// sample verdict · Article 50(1), 50(2), 50(4)
This is the document you receive. Every duty is checked against the stated facts, the legal text is quoted word for word, and the last page says what the company is not required to do.
| Question put to you | Answer on record |
|---|---|
| Reader assistant | “Ask Fernwick Media”, built in-house on a third-party model via API, live since April 2026 under your name. Answers reader questions and summarises articles for the reader who asks. |
| Is it labelled as AI? | Only in the Terms of Use. The chat window says nothing. |
| News briefs | Short summaries of your own articles, including local council decisions, written by a subscribed third-party tool. Overnight briefs publish automatically; daytime briefs are approved by the duty editor. |
| Images | From a commercial image generator: photorealistic scenes of real local places, and clearly drawn editorial cartoons. |
| Editorial responsibility | The publishing company; the editor-in-chief is named on the About page. |
If any line above is wrong, say so and the verdict is rewritten at no charge.
| Duty | Who carries it | On these facts | From |
|---|---|---|---|
| 50(1) · tell readers they are talking to an AI | You, as provider of the assistant | applies · gap | 2 August 2026 |
| 50(2) · machine-readable marking of the assistant’s output | You, as provider of the assistant | applies | 2 December 2026 |
| 50(2) · marking of the images and the briefs | The providers of those tools | not yours | — |
| 50(4) · label on photorealistic scenes | You, as deployer | applies | 2 August 2026 |
| 50(4) · label on editorial cartoons | — | not a deep fake | — |
| 50(4) · label on overnight briefs | You, as deployer | applies | 2 August 2026 |
| 50(4) · label on editor-approved briefs | You, as deployer | exception met | — |
Why you are the provider. The Commission’s guidelines give this exact case as an example of a provider: an organisation “that has developed an interactive AI system (e.g. chatbot) in-house and puts it into service in the Union for its own use and under its name or trademark.” Building on a third-party model does not move the duty to the model’s vendor.
Why the Terms of Use are not enough. The same guidelines list “Disclosures contained only in terms and conditions, URLs, or documentation” among the techniques that are insufficient on their own: they “may complement, though not replace, in-context disclosure.” Article 50(5) sets the timing: “at the latest at the time of the first interaction or exposure.”
What closes the gap. One clear notice in the chat window, before the first answer. The guidelines say a single, prominent notification “is likely to suffice in most instances”.
It reaches the assistant. The assistant generates text, and the guidelines confirm the paragraph covers text as well as audio, image and video. Its summaries do not escape through the editing exception: the guidelines list “AI-generated summaries of text” among the “semantic changes that require marking under Article 50(2) AI Act”.
The date is still ahead of you. The assistant was on the market before 2 August 2026, so the transitional rule applies:
The images and the briefs come from tools other companies built and sell. You use them as they are; you did not put them on the market under your name. Article 50(2) speaks to providers, so the machine-readable marking of those outputs is the job of the companies that make the tools. It is worth asking each vendor, in writing, how they mark what they generate. It is not a duty this paragraph places on you.
The photorealistic scenes are deep fakes. They show real local places, rendered to look like photographs, next to news stories. That is the definition, line by line. Each one needs a visible disclosure.
The editorial cartoons are not. Among its examples of content that does not constitute a deep fake, the guidelines list an “AI-generated cartoon of a pre-existing image depicting a historical event.” A drawing that reads as a drawing does not falsely appear authentic.
Your briefs are squarely in scope. The guidelines give, as an example of text that must be labelled unless reviewed, an “AI-generated summary of a human-authored article on a newspaper’s website discussing a recent decision by a town council.” That is your brief on a council decision, almost word for word.
| Condition | Overnight | Daytime | Why |
|---|---|---|---|
| i · Human review or editorial control | not met | met | Overnight briefs publish with no one reading them. Daytime briefs are approved by the duty editor, who can change or reject them. |
| ii · Someone holds editorial responsibility | met | met | The publishing company holds it, and the editor-in-chief is named on the About page. |
The guidelines set the bar for the first condition: “Fact-checking the accuracy of the content is a minimum requirement that should be part of that review.” A glance before publishing is not the same thing. If the duty editor checks the facts, the daytime briefs need no label; the overnight ones do, until someone reviews them.
The assistant’s summaries sit outside this paragraph. The guidelines place a “News summary generated by a chatbot that is only available to the user that prompted the chatbot” outside Article 50(4). They are still covered by your marking duty under Article 50(2).
This is the section nobody writes, and the reason a verdict is worth paying for.
One product, seven questions by email, no calls. Three working days after your answers, you get a written verdict on which Article 50 duties apply and what follows if they do. $200, through Upwork escrow. Questions first: andrei@logium.co.
One product, seven questions by email, no calls. Three working days after your answers, you get a written verdict on which Article 50 duties apply and what follows if they do. $200, through Upwork escrow.
Questions first: andrei@logium.co