What the EU AI Act actually requires of you, and why it is almost nothing

Somebody has probably told you that from 2 August 2026 you must label your LinkedIn posts as AI-assisted or risk a fine from Brussels. I have seen that claim in webinar invitations, in agency newsletters, and inside a compliance checklist being sold to executives. It is wrong, and it is wrong in a way anyone can check in an afternoon, because the European Commission published its own FAQ on the article everybody is citing.

The dates and the money are real. Article 50 of the EU AI Act, Regulation (EU) 2024/1689, becomes applicable on 2 August 2026, and non-compliance can attract fines up to EUR 15 million or 3% of worldwide annual turnover. What is not real is the idea that this lands on a founder writing about pricing, or on the person who writes for her. The text-labelling duty inside Article 50 is scoped to text published to inform the public on matters of public interest, and the Commission names those areas: politics and democratic processes, public administration and services, administration of justice and law enforcement.

So the honest answer to the question of what the EU AI Act requires of you in August 2026 is, for most people reading this, almost nothing. Not because the law is soft, but because it was aimed somewhere else. What follows is what the article covers, what the Commission says about human review, who genuinely has work to do before 2 December 2026, and the rule that will actually govern how you write, which belongs to LinkedIn rather than to Brussels.

What Article 50 says, and when the clock starts

Start with what is not in dispute. The transparency obligations under Article 50 become applicable on 2 August 2026. There is a limited grace period running to 2 December 2026 for the marking and detection obligation on systems already on the market, and the penalty ceiling for non-compliance sits at EUR 15 million or 3% of worldwide annual turnover. Those figures come from the European Commission, and they are the part of the story that the panic gets right.

Now notice where that fine sits in every version of this you have been shown. It arrives in the opening line, before anybody has established that an obligation attaches to you at all. That is the structure of a sales page rather than the structure of a legal analysis, and once you see it you cannot unsee it. The order should run the other way, working out whether the duty lands on you and only then looking at what happens if it does.

It also matters what has not happened yet. Enforcement practice has not materialised, and there has been no first action against a platform or a professional deployer. Nobody can tell you how a regulator will read the difficult edges of this, because no regulator has read them out loud.

1. 24 June 2026
   Law firm preparation guidance publishedSidley Data Matters, 'EU AI Act Transparency Obligations: Preparing for Compliance by 2 August 2026'
2. 30 July 2026
   LinkedIn ships its 'Seems like AI slop' report buttonThe same announcement discontinues the AI 'enhance your post' writing feature
3. 2 August 2026
   Article 50 transparency obligations become applicableNon-compliance can attract fines up to EUR 15 million or 3% of worldwide annual turnover
4. 2 December 2026
   Limited grace period endsApplies to the marking and detection obligation on systems already on the market

Serious preparation work does exist, and it does not sound like the panic. Sidley's Data Matters published a note on 24 June 2026 titled 'EU AI Act Transparency Obligations: Preparing for Compliance by 2 August 2026', and the Commission put out guidelines on transparency for AI-generated content alongside the FAQ. The text of Article 50 itself sits on the public web. Reading it takes less time than sitting through the webinar being sold to explain it to you.

Is a LinkedIn post 'information of public interest'?

This is the question the entire panic rests on, and it has an answer rather than a debate. Per the Commission's Article 50 FAQ, the text-labelling duty in Article 50(4) covers text published to inform the public on matters of public interest, and the areas named are politics and democratic processes, public administration and services, administration of justice and law enforcement. Commercial and professional content in general is not on that list.

Hold that against what a business post usually is: a founder explaining why onboarding stayed broken for two years, a finance lead describing how she thinks about renewal risk, a consultant arguing that a popular metric measures the wrong thing. None of that is politics, public administration, justice or law enforcement. It is commercial communication in a commercial market, which is governed by a great deal of other law and by the platform, but not by this particular labelling duty.

