A fair checklist for teachers and editors
You have a text in front of you and a feeling. Too smooth, too tidy, no personality. You paste it into a detector and it says 78%. What now?
Short answer: slow down. A number can start a conversation. It cannot end one.
What a score can and cannot prove
A detector measures style. It looks at how even the sentences are, how often stock phrases appear, how many connecting words are packed in. It does not see who typed the text, and it has no access to how the text was made.
So a high score only means "this text has features that machine text often has". It is not proof of authorship, and in our own honest numbers the signal is weak: on a large corpus the current analysis catches only a modest share of AI texts when the false-positive rate is held at 1%, and modern models pass more easily than early chatbots.
Never decide on a score alone. Not to fail a student, not to reject a freelancer, not to send an accusation.
Who gets flagged unfairly
Some honest writers look "machine-like" to any detector:
- people writing in a second language, who lean on simple and correct constructions;
- editors and journalists, whose polished prose has been smoothed on purpose;
- authors of formal genres: reports, letters, scientific style;
- anyone who ran their draft through a grammar tool or a translator.
If your suspect fits one of these groups, treat the score with extra doubt. We wrote about this in When the detector mistakes a human for AI.
The checklist
Go through these steps in order.
- Reread the text yourself. Are there specifics, such as names, dates, a real example, a mistake only a person would make? Or only general words?
- Compare with earlier work. Same vocabulary, same typical errors, same rhythm? A sudden jump in polish is a reason to ask, not to accuse.
- Ask for the process. Drafts, version history, notes, the sources the author read. This is the strongest evidence you can get, because it shows how the text appeared.
- Talk to the author. Ten minutes of questions about the content beats any percentage.
- Run a check, if you still want one. Read the result as a hint about where to look.
- Decide on several signals together, and write down what they were.
Using the evidence list as a conversation starter
AuthorText does not only return a number. It lists evidence with line numbers: where a stock phrase appears, where the paragraphs are nearly the same length, where connecting words pile up.
That list is useful in a conversation, because it lets you ask about a concrete place instead of the whole text. "Line 14 says 'it is worth noting that it plays a crucial role'. Tell me what you meant here." A person who wrote the paragraph can usually explain it, rephrase it, or admit they borrowed the phrase. Someone who did not write it often cannot.
Even then, an explanation is not a verdict either way. Plenty of people use stock phrases, and plenty of AI text is edited by hand.
How to talk to the author
- Start with curiosity: "I would like to understand how this text came about."
- Say what you saw, not what you conclude: "The rhythm is very even and there are several typical phrases."
- Ask for the draft history before anything else.
- Offer a short follow-up, spoken or written on the spot, about the same topic.
- Leave room for being wrong. If the author has drafts and answers the questions well, the matter is closed.
Don't say "the detector says you cheated". It said no such thing, and the author will fairly stop trusting the process.
What to do
Use the score as a reason to look closer, never as the reason to punish. Ask for drafts first, talk second, and decide only when several independent things point the same way. If you want a line-level view to guide the conversation, check the text on AuthorText and open the evidence list. For longer, regular use, see the plans.