The law focuses on Article 50, which governs how AI tools disclose their presence. If you use generative tools for reporting or production, these rules apply to your workflow now. Systems placed on the market before August 2, 2026, get a grace period until December 2, 2026, to meet the machine-readable marking rules.

Labeling content and chatbots

The first rule covers chatbots and interactive tools. Under Article 50(1), anyone using AI assistants must tell users they are talking to a machine. This notice must appear at the latest during the first interaction. If you use a bot for customer service or search, readers must see a clear notice immediately. Per Article 50(5), the notice must be noticeable, easy to understand, and clearly separated from other content. It cannot be buried inside terms of use or hidden behind menu layers.

Strict requirements also apply to synthetic and manipulated media. According to the second clause of Article 50, providers of generative systems must ensure outputs carry machine-readable markers. These signals must be integrated into synthetic audio, image, video, and text outputs. They include watermarks, metadata, cryptographic methods, or digital fingerprints. Think of this as giving the file a technical label that search engines and browsers can detect.

Deployers are required by Article 50(4) to label deepfakes clearly. These are AI-generated audio or video that looks real. The label must be visible or audible to the audience. It cannot be hidden in technical file data that humans cannot see. Such disclosures must not require special software or tools to detect them.

The ‘editorial control’ exemption for text

The most important rule for journalists concerns written articles. Any AI-generated text published to inform the public on public interest matters must be disclosed. But the law carves out an exemption designed to protect normal journalism. You do not need to label AI-assisted text if it underwent genuine human review or editorial control. A named person or entity must also accept full editorial responsibility for the piece.

Meeting this bar takes real work. Per the law, superficial spell-checking and simple grammar fixes do not qualify as genuine human review. The editor must genuinely shape and approve the finished piece before publication. And here is the trap: any AI edit made after human sign-off immediately revokes the exemption. The story then needs mandatory disclosure.

What you don’t need to label

The new rules draw a sharp line between light assistance and major automated drafting. Standard editing tasks are exempt from the Article 50(2) machine-readable marking duties. These safe tasks include basic proofreading and syntax fixes, plus minor stylistic polishing. Other permitted tweaks include language translation, noise reduction, and color correction. If a copy editor runs a spell-check or cleans up audio, no special tag is needed.

The burden shifts when an algorithm takes over the narrative framework. Structural paraphrases, new machine-written summaries, and substantive rewrites do require marking. When software generates new text blocks or condenses long documents into fresh summaries, editors must apply the proper disclosures. Your technical team needs a clear policy on where light polishing ends and algorithmic writing begins.

These compliance steps are not optional style suggestions. Violations of Article 50 transparency obligations carry serious fines. The maximum is €15 million or 3% of global annual turnover, whichever is higher. For small and medium-sized enterprises and start-ups, the fine is the lower amount of those two figures. Newsroom leaders should update internal labeling guidelines now and audit every live chatbot to stay within the law.

Written by Dominik Czarnota using the Tribune Desk AI platform. Every claim in this article was fact-checked against its sources, and an editor read, edited and approved it before publication.