The European Commission's own Article 50 FAQ scopes the text-labelling duty to text published to inform the public on matters of public interest, naming politics and democratic processes, public administration and services, administration of justice and law enforcement, which does not cover commercial or professional content generally.

I want to be careful about scope, because scope is precisely what gets lost when a claim travels. The Commission did not say the AI Act ignores commercial activity. It said this labelling duty is written this way, for this category of text. Those are two different sentences, and the distance between them is where most bad compliance advice lives. When somebody widens 'this duty covers public interest text' into 'you must now disclose AI on everything', they have not simplified the rule, they have replaced it with a different one and kept the article number for credibility.

It is the legal equivalent of an appeal for leg before when the ball pitched outside leg stump. Loud, confident, and not out.

The human review carve-out is not a loophole

People hear the word exemption and assume a technicality. Read the sentence again: deliberate examination of the substance, by a person with relevant knowledge. That is not a checkbox you tick on a form, it is a description of somebody who understands the subject reading the work against reality and deciding whether it holds. Written down like that, the carve-out stops sounding like an escape hatch and starts sounding like the job.

Here's the thing. That carve-out only fails for one kind of process, the one where a prompt goes in, output comes out, and it goes onto a publishing schedule without anyone who knows the subject reading it. If that is your workflow, the regulation is not your largest exposure. Your largest exposure is that you are publishing claims nobody has checked under a real person's name, and the first time one of them is wrong in public, a European fine will be the least of your problems.

If your workflow cannot survive the phrase 'deliberate examination of the substance by a person with relevant knowledge', the regulation was never your problem.

• Who the duty lands onAnyone posting with AI helpSystems and deployers, not authors
• Which text is coveredAll professional and commercial postsText published to inform the public on matters of public interest
• Effect of editingA label is required regardlessText under human review or editorial control does not need labelling
• Personal postingCaught from 2 August 2026Purely personal, non-professional use falls outside the deployer definition

There is a third leg to this and it is the one people skip. Purely personal, non-professional use falls outside the deployer definition entirely, per the Commission's FAQ. An individual posting on their own account in a personal capacity is therefore not a deployer in the sense the article means. The professional case is a separate question, and even there the scope test has to be met before the labelling duty bites at all.

So who genuinely has something to do before 2 December 2026?

Almost nothing is not the same as nothing, and I would rather be accurate than reassuring. If you build or operate an AI system that generates text published to inform the public on matters of public interest, you are looking at the intended target of the labelling duty, and you should be taking legal advice rather than reading a personal branding guide about it. If you operate systems that were already on the market, the marking and detection obligation carries that limited grace period ending 2 December 2026, and the date belongs in your calendar rather than in a blog post.

For everybody else, the practical obligation is not a badge. It is being able to describe how the work gets made if somebody asks, which is a much lower bar than a compliance product and a much higher one than most content operations currently clear.

The other thing to watch is how the words get interpreted rather than what they say. Enforcement practice is what turns statutory text into a working norm, and the first action against a platform or a professional deployer will set the practical disclosure standard far more firmly than the article does. The specific thing to look for is whether 'human review' is read generously or narrowly.

The rule that will actually bind you is LinkedIn's

While everyone was reading Brussels, the platform moved, and it moved further than the regulator did.

A platform that profits from posting volume just discontinued its own AI writing feature.

LinkedIn also announced that it will begin privately flagging, inside a user's own dashboard, when people believe their content is coming off as inauthentic due to heavy AI use, and that it is expanding profile verification. The private flag is the one I would watch. It is the first platform-side reputational signal about AI use aimed at the author rather than at the audience, and its visibility, appeal process and ranking consequences have not been specified.

In its policy post 'Keeping conversations real on LinkedIn', executive editor Laura Lorenzetti wrote that 'When AI is overused, especially at scale and in an automated way, it dilutes the valuable insights that real human conversations can spark', that it is 'ok to use AI to help you write' provided posts represent your own voice and perspectives, and that LinkedIn has technology systems 'trained to recognize signals of AI slop'. LinkedIn has also said it blocks hundreds of thousands of automated comment attempts daily and millions of other automation attempts in recent months, which are LinkedIn's own figures and have not been independently audited.

