Do AI Detectors Actually Work?
AI detectors claim to tell whether a piece of text was written by a person or produced by a language model. They are widely used in education and increasingly in hiring and publishing.
The short answer to whether they work: not reliably enough to be treated as evidence.
How they claim to work
Detectors mostly measure two statistical properties of text.
- Perplexity is how surprising the word choices are. Language models tend to choose likely continuations, so their output is often less surprising than human writing.
- Burstiness is how much sentence length and structure vary. Human writing tends to swing between long and short sentences more than machine output.
A detector scores the text on these measures and returns a probability.
Notice what this means: the detector is measuring predictability, not authorship. It cannot see where the text came from. It is making an inference from style, and that inference breaks in both directions.
Why false positives happen
Writing that is clear, structured and consistent scores as machine-like, because those are exactly the qualities that make text predictable.
Groups that get flagged more often than they should:
- Non native English speakers, who often use simpler constructions and a narrower vocabulary. Published analyses have found markedly higher false positive rates for this group, which raises a fairness problem on top of an accuracy problem.
- People writing in formal or technical registers, where conventional phrasing is expected.
- Anyone following a template, such as a lab report or a structured essay format.
- Writers who were taught to write plainly.
The uncomfortable implication is that a student who writes well in a plain style is more likely to be accused than one who writes messily.
Why false negatives happen
Avoiding detection is not difficult. Asking a model to vary sentence length, light editing by hand, running text through a paraphrasing tool, or simply using a newer model all reduce scores substantially.
So the people most likely to be caught are those making no effort to hide anything, while deliberate misuse often passes.
What the evidence shows
Independent testing has repeatedly found error rates too high for consequential decisions. OpenAI withdrew its own AI text classifier, citing low accuracy. Several universities have advised staff not to rely on detector scores as proof of misconduct.
Detector companies often advertise very high accuracy figures. Those are usually measured on clean test sets of unedited machine output against unedited human text, which is not what real submissions look like.
If your writing is flagged
Being falsely accused is genuinely distressing, and the burden of proof tends to land on the writer. Practical steps:
- Ask what the accusation is based on. If it rests only on a detector score, say so plainly.
- Show your process. Version history in Google Docs or Word, drafts, notes, browser history and timestamps are far stronger evidence than any score.
- Offer to discuss the content. Someone who wrote a piece can explain its choices and sources.
- Point to the published error rates, including the documented bias against non native speakers.
- Keep drafts from now on. Writing in a tool that saves version history is the single best protection.
For teachers and employers
If you use these tools, treat a score as a prompt to look more closely, never as a finding. A conversation about the work reveals more in five minutes than any detector.
Assessment designed around process, such as drafts, oral defence and in class work, holds up better than trying to detect output after the fact.
What they are reasonable for
Detectors have some value as a soft signal at scale, for instance flagging a batch of submissions for human review, as long as nobody is penalised on the score alone. The problem is not that they produce no information. It is that they are used as proof when they are closer to a guess.
How the underlying models produce text, and why their output is statistically predictable, is covered in our guide to generative AI.
Frequently asked questions
Which AI detector is the most accurate?
Independent comparisons do not show a clear winner, and results vary a lot by text type and length. None are accurate enough for high stakes decisions.
Can AI detectors be wrong?
Yes, in both directions. Human text flagged as AI, and AI text passing as human, are both well documented.
Does editing AI text avoid detection?
Usually yes, which is part of why detection is unreliable. That does not make it honest where the work is supposed to be yours.
Why do non native speakers get flagged more?
Their writing tends to use more predictable vocabulary and structure, which is exactly what detectors measure.
Can Turnitin detect AI writing?
It offers an AI detection feature with the same underlying limitations. Several institutions have disabled or restricted it.
How do I prove I wrote something myself?
Version history, saved drafts and the ability to discuss your own work in detail. Our guide on using AI well covers the line between drafting help and handing over the work.




