There is no official scale that tells you which AI-detection percentage is "bad". A detector's figure is a probability estimate produced by a model reading your text, and it moves with the length of the text, the model version, and the run itself. Treating 33% as a grade is reading a measurement as a verdict — the two are different things, and only one of them exists.
What the percentage is actually reporting
Detection tools estimate how predictable your text is to a language model. Two properties carry most of the signal:
- Perplexity — how surprised a model is by your word choices. Predictable phrasing scores low; unusual phrasing scores high.
- Burstiness — how much sentence length and structure vary. Human drafts tend to swing; generated drafts tend to even out.
Neither property measures authorship. They measure whether your prose looks statistically typical. A careful writer who edits toward plain, even sentences moves in the same direction a generator does, and a writer working in a second language can land there without ever opening a model.
Why the number jumps around
Short text is the most common reason two runs disagree. A 200-word passage gives the model very little to estimate from, so a paragraph or two of quoted material, a list of product names, or a pasted citation block can move the figure by tens of points. The same draft scored in two different tools will rarely match, because the tools are different models, not two views of one instrument.
Is there a threshold that means something?
Not an official one. Institutions set their own internal cutoffs, and they are usually conservative because the cost of a false accusation is high. What is worth knowing is that the false-positive problem is documented rather than hypothetical: a study of AI-detection tools by Liang and colleagues (arXiv 2304.02819) found that essays written by non-native English speakers were disproportionately flagged as machine-generated. If you write in a second language, that finding is the one to keep in view when a number arrives.
So the honest reading of "is 33% bad" is: 33% is not a grade, it is a reading with an error bar nobody printed. What matters is whether the tool points at specific passages, and whether anyone is treating the reading as evidence of intent.
What to do when a figure lands on your draft
- Ask for the passages, not the number. A report that highlights the segments it flagged can be argued with. A single percentage cannot.
- Produce the process, not a defence. Version history, draft timestamps, notes, sources you opened — these describe how the piece was made, which is the actual question under dispute.
- Ask which tool, which version, and which language. The three together explain most of the variance between two reports on the same text.
- Do not rewrite to chase the number. Rewriting until a score drops changes the surface of the text and leaves the substance where it was — and on a flagged passage it usually makes the writing worse.
The boundary
None of this tells you whether the piece is worth publishing. A detector answers "does this look machine-typical"; a publish decision answers whether the draft delivers what it promised, whether a reader can act on it, and whether its claims carry their evidence. Those are different questions, and a percentage has never been able to answer the second one. If that is the question in front of you, score the draft and read three answers instead.
Frequently asked questions
Is 33% AI detection bad?
There is no official scale that makes 33% a grade. It is a model estimate that shifts with text length, tool version and run, and it does not measure authorship on its own.
What percentage of AI detection is acceptable?
Institutions set their own cutoffs, and they differ. No published standard defines an acceptable figure, which is why the passages behind a flag matter more than the number in front of it.
Is 10% AI detection bad?
It is a low reading on an instrument with a documented false-positive rate, not a clean bill. The useful question is whether the tool named specific passages, not whether the figure was small.
Can a human-written text score high?
Yes. Formal, technical and non-native English writing are the documented cases, because plain and even prose resembles what generators produce.
Does rewriting lower the score?
Sometimes, and it is the wrong goal. Rewriting to move a number changes phrasing, not substance, and on a flagged passage it usually costs clarity.