Now read the platform rule and the Commission's carve-out next to each other. One says text that has undergone human review or editorial control does not need labelling. The other says it is fine to use AI to help you write as long as the post represents your own voice and perspectives. Two institutions with entirely different incentives, one a regulator and one a company that earns money when people post more, arrived at the same standard, which is that a human has to own the substance. That convergence tells you more about where this is going than either document does on its own.

• The lawEU AI Act Article 50, applicable 2 August 2026, aimed at systems and deployers, with the text-labelling duty scoped to public interest content
• The platform ruleLinkedIn's authenticity push announced 30 July 2026: slop reporting, private dashboard flagging, expanded profile verification
• The audience reactionPeer-reviewed evidence that AI disclosure reduces perceived credibility, consistent across science communication, news and advertising
• The detectorTools with documented false-positive rates on human-written text, used by third parties to accuse you with no appeal process

You will also see numbers attached to LinkedIn's changes. The claims that LinkedIn's authenticity update reduces AI content reach by up to 47%, and that AI posts get 45% less engagement, are both circulating widely. There is no LinkedIn source for either. They appear only in vendor and agency blogs with no stated method, and what LinkedIn has actually published is architecture detail about its feed, in its engineering blog post of 12 March 2026, not reach deltas by content type. If you are quoting a reach penalty to a client, you are quoting a number somebody invented.

The accusation will come from a detector, not a regulator

This is the exposure almost nobody prices. In practice the person who decides you used AI is unlikely to be an official in Brussels. It will be a competitor, a prospect, a journalist or a bored commenter running your writing through a detector, and detectors are considerably worse than their marketing implies.

• Leading detectors on human-written text, low end of documented range12%
• Leading detectors on human-written text, high end of documented range26%
• Seven detectors on TOEFL essays by non-native English speakers61.3%

Leading detectors incorrectly flag human-written text as AI between 12% and 26% of the time. A widely cited evaluation found seven detectors showed 61.3% false-positive rates on TOEFL essays by non-native English speakers, against near-zero on native English writing. LinkedIn's membership is overwhelmingly non-native English. So the instrument most likely to accuse you carries a documented bias against the way most of the platform writes, and the accusation costs the accuser nothing.

Per the International Journal for Educational Integrity (Springer, 2026) and the Jisc National Centre for AI, leading AI detectors incorrectly flag human-written text as AI between 12% and 26% of the time, and seven detectors showed 61.3% false-positive rates on TOEFL essays by non-native English speakers against near-zero on native English writing.

A smoke alarm that goes off every time somebody makes a tadka is not a safety system. It is a nuisance that everyone learns to ignore, which is the worst possible outcome, including for the people who genuinely need a smoke alarm.

Which brings me to the most-shared statistic in my industry, that 81% of long-form LinkedIn posts are AI-generated, published by Originality.ai in July 2026. Three things are true about it at once. The publisher sells AI detection, so the finding doubles as its own advertisement. The sample is 5,000 search-discoverable posts drawn from ten pages of results across 90 topic-and-date searches, which is explicitly not a feed sample and systematically over-represents content shaped for search while under-representing ordinary posting. And it was scored by the publisher's own detector at a 15% AI allowance threshold. The study page itself discloses that significant AI editing on the post was used.

I am not claiming the true figure is lower. I am saying nobody has measured it in a way that would survive scrutiny, and no detector has yet been independently audited on the kind of multilingual professional text LinkedIn actually contains. Until that audit exists, a headline of this shape is measuring the detector at least as much as it is measuring the writing.

If the law does not make you disclose, should you disclose anyway?

