AI detection statistics: the measured numbers
Every figure we publish, with the sample size it came from and how it was measured. Third-party figures are cited to their source. Nothing here is an estimate.
Most claims about AI detection are unsourced. This page is the reference: what unslop measures, on how much data, alongside the figures other vendors publish about themselves.
Everything below is either measured on our own corpus or cited to whoever published it.
Error rates on human writing
The number that matters, because flagging something a person wrote is the error that arrives attached to their name.
| detector | wrong flags on human writing | source |
|---|---|---|
| unslop, strict setting | 0.4% of documents | measured, 15,900 documents |
| unslop, public default | 0.8% of documents | measured, 15,900 documents |
| Turnitin, sentence level | 4% of sentences | published by Turnitin |
| Turnitin, document level | under 1% of documents | published by Turnitin |
unslop's strict setting is 99.6% specificity on genuine academic prose.
How ours was measured. A held-out set of 15,900 documents of real academic writing, predominantly pre-2020 arXiv papers whose provenance is not in question because they predate the models. The threshold is fixed on a validation split and the rate is reported on a test split the model never saw.
Detection rates
| what | rate | sample |
|---|---|---|
| unslop on AI academic text, public setting | 88.3% | held-out academic test set |
| unslop on OpenAI GPT output | 89.9% | 7,491 documents, GPT-3.5 through GPT-5.x |
| unslop, best-case generator family | 94.0% | measured at a fixed ~1% false positive point |
How these were measured. 14,892 generated documents spanning seven model families and five years of GPT releases, scored at one fixed threshold rather than a threshold tuned per family.
What the writing itself looks like
Two measured properties, from our own corpus. These are the signals detectors read.
| property | human academic writing | generated text |
|---|---|---|
| within-document sentence length, standard deviation | 8.7 words | 6.4 words |
| sentences opening with a connective ("Moreover", "Furthermore") | 2.4% | 5.8% |
Where wrong flags land
Published by Turnitin about their own tool, and the clearest evidence available on how flags behave inside a document:
- 54% of falsely flagged sentences sit directly beside a sentence scored as AI written
- another 26% sit two sentences away
- documents scoring below 20% show a higher incidence of wrong flags
Four in five wrong flags cluster around a genuine detection rather than scattering, which means a flagged region is wider than whatever caused it.
Who gets flagged more
Published research, not our measurement: GPT detectors misclassify writing by non-native English speakers as AI generated at elevated rates, while native samples are classified correctly (Liang et al., 2023).
The mechanism is consistency. Learned English applies its rules more evenly and reaches for the common construction more often, and evenness is what these tools measure.
AI writing in the wild
unslop scored the full text of 12,750 arXiv papers from 2021 to 2026, calibrated to the same 0.4% false positive floor. About 32% of recent submissions read as machine written, rising across the period, with computer science highest and mathematics near zero.
Method and limitations: how much of arXiv reads as AI written.
What these numbers do not cover
Stated plainly, because a figure without its scope is not much use.
The false positive rate is measured on academic prose. Casual writing, fiction and technical documentation are different registers and the rate on them is not this rate.
Detection rates are measured on English. Nothing here describes performance in another language.
Every figure is document level. Per-sentence verdicts from any detector are substantially less reliable, which is the whole content of Turnitin's 4% against their own under-1%.
Using these figures
They are free to cite. The underlying method, including how the thresholds are set and what the corpus contains, is in how the unslop AI detector works.
Run your own text against the same detector these numbers describe: free and unlimited, no account.
Sources
- Turnitin, Understanding the false positive rate for sentences of our AI writing detection capability
- Turnitin, AI writing detection update from Turnitin's Chief Product Officer
- Turnitin, AI writing detection capabilities FAQ
- Liang, Yuksekgonul, Mao, Wu and Zou (2023), GPT detectors are biased against non-native English writers
- unslop, How the unslop AI detector works: our own figures, their sample sizes and how they were measured
- unslop, How much of arXiv reads as AI written: the 12,750-paper study