Five experienced academics were handed writing by human authors and asked one question: did a machine write this? On the genuinely human text, they got it wrong 72% of the time. Not on the AI samples. On the real ones. They read writing by real people and called it machine-made, nearly three times in four.
The short version: AI detectors are accurate at the extremes and unreliable in between, and the published numbers depend on who did the testing. Turnitin reports under 1% false positives on documents with 20% or more AI writing; OpenAI retired its own detector after it caught 26% of AI text. Free AI detectors are better at this than people are, and neither is good enough to accuse anyone with. Human raters identified AI text 19% of the time, which is what guessing looks like. Two of three tested detectors tracked AI use closely and agreed with each other at an ICC of 0.96. But those same detectors read non-native English as machine writing at a 61% false-positive rate. If someone has accused you of using AI, the evidence they are holding is weaker than they think it is.
How do AI detectors work?
An AI detector estimates how likely a text is to be machine-written by measuring how predictable it is. Most tools combine a classifier trained on known human and AI samples with two statistics: perplexity, how surprising each next word is, and burstiness, how much sentence length and structure vary. Low surprise and even rhythm read as machine.
A third method does not guess at all: watermarking. A model can tilt its word choices in a pattern only a matching checker can see, which is what Google's SynthID does for text from Google's own models. A watermark proves origin when it is present, but it only covers text from a model that applied one, and it says nothing about text from any other model. Every detector you can paste text into today is doing the first kind of work, and that is the kind this article tests.
How accurate are AI detectors, by their makers' own numbers?
Better on the makers' own tests than on anyone else's. A vendor reports accuracy at the threshold and on the documents it chooses; independent researchers pick the hard cases. Read the two columns together:
| Source | Who tested | What they reported |
|---|---|---|
| OpenAI AI classifier (2023) | OpenAI, on its own model | Caught 26% of AI-written text; flagged human text 9% of the time. Withdrawn on July 20, 2023 "due to its low rate of accuracy". |
| Turnitin | Turnitin, on its own product | Under 1% false positives for documents with 20% or more AI writing; about 4% per highlighted sentence. |
| Liang et al., Patterns (2023) | Stanford researchers, seven detectors | 61.22% average false positives on essays by non-native English writers. |
| Healthcare-simulation study, PMC (2026) | Independent researchers, three free tools and five academics | Tools tracked AI use closely; human readers called human text AI 72.2% of the time. |
The vendor figures are not false. They are measured where detection is easiest: long documents with plenty of AI text. The qualifier matters. Turnitin's own figure applies only above 20% AI writing, and its own sentence-level rate is about 4%. The independent studies, covered below, measured what happens to the writers a detector is most likely to get wrong.
Are AI detectors accurate?
Partly, and the part that works is narrower than the marketing suggests. At the extremes they are good. In a 2026 study of healthcare-simulation articles, ZeroGPT scored genuinely human introductions 6.5 out of 100 and fully AI-written ones 92.5. PhraslyAI scored 5.9 and 92.4. Both moved in the right direction through every stage of AI involvement. The problem is not the extremes. It is whose human writing you hand them.
The study, published in PMC, took 30 open-access articles from two healthcare simulation journals and pulled their introductions — 500 to 600 words each, all written before 2022, so none of them could have been AI-assisted. Then it produced five versions of each: untouched human, lightly AI-edited, heavily AI-edited, AI-written from human bullet points, and fully AI-written from the title alone. Three free detectors scored every version.
| Condition | ZeroGPT | PhraslyAI | Grammarly |
|---|---|---|---|
| 100% human | 6.5 | 5.9 | 1.6 |
| Light AI edit | 20.2 | 24.8 | 3.0 |
| Heavy AI edit | 43.1 | 45.1 | 11.5 |
| AI from bullet points | 89.9 | 85.6 | 62.5 |
| 100% AI | 92.5 | 92.4 | 50.0 |
Mean AI score out of 100, by condition and tool.
Read the bottom two rows of the Grammarly column. It scored AI-from-bullets at 62.5 and fully AI text at 50.0 — it rated the more machine-written text as less machine-written. That single inversion is what drags the cross-tool agreement down: ZeroGPT and PhraslyAI agree with each other at ICC 0.96, but each agrees with Grammarly at only 0.57 to 0.60. So "AI detectors disagree with each other" is not quite the finding. The finding is that two of these three behaved consistently and one did not, and from the outside you cannot tell which one you are using.
How well can a person spot AI writing?
They cannot. Across all five conditions the human raters were right 19% of the time, which the authors describe as indistinguishable from chance. Broken down: 17% on untouched human text, 23% on light edits, 30% on heavy edits, 17% on AI from bullets, and 10% on fully AI-written text. They were at their worst on the samples that were entirely machine-written.
