Honestly, no, and neither can we. On short passages in an unfamiliar register, people perform close to chance. The EU AI Act itself concedes the point: Article 50 requires providers to mark synthetic content in a machine-readable format, which is a law written on the assumption that readers will not notice. What replaces noticing is verifying, and the four habits at the bottom of this page are what that looks like.
Below are eight passages. Five are genuine text from Regulation (EU) 2024/1689 and three were written by an AI for this exercise. Every reveal names which, and every genuine one links to the article so you can confirm we are not simply asserting it.
You are not expected to win. If you do well, notice the conditions: one subject, eight samples, and you were told in advance that half were fake. None of that is true of a message that arrives on a Tuesday asking you to move money.
Em dashes and polished punctuation
Professional editors have always used them, and every model can be told not to. Punctuation identifies a house style, never an author.
Certain favourite words
Word-frequency tells are the fastest thing to defeat, because the moment a list of them circulates, the next prompt says avoid these words. Any tell that can be published stops working once it is published.
It reads too smoothly
Smoothness measures effort, not origin. A human who revised six times and a model on its first attempt land in the same place, and plenty of AI text is now deliberately roughened.
Hands, ears and garbled text in images
A generation of image models ago this was reliable. It has not been for some time, and it is the single most confidently repeated piece of out-of-date advice in this subject.
An AI detector said so
Detectors return a probability, not a fact, and they are measurably worse on writing by non-native English speakers. People have been accused of cheating on that output. Treat a detector score as an opinion with a known bias, and never as evidence about a person.
1. Resolve the citation
Every claim of fact carries a source or it does not. Open the source. Confirm it exists, that it says what the passage says, and that it is current. Four of the eight passages above are real and each one links to its article, which is the whole difference between a quotation and an assertion.
2. Verify the person out of band
If a message, a voice or a video is asking you to do something, contact the person through a channel you already had, never the one that reached you. A number in the message proves nothing about who sent it.
3. Ask what the sender gains
Synthetic content is a delivery mechanism, not a goal. Urgency, secrecy and a payment or a password at the end are the constant across every version of this, and they were the constant long before any of it was generated.
4. Judge the claim, not the prose
Style tells you who wrote something. It never tells you whether the thing is true. A human can be wrong fluently and a model can be right clumsily, so the writing is the least informative part of the page.
Spoiled deliberately, for anyone reading rather than playing.
Passage 1
Genuine lawHigh-risk AI systems shall be designed and developed in such a way, including with appropriate human-machine interface tools, that they can be effectively overseen by natural persons during the period in which they are in use.
Genuine. This is Article 14(1) of the EU AI Act, on human oversight, word for word.
Real legislative drafting is often clumsier than an imitation of it. The doubled phrase "designed and developed" and the bolted-on clause about interface tools are the marks of a text negotiated by committee, not of a machine. Awkwardness is not evidence of a human, but polish is not evidence of a machine either.
Read Article 14 in full
Passage 2
Written by an AIDeployers shall ensure that any natural person tasked with oversight of a high-risk AI system possesses the competence, authority and support necessary to interrupt the operation of that system where its behaviour deviates from expected parameters.
Written by an AI for this exercise. There is no such provision in that form.
This is the dangerous kind of fake, because it is nearly right. Article 26(6) does require deployers to assign oversight to people with the necessary competence, training, authority and support. What was invented is the clean trigger, "deviates from expected parameters", which reads like law and is the sort of crisp condition real regulation avoids because it would have to be defined. A citation that half fits is harder to catch than one that does not fit at all.
Passage 3
Genuine lawThe risk management system shall be understood as a continuous iterative process planned and run throughout the entire lifecycle of a high-risk AI system, requiring regular systematic review and updating.
Genuine. Article 9(2) of the EU AI Act, on the risk management system.
Nothing on the surface separates this from the invented passage above it. Same length, same register, same confidence. If you got both right you were lucky rather than skilled, and the honest read of that is that surface inspection does not scale.
Read Article 9 in full
Passage 4
Genuine lawProviders of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.
Genuine. Article 50(2), the marking duty for synthetic content, which is the law that exists precisely because this exercise is hard.
The reason the EU legislated machine-readable marking is that human detection does not work. The law itself concedes the point this page is making: if people could reliably tell, there would be no need to require a label.
Read Article 50 in full
Passage 5
Written by an AIWhere an AI system is intended to interact directly with natural persons, providers shall design the system so that affected individuals are informed of its automated nature in clear and intelligible language, unless this is obvious from the circumstances of use.
Written by an AI for this exercise, although Article 50(1) does impose a duty of this general shape.
Notice how plausible fluency did the work here. It scans as law because it borrows the cadence of law. Cadence is the cheapest thing to copy and the first thing people mistake for authenticity.
Passage 6
Genuine lawThis obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.
Genuine, and recent. This sentence entered Article 4 through the Digital Omnibus, Regulation (EU) 2026/1744, in force 27 July 2026.
The hardest passages to judge are the ones that changed recently, because your instinct is calibrated on the old version. Anyone relying on memory of Article 4 from 2025 would call this a fake. Recency defeats familiarity, and it is the reason a citation with a date beats a feeling every time.
Read Article 4 in full
Passage 7
Written by an AIMember States shall ensure that national competent authorities publish, on an annual basis, an anonymised register of serious incidents reported under this Article, together with a summary of the corrective measures taken.
Written by an AI for this exercise. Serious incident reporting is real, in Article 73. A public annual register of them is not.
This is the invention worth fearing: a real obligation extended by one reasonable-sounding step. It is exactly what a confident summary of a law will do to you, and exactly why the summary is not the law. The step from what a rule says to what it ought to say is invisible in fluent prose.
Passage 8
Genuine lawHigh-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system.
Genuine. Article 12(1), record-keeping.
Short, plain and unremarkable, which is what makes it easy to dismiss as filler someone generated. Real law is frequently boring. Boring is not a tell.
Read Article 12 in full
Can you tell if text was written by AI?
Not reliably, and the research and your own score on this page will both say so. People perform close to chance on short passages in an unfamiliar register, which is why the EU AI Act requires providers to mark synthetic content in a machine-readable format rather than expecting readers to notice. What works is not inspection but verification: resolve the citation, contact the person through a channel you already had, and judge the claim rather than the prose.
Do AI detectors work?
They return a probability, not a fact, and they are measurably less accurate on writing by non-native English speakers, which has led to students and employees being wrongly accused. A detector score is an opinion with a known bias. It should never be treated as evidence about a person, and no consequence should rest on it alone.
Is the em dash a sign of AI writing?
No. Editors and professional writers have used em dashes for centuries, and any model can be instructed not to. The deeper problem is that every published tell defeats itself: the moment a list of giveaway words or punctuation circulates, prompts start avoiding them. Any detection rule that can be written down has a short life.
How do I know whether a quote from a law is real?
Open the provision. A real quotation names the article and the instrument, and the text at that citation matches. An invented one is usually a real obligation extended by one reasonable-sounding step, which is the hardest kind to catch by reading. Our EU AI Act Explorer serves every article and recital in full so a citation can be resolved in a few seconds.
What should I do if I think a message is AI generated?
Treat the question as beside the point and verify the request instead. Contact the supposed sender on a number or address you already had, never one supplied by the message. Refuse to act on urgency. If the message concerns money, credentials or intimate images, slow down, because those three are what almost every synthetic-media attack is ultimately for.
Nothing on this page is stored or sent anywhere, and there is no account. The genuine passages are short quotations from Regulation (EU) 2024/1689, each linked to its full text. The rest was written for this exercise and is labelled as such wherever it appears.