AI Content Detectors: Why the Score Is Not Proof
AI detectors measure perplexity and burstiness, not authorship. Over half of non-native English essays were wrongly flagged in one study.
A student is accused of cheating because a checker returned 94%. A freelancer loses a client over a number on a screen. A hiring manager bins an application because a tool said the cover letter was machine-written.
In every one of those cases the number came from a system that has no idea how the text was produced. AI detectors do not detect AI. They measure two statistical properties of writing and infer from them — and the inference is wrong often enough that treating the output as proof is indefensible.
Here is how they actually work, what the research says about their accuracy, and what to do instead.
AI Content Detector
Run a passage and see the signals a detector reads. Useful as a check on your own drafts — not as evidence about anyone else.
What a detector actually measures
Two things. Neither is “was this written by a machine”.
Perplexity is how surprising each word is given the words before it. Language models are built to choose likely continuations, so their output tends to be predictable — low perplexity. Human writing wanders more.
Burstiness is how much sentence length and structure vary across a passage. People write a long, winding sentence and then a short one. Models trend toward uniformity.
Low perplexity plus low burstiness produces a high “AI” score. That is the entire mechanism. It follows that any writing which is deliberately clear, consistent and simple will score as machine-made — because clarity and consistency are precisely what the metrics measure.
The research: over half of non-native essays misclassified
This is not a theoretical concern. A 2023 study published in Patterns by researchers at Stanford tested seven widely used GPT detectors against 91 essays written by non-native English speakers for the TOEFL exam.
The finding. More than half of the non-native English essays were incorrectly flagged as AI-generated. Against essays by US eighth-graders, the same detectors were near-perfect. The bias was not marginal — it was the difference between a working tool and a broken one, depending entirely on who wrote the text.
The researchers went further and showed the mechanism directly. When they used a language model to enrich the vocabulary of the non-native essays, misclassification dropped sharply. When they simplified the native-speaker essays, misclassification rose.
In other words the detectors were not identifying AI. They were identifying limited linguistic variety, and labelling it machine-generated. Anyone writing in a second language is penalised by design.
OpenAI reached a similar conclusion about its own tool. It launched a classifier for AI-written text in early 2023 and withdrew it within months, citing a low rate of accuracy.
Who gets falsely flagged
- Non-native English speakers — the largest and best-documented group.
- Technical writers — documentation is meant to be unambiguous and consistent. Varied phrasing would be a fault.
- Students following a taught structure — five-paragraph essays score as formulaic because they are formulaic, by instruction.
- Anyone writing to a house style — legal, medical and academic conventions all reduce variability deliberately.
- Autistic writers and others with a naturally consistent register, whose prose may simply not vary in the way the metric expects.
The reverse holds too. AI output that a person has edited, or that was generated from a prompt asking for varied sentence length and a specific voice, routinely scores as human. A tool that can be defeated by asking nicely is not a reliable detector.
What a percentage actually means
It looks precise. It is not. It is a confidence estimate from a model that never saw the text being written and has no access to the truth.
There is no threshold at which a score becomes evidence. 94% is not “almost certainly AI” — it is “this text has the statistical shape the model associates with AI”, which is a different statement and a much weaker one.
If you are assessing someone else’s work. Do not use a detector score as evidence of misconduct. Given the documented false-positive rate on non-native writing, doing so is not just unreliable — it discriminates in practice against the students and applicants least able to challenge it.
Better questions than “is this AI?”
The useful question was never how text was produced. It is whether the text is any good, and those are separable.
- Is anything in it specific? A number, a name, a date, a first-hand observation. Generic writing is the actual problem, whoever produced it.
- Does it cite sources, and do they exist? Fabricated citations are the single strongest signal of unedited AI output — and unlike a perplexity score, you can verify a citation.
- Is anything factually wrong? Confident, plausible, incorrect detail is the characteristic failure mode.
- Does it commit to a position? Models hedge. Text that carefully avoids saying anything is a signal in itself.
- Could the author defend it? For coursework, a two-minute conversation reveals more than any tool.
Where Google stands
Google has been explicit: it does not penalise content for being AI-assisted. It penalises content that is unhelpful, unoriginal, and produced primarily to rank rather than to serve a reader.
That distinction matters for anyone publishing. There is no ranking benefit in passing a detector, and no penalty for failing one. What is penalised — under the spam policies covering scaled content abuse — is mass-producing pages that add nothing, which describes a great deal of unedited AI output and also a great deal of human content-farm writing.
The practical implication: stop optimising for the detector and start adding what a model cannot. A tested example. A number you looked up. An opinion you will defend.
Does rewriting help?
Running AI text through a paraphrasing tool to lower a score usually makes the writing worse, not more human. Naive rewriting flattens rhythm and reaches for formal synonyms, which is exactly the profile detectors flag.
If the goal is writing that reads as a person wrote it, the reliable route is unglamorous: shorter sentences next to longer ones, a concrete example, an aside that only someone who had done the thing would include.
Questions worth asking first
Are any AI detectors accurate?
All of them are better than chance on long, unedited, purely machine-written text. None is reliable on short passages, edited text, or writing by non-native speakers. Accuracy claims from vendors are typically measured on data that resembles neither.
Can I prove I wrote something myself?
Version history is the strongest evidence available. Google Docs and Word both retain it, and a document that grew over days with edits and false starts is far more persuasive than any score. If you are a student, keep drafts.
Why do two detectors disagree?
Different training data and different thresholds. Wide disagreement on the same passage is itself informative — it tells you the text sits in the region where none of them knows.
Should I use one at all?
On your own drafts, as one signal among several: yes, it can flag a section that reads flat. On other people’s work, as a basis for a decision about them: no.