What an AI scanner is measuring
An AI detector does not know who wrote anything. It measures statistical properties of the text. The main one is perplexity, which is roughly how surprising each word is given the words before it. Machine-generated text tends to pick likely words, so it scores as smooth and regular. The tools also look at burstiness, which is how much sentence length and complexity vary.
So a detector flags writing that is statistically unsurprising. Plenty of human writing is statistically unsurprising: writing by someone using a limited vocabulary, writing in a second language, writing to a formula because that is what the assessment rewarded, and writing by anyone who has read enough machine text to have absorbed its rhythms.
That is the mechanism. Everything below follows from it.
What the evidence shows
The tools score hybrid writing backwards
Atamhenwan (Education and Information Technologies, 2026) found Turnitin over-detects light AI use and under-detects heavy use. It scored an entirely machine-written script at 60%, while text run through a humaniser came back at 0%. The student who used a little help is at more risk than the student who used a lot.
Older detectors failed non-native English speakers badly
Liang et al. (Patterns, 2023) ran genuine human essays by non-native English speakers through seven detectors. An average of 61.3% were flagged as AI, 97.8% were flagged by at least one detector, and 19.8% were flagged by all seven. Enriching the vocabulary cut the false positive rate to 11.6%, which tells you the tools were measuring fluency. This tested the 2023 generation of detectors, and vendors have since worked on it, so treat it as evidence of the mechanism rather than a current benchmark.
Simple edits defeat them
Perkins et al. (2024) measured detector accuracy at 39.5%, falling a further 17.4 points to 22.2% once basic evasion techniques were applied. Anyone trying to cheat will get past these tools. The people who get caught are the ones who were not trying.
The headline accuracy figures cannot be independently checked
Current vendors report false positive rates around 0.004%. At that rate you would need roughly 732,000 human texts to observe even 30 false positives, and every published independent dataset is one to two orders of magnitude smaller. When a vendor reports one false positive in 24,586 samples, the statistical interval around that single event puts the upper bound at more than five times the headline figure. The number may be right. It cannot currently be verified by anyone outside the company.
Why a perfect detector would still be the wrong tool
The usual assumption is that the software needs to catch up. My objection holds even if it does.
Detection casts the educator as a police officer. The work of teaching is to be a co-creator of knowledge with the student, and the moment you deploy surveillance you have changed what a teacher is for. That is a claim about what teaching is, and it stands whatever the accuracy figures say.
There is an equity dimension on top of it. Detection fails unequally, and it penalises the students least able to absorb the cost of being wrongly accused.
If you have been accused
- Ask what the score actually is. A percentage is not a probability that you cheated. Ask what the number measures and what threshold the institution uses.
- Ask for the policy. In my audit of 96 UK universities for HEPI, 41% had no public AI policy at all. If there is no published standard, there is no standard you could have breached.
- Produce your process. Drafts, version history, notes, search history, and the messy middle of the work. This is the strongest evidence there is, and it is why I tell students to keep it.
- Ask whether a human read it. A detector score is not a finding. Someone has to make a judgement and be accountable for it.
An allegation costs something even when it is withdrawn. That cost is the reason to get this right at policy level rather than case by case.
Try it yourself
The fastest way to understand the limits is to attempt the task yourself. I built these to be played in a few minutes.
Common questions
Are AI detectors accurate?
Vendors report very high accuracy, and those figures cannot currently be checked independently because the datasets needed would have to be roughly a hundred times larger than any published study. Peer-reviewed work on earlier tools found accuracy as low as 39.5%, and found that hybrid human and AI writing is scored backwards.
Can an AI scanner be wrong about my work?
Yes. These tools measure how statistically predictable your writing is, so writing that is clear, formulaic, or produced in a second language can score as machine-written. A score is a statistical property of the text, never evidence of what you did.
Should universities use AI detection software?
I argue no, on grounds that hold even if the software improves. Detection casts the teacher as a police officer, when the work of teaching is to be a co-creator of knowledge with the student. The better response is assessment that makes the process visible.
What should institutions do instead?
Publish a clear policy, ask staff to declare their own AI use to the same standard asked of students, and redesign assessment so the process carries the marks. I run audits and training on exactly this.
Getting this right at your institution
I work with universities, schools, and organisations on policy, assessment design, and staff capability.