What the research actually found
Two findings do most of the work here, and they point the same direction.
The first is a 2023 study by Weixin Liang and colleagues at Stanford, published in Patterns. They ran 91 TOEFL essays — written by real people sitting an English proficiency exam — through seven widely used GPT detectors. The detectors misclassified an average of 61.22% of them as AI-generated. Eighty-nine of the 91 were flagged by at least one detector. Eighteen were flagged by all seven. The same detectors handled essays by US 8th-graders with near-perfect accuracy.
The second is that OpenAI, the company that built the model most of this text is accused of coming from, could not reliably detect its own output. It launched an AI Text Classifier in January 2023 and quietly killed it on 20 July that year, citing a low rate of accuracy. The published numbers: it correctly identified about 26% of AI-written text and wrongly flagged 9% of human writing.
That second fact is worth sitting with. The organisation with the deepest possible access to how its own model generates text — the weights, the training data, the sampling behaviour — concluded that detecting it after the fact did not work well enough to ship.
Why detectors flag human writing
Nearly every detector works from the same two measurements, and understanding them explains most false positives.
Perplexity is how surprising each word is, given the words before it. Language models are trained to pick likely next words, so their output tends to be unsurprising — low perplexity. Burstiness is how much sentence length and complexity vary across a passage. Human writing tends to lurch: a long winding sentence, then a short one. Model output is more even.
So a detector is not asking "did a machine write this?" It is asking "is this text predictable and uniform?" Those are not the same question, and the gap between them is where real people get hurt.
Writing that is predictable and uniform for entirely human reasons includes:
- Second-language writing. A smaller working vocabulary and simpler sentence structures produce exactly the low-perplexity signature detectors punish. This is the single largest documented bias.
- Technical and scientific writing, where conventional phrasing is the point and deviation is penalised by the discipline itself.
- Formulaic academic structures — the five-paragraph essay taught in schools is close to an algorithm.
- Heavily edited prose. Grammar tools, style checkers and careful self-editing all reduce variance. Polishing your writing can make it look more like a machine wrote it.
What a detection score actually means
A score of "85% AI" does not mean 85% of the document was machine-written, and it does not mean there is an 85% chance you cheated. It means the text scored in a range the vendor has associated with AI output, based on their model and their threshold.
Two things follow. First, the number is only meaningful relative to that vendor's calibration — which is why the same paragraph routinely gets very different scores from different tools. Second, the base rate matters enormously. If a class of 300 students submits honest work and a detector has even a 5% false positive rate, fifteen innocent people get flagged. Run it across a university and the arithmetic gets ugly fast.
What we don't know
Being straight about the limits of the evidence above, because most pages on this topic are not:
- The Liang study tested detectors as they existed in 2023. Detection has moved since, and some vendors have specifically worked on the ESL bias. The direction of the finding is well established; the current magnitude for any specific tool is not.
- Vendor-published accuracy figures are marketing, ours included. Almost none are independently reproducible, because the test corpora are not released.
- There is no public, maintained benchmark that tests detectors against provably human, pre-2022 text across writer demographics. That gap is why this page leans on a three-year-old paper — it remains among the most rigorous work available.
If you have been wrongly flagged
The practical position is stronger than most people realise, because the burden of proof does not sit where an accusation implies. A detector score is one piece of weak statistical evidence, and every major vendor says in its own documentation that scores should not be used as the sole basis for an accusation.
Draft history is the most effective counter-evidence there is: Google Docs version history, Word's AutoSave revisions, commit timestamps, notes and outlines. If you are writing anything that could later be questioned, work somewhere that keeps a revision trail. It costs nothing and it settles the argument.
More on why text gets flagged and what to do about it.
Where HumanFlow sits
HumanFlow includes a detector, and everything above applies to it too. It gives you a sentence-level read of which passages look statistically machine-like, which is useful for spotting the parts of a draft that read as flat or formulaic. It is not proof of anything, and we do not claim it can tell you what another tool will say.
The humanizer exists to make writing sound more like a person wrote it. That often reduces false flags, because varied, specific prose is exactly what detectors read as human. We will not promise you a score, from us or from anyone else.