The Label It or Not Checklist

01 The tool's job, and yours

The EU AI Act's labeling rules, Article 50, have applied since August 2, 2026. On December 2, 2026, the grace period ends for tools already on the market before that date. By then, the image generator, the voice tool and the chatbot you pay for must mark their output in a way software can read.

That mark is the tool's job, not yours. Your job is three narrower things, and most of what you make with AI isn't one of them:

A bot that says it's a bot. Fakes that say they're fake. Public-interest posts that a person actually stood behind.

Not in Europe? The Act covers companies outside the EU when what their AI produces is used in the EU. Selling to Europeans, advertising to them, answering their questions with a bot: that can count.

This isn't legal advice. It's the text of the law, read plainly, with the dates checked against the official version.

02 Thing one: chatbots

If people talk to an AI, they have to be told it's an AI, at the latest at the first interaction. Strictly, that duty sits with whoever builds the chatbot. But if you put one on your website under your own name, assume it's yours.

Say it in the first bubble. Not in your terms of service, not in small gray text after five messages. The first thing they see.

Say it plainly. In Paul Weiss's reading of the Commission's guidelines, a vague line like "this system uses language models" isn't enough.

Agents say who they act for. An AI agent has to say it's an AI and who it's acting for.

03 Thing two: deepfakes

If you use AI to make an image, a video or audio that looks like something real and would pass as real, you have to say it's AI. "Real" is wider than famous faces: the law's definition covers people, objects, places and events. You don't have to mean to fool anyone. Paul Weiss reads the guidelines as an objective test: would a viewer take it as real?

What you made Label it?

An AI illustration on your blog Generally no label from you. The tool marks it.

Your product dropped onto an AI beach in an ad Grey zone. If the scene passes for a real place or a real photo shoot, label it.

Your actual storefront, rebuilt by AI with a line out the door Yes.

A voice that sounds like your founder reading an ad they never recorded Yes.

An AI "customer" giving a testimonial Yes. Or better, don't make it.

Audio you cleaned up with AI so it sounds less echoey No. That's standard editing.

Clearly a cartoon, a fantasy, a robot with a cape No. Not a deepfake.

For obviously creative or satirical work, the duty gets lighter: you still say it, just in a way that doesn't ruin the piece.

Where the label goes

Where people actually see it, the first time they see the content. In Paul Weiss's reading, a note in your description or at the very start may not be enough for a long video or a podcast, where people join halfway. You don't have to design it: the EU has made a set of icons for labeling AI-generated content.

04 Thing three: public-interest text

If you publish AI-written text to inform the public about something of public interest, think health, money, safety, the news, you have to disclose it. Unless a real person reviewed it and your business takes editorial responsibility for it.

Is it public-interest text? Health, money, safety, the news. A product description, or a chatbot answer only one customer sees, isn't.

Did a real person review it? According to Paul Weiss, running spell-check doesn't count as review.

Does your business stand behind it? No review and no editorial responsibility: disclose it.

05 Before December 2

Ask the AI tools you pay for two questions. Their answer tells you what's already being stamped on your content. The code is the Commission's voluntary Code of Practice; about 190 organisations had signed it by the end of July 2026.

Two questions for each AI vendor

Hi [VENDOR NAME],

We use [PRODUCT / PLAN].

1. Have you signed the EU's Code of Practice on labeling AI content?
2. When does marking reach the version we use?

Thanks,
[NAME]

Why your label still matters. These marks are fragile. In OpenAI's own tests, swapping a quarter of the words in a watermarked passage cut detection from about 92% to 17%. The mark is a default, which is why the law leans on your label, not just theirs.

Three checks this week

Does your chatbot say it's an AI in its first message?

Do you have anything that looks real and isn't, without a label?

Does a real person sign off on anything you publish about money, health or safety?

Fines for getting labels wrong go up to 15 million euros or 3% of global revenue, whichever is higher; for small businesses, whichever is lower.

Sources

Regulation (EU) 2024/1689, the AI Act (Official Journal) Regulation (EU) 2026/1744, the Digital Omnibus on AI (Official Journal) Paul, Weiss: EU finalises transparency rules for AI-generated content European Commission: Code of Practice on Transparency of AI-generated Content European Commission: Strong backing for the Code of Practice VKTR: OpenAI adds invisible watermarks to ChatGPT text in the EU Groweo: EU AI Act transparency rules for small businesses

Mathew's LibraryLibrary card

Free access

Keep reading free with your library card

One email opens this checklist and all 3 resources in Mathew's library. No account, no password.