How to tell whether a piece of text was written by AI

You cannot prove that a piece of text was written by AI. Not from the words on the page, not with a tool, not with a checklist. What you can do is form a judgment about how likely it is, and that judgment depends far more on context than on style: who the writer is, whether they can show you how the piece came together, and whether the claims inside it survive a check. The stylistic tells that circulate online are genuine patterns, but they are weak signals, and people have lost grades, contracts and jobs because someone treated a weak signal as a verdict.

The tells people actually notice

These are the patterns that make a reader's eyebrow go up. They are worth knowing, as long as you know what they are worth.

• Even paragraph shape. Three to five sentences per paragraph, every paragraph, with no one line paragraph anywhere and no paragraph that runs long because the writer got interested.
• Balance on every point. Each claim is immediately softened by its counterweight: "while this offers clear benefits, it is also important to consider the challenges." Nobody ever just says the thing.
• Examples with the specifics filed off. "A small business owner" rather than a man who runs a two chair barber shop and finally took card payments in his fourth year.
• A recognizable connective vocabulary. Moreover, furthermore, additionally, it is important to note, in today's landscape, navigating the complexities, delve into.
• Things arriving in threes. Three benefits, three risks, three tips, in a sentence rhythm that stays the same throughout.
• A suspiciously clean surface. No typos, no inconsistent spelling of a name, no stray double space, no formatting that drifts halfway down.
• No lived detail. Nothing that could only be known by someone who was there: a temperature, a wrong turn, the thing the colleague said, the date when it all went wrong.
• A closing paragraph that restates the piece without adding anything to it.

Why that list also describes a lot of human writing

Here is the problem. Read those bullets again and think about a press release, a university essay written to a marking rubric, a corporate policy update, a grant application, or the "about us" page of a mid sized company. Balanced hedging, even paragraphs, generic examples and formal connectives are what people produce when they are writing to be safe rather than to be interesting. That style predates chatbots by decades. It is taught in schools and enforced by editors.

Two further complications. First, a great deal of text is now a blend: a human outline, an AI draft, a human rewrite, or a human draft cleaned up by a tool. There is no clean line to detect. Second, the tells drift. Every observation that "AI always writes like this" changes the way people prompt and the way models are tuned, so any list of giveaways is a snapshot of a moving target rather than a permanent test.

Who gets wrongly accused

The cost of this guesswork is not spread evenly. People writing in English as a second or third language tend to use a smaller, more common vocabulary and more predictable sentence patterns, which is precisely the profile these judgments punish. So do students taught a rigid essay structure, writers who use grammar tools that smooth their sentences, and anyone whose natural register is plain and orderly. Meanwhile a confident native speaker who writes with jokes and digressions sails through, whether or not a machine helped.

If you are in a position to make an accusation, that asymmetry should sit uncomfortably with you. The full picture of how detection tools fail, and what to do if one has flagged your work, is covered in why AI text detectors get it wrong.

What counts as real evidence, and what does not

The bottom three rows are worth more than the top three combined. Fabricated references in particular are a strong signal, because an author who invents a source usually did not read one. The mechanics of that failure are explained in why AI invents citations.

A better procedure than guessing

1. Pick three checkable claims from the text: a statistic, a named source, a date.
2. Check them. Does the cited paper exist, and does it say what the text says it says? Advice on doing this quickly is in how to fact check an AI answer.
3. Open the file's revision history. In Google Docs it is under File, then Version history. In Word, look at the document's stored versions or the tracked changes. Note what it shows, and note that its absence proves nothing.
4. Ask the author one open question about a decision inside the piece: why they cut a section, why they picked that example, what they left out.
5. Ask what they would write next and why.
6. If you are still concerned, describe what you observed rather than announcing a conclusion. "Three of the sources here do not appear to exist, can you send them to me" is a fair question. "This is AI" is an accusation you cannot support.

Decide whether authorship even matters

Before you spend an afternoon on this, be clear about why you care. For a school assignment, authorship is the point of the exercise, so it matters. For a legal statement, a medical note, a work sample in a hiring process or a piece of journalism, it matters, because the reader is relying on a human having taken responsibility. For a product description, a meeting summary or an internal briefing, what matters is whether it is accurate and useful.

Start with the accuracy question, since it is the one you can actually answer. When you are evaluating something you found online rather than something a person handed you, the wider checks in how to spot a reliable source will get you further than any tell in this article.