This is the better question, and the evidence points somewhere uncomfortable for anyone who likes tidy ethics. Peer-reviewed evidence that AI disclosure reduces credibility is now consistent across domains. A within-subjects experiment with 433 participants, published in the Journal of Science Communication in 2026, found what its authors call a truth-falsity crossover effect: AI disclosure significantly reduced the perceived credibility of correct information while increasing the perceived credibility of misinformation.

Sit with the second half of that sentence, because it is the part that gets skipped. The label is not helping readers sort true from false. In that experiment it did something closer to the opposite, which means a disclosure badge is not a neutral act of honesty. It is an intervention with a measured effect, and the effect ran the wrong way.

Toff and Simon, writing in Political Communication in 2025, found that audiences perceive news labelled as AI-generated as less trustworthy even when the articles themselves are not evaluated as any less accurate or unfair. A 2025 paper in the Journal of Interactive Advertising found that AI disclosures raise persuasion knowledge and decrease trust toward both the advertisement and the organisation. The phrase for the pattern, the transparency penalty, comes from Nakano and colleagues' paper on reader perception shifts upon disclosure of AI authorship, posted as arXiv:2510.24011 on 23 January 2026.

	Field	What was studied	Reported effect of AI disclosure
Journal of Science Communication, 2026	Science communication	Within-subjects experiment, 433 participants	Reduced perceived credibility of correct information, increased it for misinformation
Toff & Simon, 2025	News	Audience trust in AI-labelled news articles	Less trusted, even when articles were not evaluated as less accurate or unfair
Journal of Interactive Advertising, 2025	Advertising	Trust in the advertisement and the organisation	Raised persuasion knowledge, decreased trust toward both
Nakano et al., 2026	Reader perception	Perception shifts on disclosure of AI authorship	Source of the 'transparency penalty' framing

Human ghostwriting sits in a completely different place, and the profession has actually been asked. When the Institute for Public Relations put the question to 291 PR professionals, most were comfortable leaving organisational ghost blogging undisclosed, on one condition: the thinking has to start with the executive, and the executive has to sign it off. A sizeable minority still objected. So the same industry that shrugs at a ghostwriter recoils at a model, which looks like hypocrisy until you notice what the condition is actually protecting. Nobody is policing who typed the words. They are policing whose judgement is underneath them.

Now the caveat most people writing about the transparency penalty leave out. Every study above sits in news, science communication or advertising. Nobody has yet tested whether disclosing AI assistance on a B2B thought leadership post reduces buyer trust. That question is genuinely open, and what would settle it is a controlled experiment on actual B2B buyers with the disclosure as the only variable. Until somebody runs it, anybody telling you what disclosure does to a commercial audience is reasoning by analogy from news readers, and should say so.

What I would do between now and 2 December 2026

1. Read the Commission's Article 50 FAQ yourself. It is free, it is short, and it is the primary source that everybody else is paraphrasing at you, usually badly.
2. Write down on one page how a piece gets made: who supplies the substance, who examines it, and what that person knows about the subject. That page is your answer if anyone asks, and it maps to the human review carve-out far better than any badge could.
3. Stop quoting the 47% reach penalty and the 81% AI figure. Neither survives its own sourcing, and you cannot argue that you are trustworthy using evidence that is not.
4. If you build or operate a system that generates published text, put 2 December 2026 in the calendar, because that is when the limited grace period on the marking and detection obligation for systems already on the market ends.
5. Watch LinkedIn's private dashboard flag more closely than you watch Brussels. If it ships broadly, it changes how this work gets delivered in a way the regulation never will.

Here is the test I would run on any compliance product you are offered this quarter. Ask the seller to point at the sentence in the Commission's FAQ that applies to your situation. Not the article number, not the fine, the sentence. If they can point at it, you have learned something useful and you should probably listen to the rest of what they say. If they cannot, you have learned something more useful, and it cost you nothing.

