How to tell if text was written by AI
The short answer
There is no reliable way to look at a paragraph and know a model wrote it. Detectors are unreliable, watermarks are easy to strip, and the tells people rely on are style habits, not evidence. What you can do is judge the content: check the specific claims, look for a person or process behind them, and use labels and metadata as clues rather than proof.
Why "it sounds like AI" is not a test
Most of the signals people learn, like em dashes, tidy three-part lists, or the phrase "it is important to note", appear in plenty of human writing too. They are artefacts of a particular style of editing, and style is copyable. Judging authorship from tone gives you a guess about how something was written, not who wrote it.
Detector tools are a different problem. They produce a probability score from patterns in text, and those scores move when text is edited, translated or lightly rewritten. A high score is not proof and a low score is not clearance. Any tool that reports a single confident percentage is overselling what the method can do.
The tells that actually survive
- Specifics that check out. Named sources, dates, version numbers, direct quotes with a link. AI can produce these, but it produces them wrong more often. Verifying three specific claims tells you more than any stylistic read.
- Claims that cannot be checked at all. Vague authority ("studies show", "experts agree") with no name attached is the most common pattern in low-effort generated copy.
- Uniformity at length. Long text with no digressions, no corrections, no oddly specific personal detail, and the same paragraph length throughout reads as generated more often than not.
- Errors that a human would catch. Citations to papers that do not exist, quotes attributed to the wrong person, products with features that were never shipped. This is the strongest practical signal.
- Metadata and provenance. C2PA content credentials, platform AI labels, and image metadata can confirm generation. Their absence proves nothing, since most tools strip them.
How to check a text in five minutes
Start with the most checkable claim and verify it. If the piece cites a report, find the report. If it quotes someone, find the quote. Two or three checks usually settle whether the writer had a real source, because fabrication tends to cluster.
Then look at the byline and the date. A named author with a history, a publication date, and a corrections policy is a stronger signal than any stylistic marker. Where there is no author at all, treat the content as unverified until the claims hold up.
Finally, check whether the text acknowledges uncertainty. Human writers hedge about things they do not know. Generated text tends to be confident everywhere, including about things that are wrong.
What AI labels do and do not mean
A label tells you that a platform detected generation, or that a creator disclosed it. Both are useful and both are incomplete. Platforms label unevenly, and creators can fail to disclose. In the EU, the AI Act sets transparency duties for certain AI-generated content, and Australia is under pressure to require labelling, with 84% of respondents in a recent Deloitte survey saying they want it.
The practical rule: a label is evidence for one direction only. If a label says AI, believe it. If there is no label, do not conclude a human wrote it.
If you are the one publishing
Say what the machine did and what the person did. "Drafted with AI, checked and edited by a human" is a sentence readers can act on, and it costs nothing to write. Keep the sources visible, keep a named human accountable, and do not let a tool invent a citation you did not open. That is the whole discipline, and it is more useful than any detector.