How to tell if something is written by AI

You can tell that a text resembles what a language model produces by default. You cannot tell, from the text alone, who or what produced it. Keep both halves of that sentence and the method below is useful; drop the second half and it becomes the machine for false accusations that the detectors already are. Here are the seven patterns, how to check each one, what else produces it, and what evidence actually settles the question.

Why generated text has a style at all

A language model writes by choosing, at each step, one of the words it finds most likely. Unless instructed otherwise, it drifts to the centre of everything it has read: average sentence length, average paragraph, the most common way to open a point, the safest adjective. The result is prose with no edges. Stanford researchers measured this as low perplexity, the model's own lack of surprise at its text, and showed that it is also the signature of plain, conventional human writing, which is why their seven detectors flagged 61 percent of essays by non-native English speakers. The patterns below are the visible side of low perplexity. Learn to see them and you will see them in generated text, in bureaucratic prose, in your own first drafts, and in the essays of people writing in a second language.

The seven patterns, and how to check each one

The detector on this site measures the first six of these and shows the measurement; the seventh needs a reader. No tool can tell you whether the source in footnote 4 exists. You can, in two minutes, and it is the most reliable check on this page.

A worked example

Here is a paragraph of the kind the patterns describe: "In today's fast-paced digital landscape, it is important to note that effective communication plays a crucial role in fostering collaboration. Furthermore, organizations must leverage innovative strategies to navigate the complexities of an ever-evolving market. Additionally, teams that embrace a holistic approach are better positioned to unlock their full potential." Three sentences of 19, 16 and 17 words. Two of the three open with a transition. Eleven stock expressions in 52 words. Not one noun that refers to a particular organisation, market, team or strategy. Not one contraction, person or opinion. Paste the full sample on the detector's home page (the "Try a sample" button loads a three-paragraph version of it) and five of the seven meters are flagged: rhythm, stock phrases, paragraph uniformity, openers and register.

Now here is what that proves: that the paragraph is written in the default style of a language model. It does not prove a model wrote it. A consultant who has absorbed twenty years of corporate decks writes this way unaided, and a model asked to "write like a tired blogger with opinions" does not. The patterns describe a register, and the register has human speakers.

What the patterns miss

Everything that was edited. The Stanford study gave ChatGPT a one-line follow-up, "elevate the provided text by employing literary language", and the seven detectors' catch rate on its essays dropped from 100 percent to 13 percent. A person who pastes generated text into a paraphrasing tool, or reads it once and roughens it, removes most of the seven patterns in minutes. TechCrunch tested the public detectors in February 2023 and found that most of them failed spectacularly on exactly this kind of text. So a text that shows none of the patterns is not thereby human; it is a text that shows none of the patterns.

The only method that does not depend on style is a watermark planted by the model at generation, such as Google DeepMind's SynthID, and it only answers for the models that carry it, only while the text is largely unedited, and only to whoever holds the key. For the text in front of you, from an unknown source, there is no watermark to read.

Short texts, long texts, mixed texts

The patterns need length. A paragraph has no rhythm to measure and a tweet has no vocabulary to be narrow; below a hundred words nothing on this page applies, and the serious detectors refuse such texts for the same reason. From three hundred words up, rhythm and vocabulary become visible; from a thousand, the stock-phrase count and the paragraph uniformity become hard to fake either way. The hardest case is the mixed text: a generated draft that a person reworked, or a human draft that a model polished. The patterns then appear in patches, a flat regular section between two lively ones, and the useful move is to read paragraph by paragraph rather than to score the whole. A sudden change of voice inside one piece is more telling than any average over it.

Style is not provenance

Everything above measures style, and style can be imitated, taught, edited and coached in both directions. The one method that measures provenance instead, a watermark planted by the model at generation, works by adjusting the probability of each word as the text is produced, in a pattern the model's owner can later detect; Google DeepMind describes its SynthID this way for text from the Gemini app. It answers "did this service produce this", not "did a person write this", and only for text that is largely unedited and from a model that marks it. For the ordinary case, a text of unknown origin in front of you, there is no provenance to read, and the honest conclusion of a style check is a sentence that begins "this reads like" and never one that begins "this was written by".

A checklist to keep

The evidence that actually settles it

If the question matters, the answer is outside the text. A document's version history shows a piece growing over hours or appearing whole. A conversation with the author shows whether they can explain a choice, extend an argument, defend a claim; a person who wrote six hundred words can talk about them for six minutes. Sources either exist and say what is claimed or they do not, and an invented reference is the one sign on this page that is close to decisive, because people rarely fabricate a citation by accident. Earlier work by the same writer shows what their voice sounds like. None of this is a score, and all of it is stronger than one.

How to use this without harming anyone

Use the patterns to decide where to look, and look. If you grade, ask for the draft and talk to the student before you conclude anything; the false positive rate of your own reading is unknown and the cost of a wrong accusation falls on them. If you edit, treat the patterns as notes on flat prose, whoever wrote it: shorter and longer sentences, a specific noun, a person in the room. If you are the writer and your own work trips the patterns, you have learned something about your style, not about your honesty. And if someone shows you a detector's badge as proof of anything, send them to the guide on what the published numbers actually say.