Then there is the question I cannot answer for you, which is worth more than any checklist. If a reader cannot tell the difference between something you wrote and something you approved, and no detector can reliably tell either, the only thing left to separate them is whether the thinking was yours. That is not a legal test and no regulator will ever run it. It is the one that has always decided whether anybody comes back for the second piece, and 2 August 2026 changed nothing about it.

Questions I get asked about this

Do I have to label AI-assisted LinkedIn posts from 2 August 2026?

Not on the reading of the European Commission's own Article 50 FAQ. The text-labelling duty in Article 50(4) covers text published to inform the public on matters of public interest, named as politics and democratic processes, public administration and services, administration of justice and law enforcement. Commercial and professional content in general is not on that list, and text that has undergone human review or editorial control does not need to be labelled.

What is the human review exemption, exactly?

The Commission's Article 50 FAQ states that text which has undergone human review or editorial control does not need to be labelled, and describes human review as deliberate examination of the substance by a person with relevant knowledge. That is a description of somebody who understands the subject reading the work and deciding whether it holds, not a checkbox on a form.

What happens on 2 December 2026?

A limited grace period ends on that date for the marking and detection obligation on systems already on the market. The Article 50 transparency obligations themselves become applicable on 2 August 2026.

How large are the fines under Article 50?

Non-compliance can attract fines up to EUR 15 million or 3% of worldwide annual turnover, per the European Commission. The prior question, and the one most compliance marketing skips, is whether an obligation attaches to you at all.

Does LinkedIn require AI disclosure on posts?

LinkedIn's policy post 'Keeping conversations real on LinkedIn' says it is 'ok to use AI to help you write' provided posts represent your own voice and perspectives. On 30 July 2026 LinkedIn shipped a 'Seems like AI slop' report button on posts and comments, said it will begin privately flagging in users' own dashboards when people believe their content is coming off as inauthentic due to heavy AI use, and confirmed it is discontinuing its AI 'enhance your post' feature.

Sources

1. Transparency obligations under Article 50 of the AI Act (FAQ), European Commission, n.d..
2. Guidelines on transparency of AI-generated content, European Commission, n.d..
3. Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems, EU Artificial Intelligence Act (text of Regulation (EU) 2024/1689), n.d..
4. EU AI Act Transparency Obligations: Preparing for Compliance by 2 August 2026, Sidley Data Matters, 24 June 2026.
5. LinkedIn adds a button to report AI-generated slop, TechCrunch, 30 July 2026.
6. LinkedIn adds a seems like AI slop button after blocking billions of automated comment attempts, Fortune, 31 July 2026.
7. Engineering the next generation of LinkedIn's Feed, LinkedIn Engineering Blog (Hristo Danchev), 12 March 2026.
8. Visible sources and invisible risks: exploring the impact of AI disclosure on perceived credibility of AI-generated content (JCOM_2501_2026_A09), Journal of Science Communication, 2026.
9. Or They Could Just Not Use It?: The Dilemma of AI Disclosure for Audience Trust in News, Toff & Simon, Political Communication (Sage), 2025.
10. Disclaimer! This Content Is AI-Generated: How AI-Disclosures Influence Trust in Advertisements and Organizations, Journal of Interactive Advertising, 2025.
11. Understanding Reader Perception Shifts upon Disclosure of AI Authorship (arXiv:2510.24011), Nakano, Takezawa, Matulic, Yang & Yatani, 23 January 2026.
12. Is Ghost Blogging Like Speechwriting? A Survey of Practitioners About the Ethics of Ghost Blogging, Institute for Public Relations, n.d..
13. Evaluating the accuracy and reliability of AI content detectors in academic contexts, International Journal for Educational Integrity (Springer), 2026.
14. AI Detection and assessment: an update for 2025, Jisc National Centre for AI, 24 June 2025.
15. LinkedIn AI Content Study: 81% of Long-Form Posts Are Likely AI, Originality.ai, July 2026.

Every figure above is linked to its original source. Where I am predicting rather than reporting, I say so in the panels marked My take.