AI in media — Article 50 labelling

Creating content with full source control

Knowledge and documents

Problem: Verifying where the information used in a piece comes from takes longer than writing it

How it works: A live preview of the knowledge sources used in every answer — the journalist sees which document each piece of information comes from

What you gain: Fact-checking starts from a list of sources rather than hunting for them after the fact — fewer corrections and a shorter path to publication

Multichannel content with a consistent message

Process automation

Problem: The same piece has to be prepared in several formats while keeping the message consistent

How it works: A central, versioned prompt register and queries decorated with standard editorial rules; ready-made prompts available via a shortcut in the chat

What you gain: The website, newsletter and social media versions come from one piece, according to the newsroom’s rules, not the author’s mood

Generating and editing images in one environment

Process automation

Problem: Preparing images for content requires a separate tool and a separate licence

How it works: Generating and editing images with integrated visual models in the same working environment

What you gain: The illustration is created where the text is — one licence, one log, one cost control

Testing models and prompts for content quality

Quality and standardisation

Problem: Nobody knows which prompts and models produce the best texts

How it works: A feedback loop — answer ratings and a ratings register; model evaluation (LLM-judge) and adaptive A/B testing of prompts

What you gain: The newsroom chooses the model and prompt based on ratings, not opinions — text quality improves measurably

AI cost control per newsroom and project

Cost and efficiency

Problem: AI costs spread across newsrooms without control

How it works: Budgets per team and project with alerts; a semantic cache that cuts the cost of repeated queries

What you gain: Each newsroom has its own AI budget and sees how much it has used; repeated questions do not cost a second time

Choosing the first process for AI

Adoption and skills

Problem: The newsroom does not know whether it is ready for AI or which process to start with

How it works: The Initial Assessment (CDF-Eval) with an AI Readiness Index; a Cognitive Sprint (F4) delivers the first value in the chosen editorial process in 2–4 weeks

What you gain: The first process is chosen on the basis of a diagnosis, and the pilot ends with a decision, not another pilot

Automatic labelling of AI-generated content

Compliance and oversight

Problem: The law requires labelling AI-generated content

How it works: AI Disclosure Injection (Article 50 of the AI Act) — automatic labelling recorded in the log

What you gain: The labelling obligation is fulfilled by the system, not by the journalist’s memory — and you can prove it

From when must AI-generated content be labeled?

From 2 August 2026 (Article 50 of the AI Act). For systems placed on the market before that date, machine-readable marking is required by 2 December 2026

Does text under editorial control need to be labeled?

Text published to inform the public on matters of public interest does not need to be labeled if it has undergone editorial review and a person or newsroom holds responsibility for the publication (Article 50(4)). It is worth documenting this review

Who is responsible for labeling — the agency or the brand?

The entity that publishes the content, as the deployer of the AI system. It is worth setting out the scope in the contract with the agency