ThinkFacility
AI writing tells

Writing Tells

Paste text. Get a score from 0 to 100 and every signal behind it, highlighted in place.

The score measures how much the writing leans on the habits language models overuse. It is not proof of anything. How it works and where it fails.

Text

0 words
Paste text on the left and press Score, or load the example.

What the score means

A high score means the writing leans hard on habits that language models overuse. It doesn't mean a machine wrote it. Plenty of careful writers love em dashes and lists of three.

We count fifteen things. Twelve push the score up, like em dashes and words such as "delve". Three pull it down, like contractions and specific numbers, because those are hard for a model to fake without being asked.

Nothing here is trained on a corpus. We picked the thresholds by hand, which is why you can see every count. If you think we're wrong about em dashes, the number is right there and you can ignore it.

Short text is unreliable. Under 80 words we say so on the page. Give it 200 words and the score settles down.

The tells, biggest first

TellWhat it looks likeWeight
Model lexicondelve, tapestry, testament, realm, robust, seamless, crucial, foster, navigate, underscore, showcase, journey16
Em and en dashesA dash where a comma or a full stop would do12
Even sentence lengthEvery sentence about the same length12
"Not X, but Y""It isn't about speed, it's about trust"10
Stock phrases"in today's", "it's worth noting", "plays a crucial role"8
Lists of three"clarity, precision, and impact"8
Label-colon linesBullets that open with a bold word and a colon8
Transition openersAdditionally, Furthermore, Moreover, Ultimately8
Emoji or heading openersA rocket at the start of a line that isn't a document5
Assistant sign-offs"I hope this helps", "let me know if"5
"Not only … but also"The paired intensifier4
Question then answer"Why does this matter? Because…"4
Contractions subtractsdon't, it's, we've5
Numbers and dates subtractsPrices, counts, years, measurements6
First person subtractsI, my, we, our4

Why there's no verdict

Detectors that tell you "97% AI" get it wrong often enough to ruin someone's week, and they're worse on people writing in a second language, who lean on the same formal constructions models do. We're not doing that.

This page tells you which habits your text uses and where. You decide what it means. The highlighted text is the useful part; the number is just a way in.

Common tells, explained

The em dash. Models produce them far above the human rate, partly because typing one costs you nothing if you're a model. One or two is normal. One in every second sentence isn't.

The lexicon. "Delve" became the joke in 2024, so models moved on. The current crop is quieter: robust, seamless, crucial, foster, navigate, showcase. Any one of them is fine. Six in a paragraph isn't.

Not X, but Y. It sounds like insight and costs nothing to produce. Once you notice it you can't stop noticing it.

Even sentence length. People write a long tangled sentence and then a short one. Models settle into a comfortable length and stay there. This is the tell that survives a vocabulary edit, which makes it the most useful one here.

The label-colon bullet. Bold word, colon, explanation, five times over. Ask a model to explain anything and this is the shape you get back.

The sign-off. "I hope this helps" in an essay means someone pasted straight out of a chat window.

My score is high. Now what?

Read the highlights and fix what you agree with. Swap dashes for the punctuation you meant. Cut the lexicon words and say the plain thing. Turn one "not X, but Y" into a sentence that just says Y. Write a short sentence after a long one.

Then add something only you could know. A number, a name, a date, what actually happened. That's the part no model can fake, and it's what drops the score honestly.

Nothing you paste is saved or sent anywhere.