From 2 August 2026, Article 50 of the European Union’s AI Act requires greater transparency when artificial intelligence is used to produce news content.
The regulation – as the FAQs make clear and also the European Commission’s guidelines – does not ban AI journalism, nor does it require every AI-assisted article to carry a warning. The decisive questions are what the technology has produced, how the material has been checked and who is responsible for publishing it.
In most news organisations, the publisher — rather than each reporter — is considered the “deployer” of the AI system. Journalists working under the company’s authority are not normally treated as separate deployers.
GENERATED OR MANIPULATED
A news outlet must label AI-generated or substantially AI-manipulated text published to inform the public about matters of public interest if that material has not received genuine human review or editorial control.
For example, a website that automatically publishes an AI summary of a town council decision must identify it as artificially generated. The same applies to an automatically produced article about an election, a health emergency or a listed company’s results.
A label is not required when a knowledgeable journalist or editor has examined the substance, checked the facts and sources, and has the authority to change or reject the text. The publisher or another identifiable person must also assume editorial responsibility.
GRAMMAR CHECK
A quick grammar check is not enough. Nor is approval by a second AI system. If AI substantially rewrites an article after the human review, the exemption is lost and the final version must be checked again.
The rules are stricter for deepfakes created or altered by the newsroom. If a podcast clones a presenter’s voice, or a video realistically depicts a politician making a statement that was never made, its artificial origin must be disclosed — even if an editor approved it.
The warning must be clear, visible or audible when readers or viewers first encounter the material. Hidden metadata alone is insufficient.
Routine assistance does not automatically trigger a label. Spell-checking, minor stylistic corrections, transcription, limited cropping, colour adjustment or face-blurring are treated as standard editing when they do not change the meaning of the content. An AI translation can also remain unlabelled if it receives proper human review.
PUBLIC-FACING CHATBOT
A newsroom operating its own public-facing chatbot or AI avatar must make clear from the start that readers are interacting with a machine. People must also be informed if a media organisation exposes them to an emotion-recognition or biometric-categorisation system.
Technology providers have a separate duty to place machine-readable markers in synthetic content. A newsroom that develops and releases an AI system under its own name may therefore become a provider as well as a user.
Breaches can expose operators to fines of up to €15 million or 3% of their worldwide annual turnover. Human oversight is therefore more than an ethical safeguard: from 2 August, it can determine whether an AI label is legally required.
