Question 1

Is this an AI detector?

Accepted Answer

Not in the sense of naming an author, and the reason is a number we measured rather than a disclaimer. Two things it does find reliably: the style habits of language models, which put all 35 of our AI test texts above the lowest band while 28 of the 35 human texts from the year 2000 stayed inside it, and the invisible characters that "humaniser" tools insert, which it catches without a single false positive across the whole corpus. What it cannot do is rule out a person who writes plainly. On 6,000 passages of genuine human business email from 2000 and 2001, a quarter reached the second band. A trained detector such as Pangram or GPTZero reports a false positive rate thousands of times lower, because a model trained on millions of documents can weigh things a rule cannot. So read a high score as "this reads like AI", never as proof of who typed it.

Question 2

Can it catch AI text that was written carefully?

Accepted Answer

Often not, and we measured that. We ran the checker over 35 texts published in the year 2000, long before AI could write: speeches, news reports, press releases, reviews, columns and how-tos in five languages. A current AI model then wrote 35 texts for the same briefs. The texts from 2000 scored 0 to 30, and 28 of the 35 landed in the lowest band, with none reaching the third. The AI texts scored 12 to 64: every one of them came out above the lowest band, and 20 of the 35 reached the top two. Unedited AI writing in the usual chatbot register, the kind people paste in here, scores 27 to 76. A high score means the text carries the habits; a low score only means it does not.

Question 3

Which words make text sound like AI?

Accepted Answer

The usual suspects are delve, tapestry, testament, pivotal, seamless, leverage, foster, underscore, meticulous, multifaceted and realm. On their own they are ordinary English words; AI models simply use them far more often than people do.

Question 4

Why does AI use so many em dashes?

Accepted Answer

Language models learned from edited text where the em dash is common, and they use it as an all-purpose connector. Two or three per paragraph is a strong tell. Replacing most of them with commas, colons or full stops makes a text read more naturally.

Question 5

How do I make my text sound less like AI?

Accepted Answer

Cut stock phrases and transition words, swap typical AI words for plain ones, vary your sentence length, add concrete details only you know (numbers, names, what actually happened) and end on a real point instead of a summary.

Question 6

Can a text written by a person get a high score?

Accepted Answer

Yes. Formal business writing, marketing copy and school essays often share the same habits. That is why the score describes style, not authorship. Use it to edit your own work, never as evidence against someone else.

Question 7

What does the sentence rhythm chart show?

Accepted Answer

Each bar is one sentence, and its height is the number of words. People naturally mix short and long sentences, so their chart looks jagged. AI text often produces bars of similar height, which reads as monotonous.

Question 8

Does it work in other languages?

Accepted Answer

Yes. There are versions in Dutch, German, French and Spanish, each with its own lists of typical AI words and phrases for that language.

Question 9

Can it tell whether an email was written by AI?

Accepted Answer

Yes, and here the checker is deliberately strict. A polite business email is built from set phrases: the opening pleasantry, the offer to answer any questions, the thanks, the closing line. To find out how often people really reach for them, we measured 96,000 sent messages from a business email archive of 2000 and 2001, written before any model could draft one: 94 percent of 150-word stretches contain none of those courtesies, 99.75 percent contain at most one, and not one stretch in 6,000 contained more than two. Every courtesy the checker finds therefore counts, and three of them already puts a text beyond anything in that archive. AI-written emails scored 45 to 72 in our own test. The point is not to accuse the writer: a line that reads as filler is worth cutting, whoever typed it.

Question 10

What are hidden characters, and why does the checker flag them?

Accepted Answer

Some tools that promise to make AI writing undetectable tamper with the characters themselves: they push invisible marks inside words, or swap a Latin letter for an identical-looking one from the Cyrillic or Greek alphabet. The text still looks normal to you, while the machinery a detector relies on breaks. The checker finds both and marks the word they sit in. If it flags your text and you used no such tool, the usual cause is a copy out of a PDF, which leaves a soft hyphen at every line break. Either way they are worth removing, because invisible characters break search, copy-paste and screen readers.

Question 11

Is my text stored?

Accepted Answer

No. The analysis runs entirely in your browser. Nothing you paste is sent to a server or saved.

Does This Sound Like AI?

How this checker was tested

Writing from before AI could write

The same brief, written by a model

One last pair per language

What it still cannot do

Frequently Asked Questions