AI Policy

Label Your AI Text or Give It an Editor

The Commission published the final Article 50 transparency guidelines on 20 July 2026. From 2 August, a deployer who publishes AI-generated text on a matter of public interest without human editorial review has to disclose it.

AI PolicyBy Servola Tech Desk2026-07-214 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Key takeaways

The European Commission published the final Article 50 transparency guidelines on 20 July 2026, thirteen days before the obligations apply, and those duties do not stop at the model makers: from 2 August a deployer who publishes AI-generated text on a matter of public interest without human editorial review has to say so.
Human editorial review is no longer a quality preference, it is the switch that turns the disclosure duty off. Decide this week which of your published channels has a named human reviewer and which does not, because that list is the boundary of your exposure.
You cannot do the machine-readable marking yourself. Marking AI-generated or manipulated output so it is detectable in machine-readable form is a provider duty, so your compliance depends on the tool you bought. Put the question to every generative vendor in writing before 2 August.
Deepfakes, emotion recognition and biometric categorisation carry deployer disclosure duties too. If you run any of them in a European workplace or customer flow, the person in front of the system has to be told, and that is a process change rather than a settings change.
The Code of Practice on Transparency of AI-Generated Content is voluntary and has been final since 10 June 2026. Signing it does not replace the law, but it is the cheapest available evidence that you took the marking requirement seriously before anyone asked.

The document that landed thirteen days early

The Commission does not often publish guidance with this little room to spare. On 20 July 2026 it approved and published the Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act. The obligations those guidelines explain apply from 2 August 2026. That is thirteen days between the final text and the date it describes.

The draft had been public since 8 May and the consultation on it closed on 3 June, so none of this is a surprise to anyone who was reading. What changed on Monday is that the interpretation stopped being provisional. Guidelines of this kind do not create obligations, because Article 50 already did that. They set out how the Commission and the national authorities intend to read it, which is the part a European operator needs before it commits money to a process.

Two duties, and they land on different desks

The split is the whole story. Article 50 puts one set of duties on providers and a different set on deployers. Providers have to make sure a system built to interact directly with a person tells that person they are dealing with an AI, and they have to mark AI-generated or manipulated output so that it is detectable in a machine-readable form. Deployers carry a narrower but more personal list: tell people when they are looking at a deepfake, when they are exposed to an emotion recognition or biometric categorisation system, and when they are reading AI-generated text published on a matter of public interest without human review or editorial control.

Most European businesses have spent the past year reading the AI Act as a supplier problem, and our own coverage of the 2 August date has pointed the same way, because the general-purpose model rules that bite on that day are aimed squarely at the labs. The guidelines close that comfort. The deployer half of Article 50 attaches to the company that puts the system in front of its own public, and there is no size threshold underneath it.

The marking you cannot do for yourself

The machine-readable marking requirement is worth isolating, because it is the one duty an operator cannot discharge by changing its own behaviour. Embedding provenance so that a third party can detect that output was machine-generated has to happen where the output is produced, which is inside the vendor system. If the tool you licensed does not mark what it emits, no policy written internally will put the mark there.

That turns a technical question into a procurement question with a date on it. Ask every generative vendor in your stack, in writing, whether their output carries machine-readable marking today, in what form, and whether they have signed the Commission Code of Practice on Transparency of AI-Generated Content. Those answers are cheap to collect now and expensive to reconstruct after an authority asks for them. A vendor that cannot answer inside a fortnight is telling you something about the next twelve months.

The carve-out is really an instruction

Read the deployer duty on AI-generated text closely and you find a condition rather than a prohibition. The disclosure obligation attaches when text on a matter of public interest is published without human review or editorial control. Where a named person reviews the piece and takes responsibility for it before it goes out, the duty does not arise the same way. The law is describing an editorial process and pricing the absence of one.

For anyone running a content operation, a corporate blog, a newsroom or a customer-facing knowledge base, that is a cleaner instruction than most regulation delivers. Work out which channels have a human editor of record and which do not. The ones that do not are either labelled or given a reviewer. There is no third option that survives an inspection, and building the reviewer role is the version of this that also improves the output.

What to have in place before 2 August

Three artefacts, none of which needs a lawyer to start. A one-page inventory of every place your organisation shows AI output to a person outside it, recording for each whether a human reviews it before publication. A written answer from each generative vendor on machine-readable marking. A short internal note recording who decided which channels get a label and when, because a dated decision is most of what an authority wants to see from a company of ordinary size.

None of that is expensive. The expense arrives if 2 August passes with nobody in the building able to say which of your public-facing text is machine-written and who checked it. Thirteen days was not much notice. It is enough for a list, and a list is where the next conversation will start.

Frequently asked questions

Do the guidelines create new obligations?

No. Article 50 of the AI Act creates the obligations and they apply from 2 August 2026. The guidelines published on 20 July 2026 set out how the Commission reads the scope of those obligations for providers and deployers. They are guidance rather than binding law, but they are the clearest available statement of how enforcement will approach the text, which makes them the practical standard to plan against.

We only use AI to draft, and a person edits everything. Do we have to label it?

The deployer duty on AI-generated text attaches to material on matters of public interest published without human review or editorial control. If a named person genuinely reviews the work and takes editorial responsibility before publication, that is the condition the obligation turns on. Write down who that person is for each channel, because the claim is only as good as your ability to evidence it later.

Is signing the Code of Practice worth it?

The Code of Practice on Transparency of AI-Generated Content was finalised on 10 June 2026 and is voluntary, open to providers and deployers of generative AI systems. Signing does not replace compliance with Article 50 and does not immunise anyone. It is useful as dated evidence of intent, and as a way to align on marking techniques with the rest of the market instead of inventing your own.

Thirteen days is short notice for a rule and long notice for a list. Write the list.

EU AI ActArticle 50European CommissionAI transparencydeepfakesCode of Practice

America Weighs an AI Watchdog the Labs Would Fund

Treasury helped shape a plan for an SEC-supervised AI watchdog modeled on FINRA. Why it is the structural opposite of the EU AI Act, and what it means for owners.

3 min read

AI Policy

The AI Rulebook 29 Nations Signed, and the EU Did Not

On 16 July 2026 in Shanghai, 29 countries signed to establish WAICO, a new intergovernmental AI body. No EU member joined. Why owners should plan for governance fragmentation.

4 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.