Since August 2, 2026, Article 50 of the EU AI Act requires some AI-generated content to be labeled. If your startup uses AI to draft posts, articles, emails, ads or videos, the practical question is which of those need a label, and whether a quick read before publishing is enough. Advice online disagrees. This guide works from the European Commission's final guidelines of July 20, 2026 and their own examples: who is responsible, which channels are usually in or out of scope, what review actually exempts text, and where a label must appear. It ends with a fictional worked example and a review log you can copy. It is general information, not legal advice, and it is useful without a Soloop account.
The short answer
For most startup marketing, the text rule is narrower than many checklists suggest. The Commission lists a company's advertisements and product descriptions as outside it, and one-to-one emails are not published at all. Two kinds of content are different.
- Realistic AI images, audio or video of real people, places or products that could pass as genuine (deepfakes) need a visible label. A digital avatar of your founder and an AI product image that makes the product look better than it is are both deepfakes in the Commission's examples.
- AI-written articles or posts that inform the public about an economic, regulatory, scientific, health, safety or similar development need a label, unless a competent person substantively reviewed and fact-checked them and a named person or company holds editorial responsibility.
- AI tools that talk to people, such as chatbots and agents that send messages, must disclose that they are AI. That is mainly the tool maker's design duty.
- Purely personal use is outside the rules; using AI for your business is not.
Who is responsible for what
Article 50 splits duties between the provider, which is the company that makes or markets the AI tool, and the deployer, which is anyone using an AI system under their authority for professional purposes. If you use an AI writing, image or agent tool for your startup, you are a deployer for what you create and publish with it.
Design AI systems that interact directly with people so the people know it is AI (Article 50(1)), and mark generated audio, images, video and text in a machine-readable way (Article 50(2)). Generative systems already on the market before August 2, 2026 have until December 2, 2026 for the marking duty.
Commission guidelines on Article 50Disclose deepfakes, and disclose AI-generated or AI-manipulated text published to inform the public on matters of public interest unless the human-review exception applies (Article 50(4)). Any activity that brings a regular economic benefit or is part of a business or freelance activity counts as professional.
Commission guidelines on Article 50A deployer outside the EU is covered where it foresees its AI output being used in the EU, for example by directing or authorizing distribution there, including by posting deepfakes on the globally accessible internet. It is not covered when content reaches EU audiences through channels it could not foresee and does not control.
Commission guidelines on Article 50The obligations apply from August 2, 2026. Content published before that date does not need to be labeled retroactively, but text generated earlier and published on or after August 2, 2026 does, if it is in scope.
Commission guidelines on Article 50Up to €15 million or 3% of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises, including start-ups, the lower of the two applies.
Commission guidelines on Article 50Channel by channel: what usually needs a label
The table applies the Commission's own examples to the content a small team typically produces. "Usually" is deliberate: whether text informs the public on a matter of public interest depends on its content, not its format.
| Dimension | Usually needs a label? | Why, according to the Commission's guidelines |
|---|---|---|
| Ads and product descriptions | No, for the text rule. | Listed as outside the text rule, unless they contain claims about health, consumer safety or sustainability. Misleading AI imagery in an ad is a separate deepfake question. |
| One-to-one sales or interview emails | No, for the text rule. | Private professional correspondence is not "published". If an AI agent writes and sends messages itself, the disclosure rule for interactive AI applies instead (last row). |
| Landing pages, changelogs, product how-tos | Usually no. | Inference: commercial information about your own product is closer to a product description than to a matter of public interest. Health, safety or sustainability claims change that. |
| Short promotional posts (launch, hiring, event) | Usually no. | Commercial, and short texts that do not materially communicate knowledge, opinions or facts are not treated as informing the public. |
| Articles, threads or reports on an economic, regulatory, scientific, health or safety development | Yes, unless the human-review exception applies. | The guidelines' in-scope examples include a lifestyle article comparing diets for a disease and a listed company's investor reports. Topics relevant to public debate count. |
| Realistic AI avatar or cloned voice of a real person | Yes. | A realistic avatar of a company CEO congratulating staff is a deepfake example. Human review does not exempt deepfakes. |
| AI product image that changes how the product looks | Yes. | A product image that could mislead about its appearance, characteristics or use is a deepfake. A real product against an AI-generated background is not, if it does not mislead. |
| Clearly fictional or cartoon visuals | No. | Talking mice in a cheese advertisement and a sphinx flying over the Eiffel Tower are the guidelines' examples of content that is not a deepfake. |
| Internal documents and team chat | No. | Organization-internal texts are not published. |
| AI chat or agent that talks with people | It must say it is AI. | Mainly the tool maker's design duty: agents should disclose that they are AI and whom they act for. Outputs a human properly reviews and sends as the main interlocutor are outside this rule. |
What review actually exempts text
The exception for in-scope text has two conditions, and you need both: the text has undergone human review or editorial control, and a natural or legal person holds editorial responsibility for publishing it. The guidelines set a higher bar than "someone looked at it".
