AI

AI disclosure is becoming a norm, not a policy choice

The EU AI Act's Article 50, IFCN transparency, and newsroom AI policies are converging on the same floor: disclose AI involvement or lose the claim to editorial control.

Two years ago, AI disclosure in newsrooms was a policy debate. In 2026 it is converging into a compliance baseline. The EU AI Act’s Article 50, the IFCN’s transparency requirements, and a wave of published newsroom policies now point in the same direction: content generated or substantially manipulated by AI must be labelled as such — with narrow exceptions for genuinely editorial use.

What Article 50 actually requires

The EU’s transparency article obliges deployers of AI systems to label synthetic and manipulated content — text, image, audio, video — as artificially generated or manipulated. For newsrooms the operational edge is the exception: content produced under human editorial control and carrying editorial responsibility can escape the label only when the human control is real, not nominal. A journalist who drafts with AI assistance, verifies every claim, and takes responsibility for the published piece arguably stays within editorial control. An automated digest pipeline does not.

Why disclosure also serves the publisher

Beyond compliance, disclosure solves a positioning problem. A newsroom that labels its automated output can honestly distinguish it from its human reporting — protecting the premium attached to actual journalism. A newsroom that hides AI involvement converts a product choice into an integrity failure waiting for discovery.

The working rules

  • Disclose the level, not just the fact. “Headline variants generated with AI, reviewed by an editor” is more informative than “AI-assisted.”
  • Label at the artifact. Disclosure on the article/image/video, not only on an about page.
  • Never feed the machine what you can’t defend. Unpublished sources, embargoed material, and confidential documents don’t go into external AI systems — the disclosure problem is dwarfed by the source-protection breach.
  • Name the accountable human. Automated output still needs a byline of responsibility — an editor, not “the system.”
  • Log it. AI involvement in the editorial record is what makes the disclosure claim auditable later.

The harder question nobody’s solved

Where is the line between “AI assistance” and “AI generation”? Spell-check, translation, transcription, summary, draft — the spectrum is continuous. The workable test most policies converge on: would a reasonable reader feel deceived if they knew how much of this was machine-made? If yes, disclose. If the answer is “we’re not sure,” disclose anyway — the cost of over-disclosure is a label; the cost of under-disclosure is trust.

MediaMeter context: AI disclosure is LEG-08 and ETH-10 in our standards registry, and automated-news (type #52) requires an ai-disclosure element in the taxonomy. Detection of undisclosed AI content is specified as a scoring impact with mandatory human review.

MediaMeter Editorial

Written by the MediaMeter editorial team. We publish our methods, limitations, and corrections.

https://mediameter.eu/resources/blog/ai-disclosure-norms-emerging-standard/
Newsletter

Monthly notes on journalism quality.

One email a month. Methods, limitations, and what we are learning. No product spam.