A student in my extended circle was accused of using AI on an essay she wrote by hand, on paper, before typing it up. The detector said 89% AI. She had no way to prove otherwise.
That situation is more common than the tools admit, and understanding why requires understanding what these detectors actually measure — which is not what most people assume.
What detectors actually measure
They do not detect AI. They measure two statistical properties of text and infer from there.
| Property | What it means | What scores “AI” |
|---|---|---|
| Perplexity | How predictable each word is, given the ones before it | Low — the obvious next word, every time |
| Burstiness | How much sentence length and complexity vary | Low — sentences of similar length and rhythm |
Language models pick likely words, so their output tends to be predictable and evenly paced. Detectors look for that signature.
Who gets falsely flagged most
The false positives are not random. They cluster:
- Non-native English speakers. Writers working in a second language use more common vocabulary and simpler constructions. That is exactly the low-perplexity signature. Multiple studies have found this bias, and it is the most serious problem with these tools.
- Technical and academic writing. Formulaic by design, standard phrasing, consistent structure.
- Anyone who edits heavily. Polishing removes the very irregularity detectors treat as proof of humanity.
Why you cannot appeal with a number
Detectors output a confidence percentage. That number feels like evidence. It is not — it is a probability estimate from a model that has never seen your writing before and does not know your voice.
OpenAI shut down its own AI detector in 2023, citing low accuracy. The company with the most to gain from a working detector could not build one.
If you are accused
- Produce process evidence. Version history in Google Docs or Word, drafts, notes, browser history. This is the only defence that works, which is why you build it before you need it.
- Ask which detector and what threshold. Different tools disagree wildly on the same text. Run the accuser’s own text through it — the result is often illuminating.
- Cite the false positive rate. No vendor claims 100% accuracy. Ask what theirs is and whether a single number is sufficient grounds.
- Offer to be re-examined. Write something comparable under supervision.
If you are the one checking
Do not use a detector score as a verdict. Use it as, at most, a prompt to have a conversation. Ask the writer to explain their argument, walk through their sources, or expand a section. Someone who wrote it can. Someone who did not, usually cannot.
That conversation takes five minutes and is more reliable than any tool on the market.
Does “humanising” text help?
Tools that add typos and vary sentence length to defeat detectors do work, in the narrow sense that scores drop. They also make the writing worse. You are degrading your work to satisfy a broken measurement — which tells you the measurement is the problem.
Common questions
Which AI detector is most accurate?
None are reliable enough for consequential decisions. They vary in false positive rate, but every one of them will flag some human writing, and none can prove authorship either way.
Can Google detect and penalise AI content?
Google has said explicitly that it rewards helpful content regardless of how it was produced, and penalises low-quality content regardless of how it was produced. Method is not the criterion; quality is.
How do I protect myself as a writer?
Write in a tool that keeps version history and leave it on. That single habit is the most effective protection available, and it costs nothing.
