Human draft, machine polish: what a score sees
There is a difference between who had the thought and who typed the final sentence.
Between a person writing something and a machine writing something there is a large middle, and it is where most real documents sit. Someone outlines, argues and decides; a model tidies the prose. A score responds to the tidying and says nothing about the decisions, which is why arguments about detection rarely end with a number.
What this page is for: reading a detection score in context. It is not offered as a way past a checker, and no figure or example here should be taken as a promise about what any detector will report; scores move, and services disagree.
What changes when a model edits
- Revision with a model is not generation. The claim, the structure and the choice of what to include usually came from a person, and a detector sees none of that.
- Surface features move in a recognisable direction. Sentence lengths even out, connectives repeat, and the small irregularities that mark a particular writer flatten.
- Those are the features these tools measure. The score tracks the polish, not the thinking behind it.
- The score cannot separate polish from generation, because both leave the same surface. A heavily edited human document and a generated one can present almost identically.
- Nothing in the number tells you who chose the claim. That question is answered by asking the writer or by reading drafts, not by reading a percentage.
- Disclosure is a policy decision. Whether assisted editing must be declared is a rule people agree on, not a measurement problem.
Questions that get you further than a score
- Ask what the writer did, in order. Draft, then model, then human edit reads differently from model, then human, and the score often shows which shape it was.
- Separate the claim layer from the sentence layer. Who decided what the document argues is a question you can answer; who typed the sentences is one the score answers loosely.
- Check whether the edits changed meaning. Assisted editing that preserves the argument is a different case from editing that supplies it.
- Compare against an earlier draft if one exists. Version history is stronger evidence than any single reading of the final text.
- Apply the policy you already have, and apply it to everyone. If the rule is about declared assistance, enforce the rule; if it is about machine authorship, say what evidence would establish that.
- Write the reason down. A high score is not a reason, and the next person to read the file should be able to see what you weighed.
Questions people ask
If a model rewrote the sentences, is it AI-written?
That depends on the rule you are applying, and reasonable rules differ. What is not ambiguous is that the surface changed and a detection score will respond to that change. Whether the document is then treated as assisted or as generated is a policy question.
Can a detector distinguish polished human text from generated text?
Not reliably, because it does not measure the process. It measures how uniform the prose is, and a thorough polish and a generation both make prose uniform.
Does that make detection useless?
It makes it useful for what it is: a screen that points at passages worth a closer look, and a poor basis for a finding about a person.
What should a policy say about assisted editing?
Whatever the institution can check and apply evenly. A rule about declared assistance is enforceable; a rule about how much help is too much, measured by a score, is not.
What evidence settles the question?
Drafts, version history and the writer's own account. The score belongs beside those, not in place of them.