Seven signs text was written by AI

These are countable. You can run them past a tool and get a number rather than an impression, which is what makes them more useful than vibes-based detection.

1. Em dashes at a steady rate. Chat models produce the em dash (—) far more consistently than people do, because typing one requires a keyboard shortcut most writers never learned. One or two in a long document is normal. Roughly one per paragraph, evenly spread, is not. The em dash guide covers why the habit is so entrenched.

2. Curly smart quotes and typographic punctuation. AI output tends to arrive with directional quotation marks (" "), curly apostrophes ('), and a real ellipsis character (…) rather than three full stops. Word processors also produce these automatically, so they are only meaningful in contexts where autocorrect was never involved — a plain-text field, a code comment, a CSV file.

3. Invisible Unicode characters. Zero-width spaces, non-breaking spaces, and narrow no-break spaces frequently survive a copy-paste out of a chat window. They are undetectable by eye and are the single strongest mechanical marker, because almost nothing a human types by hand introduces them.

4. Emoji-headed bullet lists. Structured answers with a decorative emoji at the start of each bullet, section headers built from bold text rather than real headings, and lists nested three levels deep are a house style that chat interfaces encourage and most human writers do not use.

Style-level patterns are much easier to get wrong, because plenty of careful human writing shares them. Treat these as prompts to look closer, never as findings.

5. Hedged, evenly balanced framing. Sentences that present every question as having two reasonable sides, paragraphs that open by restating the question, and closing lines that summarise without adding anything. This comes from training that rewards being helpful and non-committal at once.

6. The "not just X, but Y" construction. Along with "it's worth noting that", "delve into", "navigate the complexities of", and "in today's fast-paced world". These phrases are not banned words — they are ordinary English. What stands out is how often they cluster together in a single short piece.

7. Unusually uniform structure. Paragraphs of near-identical length, sections with a suspiciously equal number of bullets, and a conclusion that mirrors the introduction. Human drafts are lumpier; a paragraph runs long because the writer had more to say about that particular thing.

The vocabulary signals in particular decay quickly. Every list of "AI words" that circulates gets absorbed into how people write and into how the next generation of models is tuned. Mechanical signs are more durable because they are byproducts of encoding, not of style.

The signs above are about characters and phrases. There is another layer that sits above the sentence — how the whole piece is shaped — and it is the layer people most often point at when they say something "reads like AI".

• Uniform paragraph length. Three to four sentences, over and over, with no one-line paragraph for emphasis and nothing running long because the writer got interested.
• The opener that restates the question. "Great question. Choosing a database involves several considerations." A paragraph of throat-clearing before any content, inherited from a chat format where restating the prompt is a reasonable way to begin.
• The closer that adds nothing. "In conclusion", "Overall", "Ultimately" followed by a summary of what you just read. Fine in an essay with an argument to land; odd on a 300-word answer.
• Both-sides framing by default. Every trade-off presented as balanced, every recommendation hedged with "it depends on your specific needs", even where one option is plainly better for the case described.
• Everything as a list. Headings, bullets, and numbered steps applied to material that has no sequence and no parallel structure — the bullet points to paragraph tool exists because converting that back by hand is tedious.

Now the important part. These are the weakest signals on this page, weaker than the stylistic ones, and it is worth being blunt about why.

Every item in that list is taught. Uniform paragraphs and a restated opener are what school essay instruction asks for. "In conclusion" is a standard academic move. Balanced framing is professional caution — it is how a consultant, a lawyer, or a doctor is trained to write, because overclaiming has consequences. Heavy list use is house style across technical documentation, corporate reporting, and most internal wikis. A model writes this way because people write this way; the pattern came from human text in the first place.

That is also why detectors that score on structural features produce false positives on exactly the writing that is most carefully edited, most formal, and most often produced by non-native English speakers who learned English through formal instruction. Uniformity is a property that rewards drilling, and drilling is what a lot of people have done.

So this section is here to help you recognise a register, not to give you a test. The em dash guide makes the same point about punctuation, and it holds harder here: a document can hit every structural pattern above and be entirely human, and a document can hit none of them and be entirely generated. Nothing on this page, and nothing this site's tools report, establishes who wrote something.

A practical order of operations, cheapest check first:

1. Scan for invisible characters. This is the highest-signal, lowest-effort check. Paste the text into the Invisible Character Viewer, which deletes nothing and labels each hidden character in place — [ZWSP], [SHY], [BOM], [LRM], [NBSP], [NNBSP] — so you can see how many there are and where they sit, including the non-breaking and narrow no-break spaces. The Invisible Character Scanner is the other half of the pair: it reports totals for hidden characters, zero-width characters, and soft hyphens and strips them, leaving no-break spaces in place.
2. Count the mechanical artefacts together. The AI Text Cleaner reports em dashes, smart quotes, invisible characters, and emoji in a single pass, so you can see whether several signals co-occur rather than chasing one at a time.
3. Read for the stylistic patterns. Only after you have the counts. Numbers first stops you from talking yourself into a conclusion.
4. Weigh the context. Text pasted from Word will carry smart quotes regardless of who wrote it. Text typed into a plain textarea will not.

Co-occurrence is what matters. A single em dash means nothing. Em dashes and smart quotes and two zero-width spaces and emoji bullets, in a document that was typed directly into a plain-text form, is a genuinely unusual combination.

Every sign above has an innocent explanation.

Word converts double hyphens into em dashes, and both Word and Google Docs convert straight quotes into curly ones automatically. Copying from any modern website can carry invisible characters. Professional editors write in exactly the balanced, hedged register that reads as machine-generated. Non-native English speakers often write in a more formal, structurally regular style, which is precisely why commercial AI detectors flag their work at markedly higher rates.

There is also the obvious circularity: a person can paste AI output into a text editor, run a cleanup pass, and remove every mechanical sign in about ten seconds. The signals that are easiest to measure are also the easiest to erase.

So the honest position is this. Formatting analysis tells you what characters are in a document. It cannot tell you who put them there. Our tools deliberately report counts and never render a verdict, because a verdict would be a guess wearing a number as a disguise. If you need to know whether a person wrote something, the answer comes from talking to them about it — draft history, sources, and the ability to explain their own argument — not from counting dashes.