- A person with relevant knowledge of the subject examines the substance, not just the wording.
- Every factual claim, number and quote is checked against a source. The guidelines call fact-checking the minimum.
- The reviewer has the authority to approve, change or reject the text, and uses it.
- Spell-checking, grammar fixes, an editorial policy on paper, automated review, AI reviewing AI or a cursory sign-off do not count.
- No substantive AI edit happens after sign-off. If you run the text through AI again, review it again; otherwise the exception is void.
- A named person, role or your company holds editorial responsibility, and its identity and contact details are easy to find, for example on your site's legal or About page.
Where and how to label
When a label is needed, it has to be clear and distinguishable and given at the latest at the first exposure: at the start of the text, at the start of a video, or when someone scrolls onto a deepfake in a feed. People should be able to see or hear it without special tools, so the machine-readable mark your tool adds is not enough on its own. Accessibility rules that already apply to you also apply to the label.
For artistic, creative, satirical or fictional work, the disclosure can be lighter so it does not spoil the work, but it is still required. The Code of Practice on transparency of AI-generated content offers a voluntary way to show compliance.
Example wording, which is our suggestion rather than official text: "This article was written with AI and has not been editorially reviewed." For a video: "The presenter in this video is an AI-generated avatar of [name]."
Worked example: one fictional week of content
Everything below is fictional: the company, the content and the decisions. It shows how to apply the table; it is not a legal assessment, a customer result or output from Soloop.
Tallyroom, an invented tool for small accounting firms, is based in the United States and sells to firms in Spain and Germany. Its founders used AI to draft six pieces of content in one week.
| Dimension | In scope of Article 50(4)? | Decision |
|---|---|---|
| 1. Search ad copy | No: advertisement, no health, safety or sustainability claims. | Publish without a label. |
| 2. Five interview emails, each reviewed and sent by the founder | No: not published. | Send. No label under the text rule. |
| 3. Blog article on new e-invoicing rules for small firms | Yes: informs the public on a regulatory and economic development. | The co-founder, a qualified accountant, checks every claim against the official text and approves it; the company is named as responsible on the legal page. Exception applies, no label; keep the record. |
| 4. X thread summarizing the same rules, posted after a skim | Yes: same topic. | A skim is cursory review. Add the label at the start of the thread, or review it properly first. |
| 5. Demo video with an AI avatar of the CEO | Yes: deepfake. | Label at the start of the video. Review does not exempt it. |
| 6. Real product screenshot on an AI-generated background | No: not misleading about the product. | Publish without a label. |
Copy the content review log
Copy this log into your own document or spreadsheet; no account is required. Fill in one block per piece of AI-assisted content you publish.
| Dimension | Record | Decision |
|---|---|---|
| Piece | [Title or link] / [channel] / [date published] / [AI tool used and for what] | [Ad or product description / not published / informs the public / deepfake / other] |
| Scope | [Topic in one line] / [EU audience foreseeable? yes or no] | [In scope / out of scope, and why] |
| Review | [Reviewer and their relevant knowledge] / [claims checked and sources] / [changes made] | [Exception applies / label needed] |
| After sign-off | [Any AI edits after approval? If yes, re-reviewed when?] | [Label text and position, if labeled] |
| Responsibility | [Person, role or company responsible] / [where its identity and contact appear] | [Next review of this piece, if it will be updated] |
Where Soloop fits
Soloop is an AI cofounder that runs individual go-to-market tasks for a founder. According to the Soloop manual, it can write X posts and replies and publish them through a connected account, send outreach through a connected Gmail account, prepare Reddit posts and launch Google Ads campaigns after setup and payment. When you use it for your company, you decide what is published, so the deployer duties above are yours.
This guide does not assess Soloop's or any other tool's obligations as a provider, and Soloop does not decide for you whether a piece is in scope. A bounded first request: Draft an X thread about [development] for [audience], with a source link for every factual claim, and list the claims I need to check. Do not publish anything until I have reviewed it.
Check the current publishing steps and connections in your workspace, because they can change.
Sources, limitations and refresh
The rules, examples, dates and fines come from Article 50 of the AI Act and the Commission's guidelines of July 20, 2026, read on October 19, 2026. The channel table's "usually" judgments, the label wording, the log and the fictional example are editorial guidance, not legal advice. National authorities interpret the Act; your situation, your contracts and other laws may change the answer.
The Commission describes the guidelines as a first interpretation that it will review. This guide will be reviewed by January 19, 2027, or sooner if the guidelines change.
Commission guidelines on the transparency obligations for certain AI systems under Article 50, C(2026) 5054 final, July 20, 2026.
European CommissionArticle 50, paragraphs 1 to 7: interactive systems, marking, emotion recognition, deepfakes and public-interest text, timing, codes of practice.
EU AI Act, Article 50Voluntary code on marking and labelling AI-generated content; adherence can be used to demonstrate compliance.
European Commission: Code of PracticeThe Soloop manual describes X publishing, Gmail outreach, Reddit drafts and Google Ads campaigns.
Soloop product manual