidden-tag marking duty for generative AI systems already on the market before that date (Cooley, FAZ). New systems must comply immediately.

National authorities handle enforcement. Germany's Bundesnetzagentur already accepts complaints through a free online form (Bundesnetzagentur). Ireland has stood up a national AI Office as a coordination hub (RTÉ). Spain's AESIA is operational (AESIA). Fines for transparency breaches can reach up to €15 million or 3% of global annual turnover, below the €35 million / 7% ceiling reserved for the Act's most serious prohibitions (Commission summary, Handelsblatt).

What to watch now: whether regulators act on early complaints, whether those hidden file tags survive the normal internet cycle of compression, screenshots and re-uploads, and whether audiences learn from disclosures or simply scroll past them.