Which tool to choose
Automatic detectors exist, but they have to be understood before being used.
- Online AI text detectors: they give you a percentage of "AI probability." Handy for a first suspicion, but fragile: a rewrite, a translation, or a good manual revision wipes out their reliability. Never use them as proof.
- The AI itself as an analyst: you paste the text and ask for a style analysis. Useful for reasoning about the clues, but no more reliable than the detectors on the final verdict.
- Your trained eye: for the texts you know (your students' assignments, a colleague's emails) the comparison with the person's habitual way of writing is the most solid signal you have.
How to do it
Look for the stylistic clues, then weigh them together. A single signal isn't enough.
- Assess the "suspicious smoothness." Is the text grammatically perfect, well-structured, without a slip, but strangely flat? AI writing tends to be correct and impersonal, devoid of the stumbles and idiosyncratic choices of a real person.
- Look for the absence of concrete experience. Are the examples generic and interchangeable ("for example, in many companies...") instead of specific and lived ("on Tuesdays our warehouse worker...")? The AI speaks in categories, direct experience speaks in details.
- Note the recurring tics. Sentences opening with "It's important to note that," perfectly balanced three-item lists, the "not only... but also" construction, closings with a moral summary: these are frequent stylistic fingerprints in generated texts.
- Compare with a reference text, when you have one. If you have a text the person wrote in the past, set them side by side: the sudden jump in register, vocabulary, or sentence length says more than any detector.
Feedback: if you accumulate three or four concordant clues (smoothness, generic examples, recurring tics, a register different from the usual), the suspicion is well-founded. If you find only one, it's not enough: a human also writes in an orderly way.
A concrete example
A professor receives a surprisingly fluent essay from a student who usually writes with difficulty. He doesn't trust the first online detector he uses, and rightly so: it gives 95% AI probability, but that same tool is known to fail human texts too. So he reasons about the clues. The essay is perfect but cites nothing from the syllabus covered in class, the examples are generic, and it completely lacks the clumsy but sincere voice of the previous assignments. Instead of accusing on the basis of a percentage, he summons the student and asks him to explain aloud two passages of his own text. There the truth emerges, with no need for software.
When it does NOT work (and how to fix it)
If the detector gives a clear-cut verdict
A detector that declares "100% AI" or "100% human" is giving a confidence it doesn't possess. These tools have real error rates and have been caught failing texts written by people, including classic works. The remedy is to never treat the percentage as proof: use it as a cue to verify with the stylistic clues and, where possible, with a direct comparison.
If the text was rewritten by hand
A text generated by the AI and then reworked by a person loses almost all the detectable traces. Here no tool helps. The only sensible approach, in a context like school, is to shift the verification onto the process: asking for intermediate drafts, an oral explanation, a link with what was done in class.
If you have to make a decision that has consequences
When a grade, an evaluation, or an accusation depends on your judgment, don't rely on any software as proof. The remedy is procedural: gather several converging clues, give the person the chance to explain themselves, and base the decision on the comparison with their habitual work, not on a percentage.
A tip from someone who actually uses it
If you manage assignments or other people's texts, shift the attention from the product to the process. Asking for a draft, a note, an oral explanation of a chosen passage reveals in a few minutes what no detector can guarantee. Whoever actually wrote a text knows how to tell you how they got there; whoever just pasted it freezes.
Frequently asked questions
Is there an invisible watermark that reveals AI texts?
There's been long talk of statistical "watermarks" hidden in the text, but in practice there's no official, tamper-proof mark. Those signals, where they exist, are fragile: a paraphrase, a translation, or a revision erases them. Don't count on it.
If a text is well written, does it mean the AI made it?
No, and it's the misunderstanding that does the most damage. Writing well isn't a fault nor a proof: many people write clearly and correctly. Accusing someone just because the text is fluent is unfair and wrong. The suspicion arises from the set of clues, never from quality alone.