The error rates are the part worth sitting with. On text written entirely by humans, the raters produced a 72.2% false positive rate. On text written entirely by AI, a 76.9% false negative rate. And the agreement between the human raters and each of the three tools was ICC = 0 — not weak agreement, none.
These were not random readers. They were healthcare simulation educators and researchers with between one and 25 years of experience, reading articles from their own field, where they would know the conventions better than almost anyone. That is the ceiling on human judgement here, not the floor.
Why detectors flag writers whose first language is not English
Because they are largely measuring how predictable your word choices are, and writing in a second language makes your word choices more predictable. In 2023, Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou ran seven widely used GPT detectors over 91 TOEFL essays by non-native English writers and 88 essays by US eighth-graders. On the eighth-graders, the detectors were close to perfect. On the TOEFL essays, the average false-positive rate was 61.22%, and all seven detectors unanimously flagged 18 of the 91 essays — 19.78% — as machine-written.
The mechanism is perplexity: a measure of how surprising each next word is. Someone writing in a language they learned second reaches for the common, safe word, because that is what working carefully in a second language looks like. Low surprise reads as machine.
The researchers then proved it was style and not substance. They had ChatGPT rewrite the same TOEFL essays with richer vocabulary — same authors, same arguments, same essays — and the average false-positive rate fell from 61.3% to 11.6%. Nothing about the authorship changed. Only the vocabulary did. Their paper, "GPT detectors are biased against non-native English writers", ran in Patterns, and it ends by arguing these tools should not be used in evaluative or educational settings at all.
I have a personal stake in this one. English is not the language I think in first. When a tool scores my writing as machine-made, I have no way to know whether it is detecting a machine or detecting me.
What to do if you have been accused of using AI
Ask four questions, in this order.
- Which tool, and what score? A screenshot of a percentage is not a finding. Different tools disagree with each other on the same text, and one of the three tested above rated fully AI writing as less machine-like than partly AI writing.
- What is that tool's false positive rate? If whoever is accusing you cannot answer, they are reading a number they do not understand. A tool that is right about AI text and wrong about a large share of human text is not evidence against a specific person.
- Does the accusation account for how I write? If you write in a second language, the published research says the tool is biased against you by a measured, published margin. That belongs in the conversation.
- Can I show process instead of prose? Drafts, version history, research notes, timestamps, the messy middle. Process is checkable. A style score is not.
One thing not to do: rewrite your work to beat the tool. Raising your perplexity on purpose is exactly the bypass the Stanford team demonstrated, and it works — which is the clearest proof available that the tool is scoring vocabulary rather than authorship. You would be making your writing worse to satisfy an instrument that has already been shown not to measure the thing it claims to.
How we use AI detectors at Skryvo
As a delta, never as a verdict. A draft that scores 95% before an editing pass and 30% after tells you the pass landed. The same score read as a judgement on a finished page tells you nothing you can act on, and chasing a lower score is its own trap. What makes AI writing recognisable is a set of countable habits — uniform sentence rhythm, watermark vocabulary, symmetrical structure — and those are worth fixing because they make writing worse, not because a detector objects to them.
The check we actually gate on is whether the claims are true, because most new pages are already a mix of AI and human writing. When we ran 20 AI-written posts through our fact-check pass, six in ten checkable claims could not be traced to any source. That question has a right answer you can go and verify. "Did a machine write this" does not, and two studies now say nobody — person or tool — can answer it reliably enough to act on. Google's own position points the same way: its guidance rewards people-first content however it was produced, and the signals that do move rankings are the author E-E-A-T signals behind AI-assisted content that ranks.
What this evidence does not prove
Both studies are small, and we would want this stated by anyone quoting them at us.
- The detection study covered 30 articles from a single field, scored by five raters using three free tools, with text generated by one model over four weeks in 2025. It is a signal, not a law.
- The bias study dates from 2023 and tested the detector generation available then. Detectors have changed since; whether the bias has is not something either paper can tell you.
- Neither tested the paid institutional detectors used in universities and newsrooms, which are the ones most accusations actually come from. The only figures for those come from the vendors themselves.
- TOEFL essays are a specific genre written under exam conditions. They are not a stand-in for all non-native English writing.
What survives all of that: no published evidence supports using any of these scores to accuse a named person, and a good deal of published evidence says the people most likely to be wrongly accused are the ones writing in their second language.
The check that actually works
Stop asking whether a machine wrote it. Ask whether it is true. Run your own draft through a fact-check pass and see how many of its claims survive contact with a primary source — that number is real, it is yours, and unlike a detector score, you can do something about it. It is free to try and there is no card required. The broader question is whether a page is slop at all — and the four-signal test for AI slop checks what is absent, not how it was typed.
First published on skryvo.ai.
Top comments (0)