Why percentage scores fail
You have probably seen one. Your essay is 23% AI-written. Or: this submission scores 78% likely to contain AI.
The number looks like a measurement, in the way a cholesterol reading is a measurement, and that is most of the problem with it. We used to show one ourselves. We took the last of them out in version 4.49.0, and this page is the reasoning.
What the number is actually doing
A detector scans for statistical patterns, gives passages a score, and averages them. What comes back is an average of guesses, presented in the same format as a laboratory result.
That would still be useful if the average carried the information you needed. It does not, for three reasons.
It hides where the problem is
Your 2,000-word essay comes back at 15%. Now what?
You do not know whether that is three hundred words scattered evenly through the piece or one paragraph that reads as though somebody else wrote it, and those are entirely different situations calling for entirely different afternoons. You do not know whether the flagged material is carrying your argument or joining two paragraphs together. You cannot tell a passage the detector was almost certain about from one it barely noticed, because both have been folded into the same average.
So you know something is wrong somewhere, which is roughly the least useful shape information can arrive in.
It treats every signal as though it weighed the same
Some patterns are distinctive. "Importantly, this demonstrates X. Significantly, this shows Y. Notably, this proves Z." Three sentences in a row opening on an intensity adverb is specific and genuinely uncommon in writing by people.
Other patterns are ambiguous, and hedging is the obvious one. "It could be argued that climate change affects agriculture" is what students are taught to write, and it is also what models produce, and no amount of statistics separates those two without knowing something about the writer.
An average adds both to the same total. That is how an essay written entirely by a careful student ends up at 35%, on the strength of hedging they were marked down for not using the year before.
It knows nothing about context
A geography student quoting IPCC language will match "scientific register" patterns, because they are quoting the actual source. The pattern match is real and it means nothing.
Reading whether a passage sits consistently with the rest of the document catches this, since a quotation is meant to read differently and an inserted paragraph is not. An average cannot make that distinction, as it has already thrown away the thing the distinction depends on.
Why the number persists anyway
Not conspiracy. It is simply easier in four ways at once: easy to explain to somebody in a corridor, easy to compute, easy to compare between documents, and easy to defend afterwards, because a tool that reported a number and turned out to be wrong can say the number was accurate and the reading of it was not.
Every one of those conveniences belongs to whoever built the tool. None of them belong to the person holding the essay.
What we show instead
Passages, marked one of two ways.
Grey for a passage that reads the way machine writing reads. Gold for one carrying the marks of a person. Click either and you get the reasoning that put it there, written out, including where the two reviewers disagreed with each other before settling.
That is a longer thing to read than a number. It is also the only version of this that tells you what to do next, and telling you what to do next is the entire job.
What this costs us to say
Two things, and it is fairer to name them than to pretend the trade is free.
The first is that you cannot compare our result to last week's result at a glance, because there is no figure to compare. The second is that a marked passage takes a minute to read where a percentage takes a second, and on a stack of forty submissions that minute is real.
We think it is worth it, because the second thing a percentage buys you is the ability to argue about the percentage, and we have watched enough of those conversations to want no part in supplying the ammunition.
How to talk about a result
If you are explaining an analysis to a tutor or a panel or your mum, describe what was marked rather than reducing it to a figure.
"Three passages came back marked. One of them is my conclusion, which I had written in a hurry and have now rewritten. The other two are quotations, and I have shown where they came from."
That is a sentence somebody can respond to. A number is a sentence somebody can only agree or disagree with.
What to ask of any detector
Whether it shows you where. Whether it tells you why, in words. Whether you can see the flagged text with the paragraphs either side of it still attached. Whether you can revise without leaving the tool and losing your thread. And whether it distinguishes a strong signal from a weak one, or quietly averages them together.
We would rather you asked those questions of us than took our word for any of it.
The part that has not changed
Removing the number does not make detection certain, and nothing on this page should be read as claiming it does. A mark still means a passage reads a certain way rather than telling you who wrote it, and that limit sits underneath every detector including ours.
What changes is what you can do about it.