How does an AI detector decide who wrote something?
An AI detector decides whether a piece of writing, or sometimes a picture, looks machine-made. It gives a score, not proof. It compares the work with patterns from examples of person-made and machine-made work, and flags anything over a line. No line is perfect: some real work gets flagged, and some machine-made work slips through.
In one line: A program that scores writing or pictures for how machine-made they look, and flags the ones above a line.
Last reviewed: 25 October 2026 3 min read
What decision does it make?
AI is short for artificial intelligence. An AI detector makes one yes-or-no choice about a piece of work: does it look like a machine made it? Some detectors check writing and some check pictures. This page uses writing as the example.
| Decision | Possible answers |
|---|---|
| Was this made by a machine? | Yes: flag it as machine-made. No: do not flag it. Either way, the answer is a guess from a score. |
What does it look at?
- Word patterns. It checks which words and phrases appear and how the sentences are built, and compares them with the examples it learned from.
- How predictable the words are. Some detectors measure how unpredictable the text is, a measure called perplexity.
- The examples it learned from. A detector trained mostly on one kind of writing, such as school papers, can do much worse on other kinds.
How does it learn the rule?
- Collect two piles. People gather writing known to be by people and writing known to be machine-made.
- Find the differences. The program learns which patterns show up more in one pile than the other.
- Score new writing. Each new piece gets a score for how machine-made it looks.
- Set the line. Someone picks the score that counts as flagged. A lower line flags more work. A higher line misses more.
Where can it go wrong?
- Real writing gets flagged. A person's own work can get a high score. That is a wrong yes, and it can feel very unfair.
- Machine-made writing slips through. If someone rewrites machine-made text on purpose to fool a detector, the score can drop below the line. That is a wrong no.
- A score is not proof. In the United States, the Federal Trade Commission ordered one detector company to stop making accuracy claims it could not back up with evidence.
- No line is perfect. Move the line down and more real writing is flagged. Move it up and more machine-made writing gets through.
Try it
Toy example with made-up scores: move the line. Try to flag every machine-made text without flagging anything a person wrote. Is there a line with zero mistakes?
A toy example: ten short texts with a made-up machine-made score and who really wrote each one. The numbers are invented for practice.
| Item | Values | Right bin |
|---|---|---|
| A short product description, made by a program | Machine-made score: 90 | Flagged as machine-made |
| A five-paragraph essay, made by a program | Machine-made score: 85 | Flagged as machine-made |
| A science summary, written by a student | Machine-made score: 80 | Not flagged |
| A list of study tips, made by a program | Machine-made score: 70 | Flagged as machine-made |
| A book report, written by a student | Machine-made score: 65 | Not flagged |
| A thank-you note, written by hand and typed up | Machine-made score: 55 | Not flagged |
| A story made by a program, then edited by a person | Machine-made score: 45 | Flagged as machine-made |
| A poem with odd, surprising words, made by a program | Machine-made score: 30 | Flagged as machine-made |
| A diary entry full of jokes and spelling slips | Machine-made score: 15 | Not flagged |
| A text message to a friend | Machine-made score: 10 | Not flagged |
Check it
What does an AI detector give you?
Answer: A score for how machine-made it looks. It gives a score and compares it with a line. A score is a guess, not proof.
You move the line lower. What happens?
Answer: More person-written texts get flagged. A lower line catches more machine-made writing, but also more writing by people.
A detector flags an essay that a student wrote alone. What is the fairest next step?
Answer: Look at the drafts and notes, and talk with the writer. Detectors make mistakes both ways, so a person should look at more than the score.
For older readers
An AI detector is a binary classifier. Some are trained on labeled examples of person-written and machine-written text. Others use a language model to measure how predictable the words are, a signal called perplexity. Flagging a person's work is a false positive; missing machine-made work is a false negative. Moving the decision threshold trades one for the other. A detector trained on one kind of writing can do much worse on another, so any accuracy claim only means something for the kind of writing it was tested on.
Questions
They can be right often and still be wrong. They give a score, not proof, and every line flags some real writing or misses some machine-made writing. How well one works depends on the kind of writing it was tested on.
No. A score is a clue at most. A fair check also looks at drafts and notes, and includes a conversation with the writer.
Its patterns scored above the line. A detector can give a high score to writing a person made, even when a person wrote every word. Keep your drafts and notes, and show them to the person who checks your work.
Some are built for pictures instead of words. They work in a similar way: a score and a line, with mistakes possible both ways.
Sources
- FTC: final order on an AI content detector's accuracy claims (opens another website)
- University of Washington CSE 416: Classification (confusion matrix and changing the threshold) (opens another website)
- NIST: reducing risks posed by synthetic content (opens another website)
- FTC: order requiring an AI content detector company to back up its accuracy claims (opens another website)