Why AI detectors give different scores for the same text

Agreement on passages counts most

No single AI detector gives the true score for a text, so three different scores for one essay are normal. Nothing is broken. Each detector is built by its own team, looks for its own patterns and has its own way of turning what it finds into a number. So if three detectors gave your essay three different scores, which one should you trust? None of them on its own. The places where they agree carry more weight than any single number, and even then, agreement does not prove how the essay was written. Below you will find where the differences come from and what to do with the numbers on your screen.

Each tool measures something different

Every detector reads a text and gives back an estimate of how much the writing resembles text made by AI. The task is the same on every site. The way of doing it differs from company to company. One team may pay most attention to how single sentences are built. Another may weigh the whole text, from the first line to the last. And each team builds its tool from its own examples of human writing and AI writing, which means each detector has its own picture of the two.

The number at the end differs as well. On one site, a percentage may stand for the share of sentences the tool marks. On another, a number of the same kind may say how sure the tool is about the whole text. Two scores that look alike can therefore mean quite different things, and setting them side by side compares results that were never measured on the same scale. For a plain account of what detectors look for, read the ScanForAI page on how AI detection works.

Different test texts, different results

Before a company says how well its detector works, it tests the tool on a set of texts. Each company picks its own set. A detector that does well on long school essays may do worse on short emails, poems or lab reports. ScanForAI’s own view is that no one can honestly say which detector is the best, because the ranking changes with the texts chosen for the test. A tool that comes first in one comparison can fall far down in the next.

Your essay is a test of its own, and the result depends a lot on what kind of text it is, how long it is and how much it was edited. A detector that does well on history essays may misread a personal statement, which reads quite differently. Length matters as well. Two or three paragraphs give any detector very little to go on, and we treat anything under about 300 words as weak evidence, whatever score comes back. Editing mixes things up again. Perhaps you rewrote your opening twice, and a grammar tool changed a few sentences in the middle. One detector may read that mix very differently from the next.

What a fair comparison would need

For three scores to be worth comparing, the three detectors must have read the same thing. This can go wrong without anyone noticing. Did each site get the title, the quotes and the references, or did you leave something out on one site because the box was too small? Some sites read only up to a certain length. Paste the same version into every tool at full length, and check on each site that the whole essay was read, not only its start.

Two more things would make it fair, and you will rarely have them. You would need to know what each number stands for, and how each detector does on texts of your type and length with about as much editing. Look for the first in each site’s help pages; without both, three percentages from three sites say less than they seem to.

What to do with three numbers

Do not add the numbers together or take an average of them. Scores that mean different things do not make a better score when you combine them.

Open the three results side by side, in three browser tabs or as three printouts, and go through your essay one paragraph at a time. At each paragraph, note which detectors mark it. When two detectors that work in different ways mark the same passages, that means more than the score of either one alone. Suppose one tool gives your essay 80% and another gives it 35%. The numbers do not match, but both may mark the same two paragraphs near the middle. Those are the paragraphs a reader might read more closely, and they are also the ones your drafts can speak for.

If nothing lines up, because the tools mark different passages or one marks a great deal and the others almost nothing, the evidence is thin or mixed. Read a result like that as “not enough to say”, and never as “human”.

Before you close the tabs, write down which detector gave which score and on which day, and keep a copy of the essay exactly as you scanned it.

You can scan the same text on ScanForAI for free, compare which passages each tool marks rather than the percentages, and read the confidence line that comes with the result. We explain on the ScanForAI accuracy page why agreement on the same passages counts for more than any one number. Whatever the results show, keep your essay and its files exactly as they are.

Why none of them is proof

No detector is 100% accurate. It sees the finished words and nothing else, so it can only estimate from them. Who typed the essay, which notes came first and how many drafts there were leave no trace in those words. For that reason a score must never be used to accuse anyone, and it must never be the only reason for a decision about a person. Each score is one signal among several.

If you teach or edit, take the scores as a reason to read and to talk. Read the text yourself. Ask the writer how it was made, and ask to see drafts or notes. Reading has its own limits, and the guide on how to tell if a text was written by AI explains what careful reading can and cannot show. If you are the writer, your drafts, notes and saved versions say far more about your essay than any of the three numbers. Whoever holds the scores, it is a person who weighs them with everything else and then decides.

Everything above depends on two things. You can see the full results, and you can run the text yourself. Neither is true when someone else ran one detector and gave you one number: no passages are marked, and nobody has told you what the number means. Putting the essay through other tools will not show how that number was reached. A written question will do more. Ask which detector was used, what its score stands for and what else the decision rests on. Then take your drafts and notes to the meeting about the essay.

This guide is general information for students and teachers. It is not legal advice, and your school’s own rules come first. A score from any AI checker is a signal, never proof.

Cite this guide

ScanForAI team (2026). Why AI detectors give different scores for the same text. ScanForAI. https://scanforai.com/guides/why-ai-checkers-give-different-scores

You may share and adapt this guide if you credit ScanForAI and link to it. See Reuse and citation.

AI detection is not 100% accurate. No detector, including ScanForAI, can prove if a text was written by a person or by AI. Every result is an estimate based on writing patterns. It can be wrong in both directions. Never use it as the only reason for an important decision about a person, such as a school decision, a punishment or a hiring choice. Never use it to accuse someone.