AI-Written Text: When Do You Have to Disclose It?

Three separate questions that keep getting mixed up

When someone asks "do I have to say I wrote this with AI", they are really asking three different things at once. Is it legally required? Does the surface I am posting to require it? And is it fair not to say so? The three answers differ, and none of them follows from the others.

The most common error runs in both directions: people either think the legal duty is far wider than it is, or assume that if there is no legal duty there is no rule at all. In reality the law draws a very narrow line while platforms and professional bodies draw much wider ones — and for most people the second is what actually applies. (This post is an explainer, not legal advice; for a specific case, ask a lawyer.)

The law: what the EU says

The EU AI Act's labelling rule for text has been applicable since 2 August 2026, and it places the duty on the user — the "deployer", in the regulation's own term. The wording, however, ties three conditions together, and all three have to hold.

First: the text has to be published. Second: the purpose of publication has to be informing the public. Third: the subject has to be a matter of public interest. The European Commission's guidance lists what belongs there: politics and democratic processes, public administration and public services, the administration of justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, scientific and cultural developments relevant to public debate.

And there is an exemption that, in practice, carries most cases. The duty does not apply where the content has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for its publication. The Commission's guidance defines both halves: human review is "deliberate examination of the substance of the content by one or more natural persons possessing relevant knowledge", and editorial control is a responsible editorial practice holding the authority to approve, alter or reject.

One important narrowing belongs here too: the guidance states explicitly that superficial, purely formal or procedural checks — spell-checking or grammatical correction, for instance — do not count as human review. A routine, pro-forma nod is not enough. But where someone genuinely read it, checked it and answers for it, the labelling duty does not arise.

What it therefore does not cover

Five common cases the labelling duty does not reach, given the conditions above. This is not a loophole; it is the scope of the rule.

• The work email. It is not published, so the first condition already fails.
• The internal document. Likewise: a report, minutes, a deck for the team.
• Marketing material and advertising. Marketing and advertising do not appear among the public-interest subjects in the Commission's list.
• A company or personal blog, insofar as it does not inform on a matter of public interest. A product announcement is a different category from a health or regulatory explainer.
• Edited journalistic content with a responsible editor. This is precisely the exemption the regulation names — but only where the review is real.

Machine-readable marking is the provider's job, not yours

There is a second duty in the same article that gets routinely conflated with the first: machine-readable marking. That one does not apply to you but to the provider of the system — the company operating the model. They have to ensure the output is detectable by machine as artificially generated.

Two further details belong here. One is the exemption where the system "performs an assistive function for standard editing" or does "not substantially alter" the input data or its semantics — so a grammar corrector falls outside. The other: providers of systems already on the market before 2 August 2026 were given until 2 December 2026 to implement it. The voluntary code of practice on this was finalised in June 2026 and had around 190 signatories by the end of the summer; it treats free-form text separately, because — as it puts it — text cannot carry metadata, and it exempts very short text, under roughly 150 words, from watermarking altogether. For images the same duty looks quite different, because a picture can carry both metadata and a hidden mark — watermarks and provenance goes through that in detail.

Elsewhere in the world

Labelling is not globally uniform, and the differences are substantive.

• China. The measures on labelling synthetic content have been in force since 1 September 2025 and, alone among the major jurisdictions, require visible labelling of text too — a notice placed at the beginning, end or middle of the text — while also requiring users to label content themselves when they publish it.
• United States. California's transparency act became operative on 2 August 2026, but it applies specifically to image, video and audio content. Text is not within its scope.
• United Kingdom. As a January 2026 parliamentary briefing puts it, there is no UK legislation requiring AI-generated content to be labelled.

Where the rules are stricter than the law

In practice most people are bound not by legislation but by the surface they are posting to. Here the rules are wider, and they are about declaration rather than prohibition.

• Academic publishing. The governing bodies are unanimous: AI cannot be an author, because it cannot take responsibility, while its use must be disclosed. The leading medical editors' body opened a dedicated section on the subject in January 2026 and states that non-disclosure may in some circumstances be construed as misconduct. Publishers differ, though: some would have every instance of language polishing declared, while others explicitly permit language smoothing without one. All the major publishers exempt basic spell-checking.
• Newsrooms. An analysis of the policies of 52 news organisations found roughly 90 percent requiring disclosure where AI went into a piece. One large public service broadcaster puts a visible label on such content, and from March 2026 a major British daily has allowed generative AI for direct publication only in exceptional cases, with senior editorial permission.
• Book publishing platforms. The largest self-publishing platform requires a declaration for AI-generated text and distinguishes "AI-generated" from "AI-assisted": if the machine created the content it counts as generated even after extensive later editing, whereas if you wrote it and the machine only improved it, no declaration is needed. One important detail: that declaration is internal, and readers never see it.
• Freelance platforms. Here there is typically no proactive duty: the rule is that if the client asks, or has stated up front that they do not want AI, you must say so. The trigger is the client, not the tool.

What is conspicuously absent: none of the major social platforms labels AI-written text. Their labelling schemes cover images and video. Anyone waiting for the platform to handle this on their behalf will be waiting a long time.

What the research says about trust

Here comes the part the legal debate usually obscures. Disclosure has a price, and it has been measured. A series of sixteen experiments with 27,491 participants found that readers consistently rate the same text lower when they believe a machine wrote it — and the mechanism is not perceived quality but perceived authenticity. Another study, across thirteen experiments, measured a 20 percent drop in trust among clients and 16 percent among students.

That this is a label effect and not a quality difference is shown most cleanly by a third experiment: when machine text was presented as human-written, evaluators preferred it 68.4 percent of the time, against 51.2 percent in blind comparison. Same text, different label, a seventeen-point swing.

Two findings point the other way, though, and these two are the practically useful ones. First: being exposed is worse than disclosing. In that same research, trust fell further when someone else found out than when the person said so themselves. Second: in an experiment on news texts, the trust penalty from disclosure was largely cancelled out by listing the sources. So the problem is not the label but the absence of verifiability — and that can be supplied.

A decision rule you can use

Four questions, in this order. Most texts drop out at the first.

1. Am I publishing this, to the public, on a matter of public interest? If any of those is no, the EU labelling duty does not touch it. Go to the third question.
2. If yes: has a knowledgeable person read and approved it, and do they answer for it? If so, the regulation's exemption applies. If the "approval" was a glance, it does not — and then you either label it or genuinely review it.
3. Does the surface I am posting to require a declaration? An academic journal, a book platform, a client, an employer's policy. This is the question that affects most people most often.
4. Could I defend it if someone asked tomorrow? They will not ask about the label; they will ask whether the claims in the text are yours and whether you checked them. Checking before you publish sets out a method for that, and this is the question that matters in practice.

What to take away

The EU labelling duty is tied to three simultaneous conditions — published text, informing the public, on a matter of public interest — and does not arise where there is genuine human review and editorial responsibility behind it. Machine-readable marking is the system provider's job, not yours. In practice, though, what will find you is not the law but the policy of a journal, a platform or a client — go and read those. And if you do disclose, put your sources beside it: by measurement, that cancels most of the trust penalty. Being exposed is always worse than disclosing.