How to tell if something was written by AI

Start with what you can see, then measure what you can't — and keep the conclusion proportional to the evidence.

Someone hands you a text — an essay, an application, a report — and something feels off. It is fluent but strangely weightless. Before you reach for a detector, it helps to know what you are actually looking for. AI-generated writing is not bad writing; it is often smooth, grammatical, and well organised. The giveaway is not errors. It is a particular kind of sameness that comes from how language models work: they generate text by repeatedly choosing the most probable next word, and probability, repeated hundreds of times, leaves a texture.

The visible signs

Read the text once through and watch for a cluster of these traits. Any one of them appears in human writing too; it is the combination that raises the question.

• Even, metronomic rhythm. Sentences land in a narrow band of lengths, paragraph after paragraph, with none of the short punches or long spirals people naturally produce.
• Generic transitions and scaffolding. “Furthermore,” “Moreover,” “In conclusion,” “It is important to note” — the connective tissue of a text that is assembling points rather than thinking through them.
• Balance without commitment. Every side gets its paragraph, every claim is hedged, and nothing is ever quite asserted. The text sounds reasonable about everything and convinced of nothing.
• Confident specifics that dissolve on contact.References to studies that don't exist, statistics with no source, quotations nobody said. Invented citations are one of the most checkable traces a model leaves behind.
• A voice with no biography. No odd detail, no unnecessary aside, no phrase only this author would use. Human drafts carry the writer's history in them; generated text carries the average of everyone's.

The signs you can't see

The stronger evidence is statistical, and it is not visible to the eye. Two properties matter most. The first is predictability: given the words so far, how surprising is the next one? Model-generated text sits at the high-probability end of that scale almost everywhere, because that is literally how it was produced. The second is variation: human writing alternates between predictable stretches and sudden leaps, while generated writing holds a steadier level of predictability from start to finish. A checker measures both across the whole document and sentence by sentence, then weighs how consistent the pattern is. A single odd sentence proves nothing; three hundred sentences that are all equally “typical” is a pattern no human habit explains as well.

A practical order to check in

First, read. If the surface signs are absent and the text carries specific, verifiable detail in a distinctive voice, you may already have your answer. Second, verify the facts. Check the citations, the names, the dates. Fabricated references are strong evidence regardless of what any score says, and real, accurate, obscure detail is evidence in the author's favour. Third, run a checker — the one below, or the full version on HalluSpot — and read the result as a band, not a verdict. Fourth, ask the author. How a person talks about “their” text — the choices they made, the parts they struggled with, the drafts that came before — is often the most informative test of all, and the only one that treats them fairly.

Where this goes wrong

Two failure modes matter. Formal, formulaic human writing — lab reports, legal summaries, application essays written to a template — can look machine-like because the genre flattens exactly the variation detectors measure. And edited AI text moves the other way: a human who rewrites a generated draft in their own words removes the statistical traces along with the phrasing. That is why no single step above is decisive, and why the honest output of the whole process is a degree of confidence, not a conviction. If the stakes are someone's grade, job, or reputation, the evidence has to be gathered like it matters — because it does.