AI content detectors claim to tell you whether text was written by a machine. They are enormously popular and considerably less reliable than their confident percentage scores suggest.
Here is how they work, why they fail, and what to do if you are accused.
How they work
They do not detect AI. They measure statistical properties of the text and infer from them.
Perplexity — how predictable each word is given the ones before it. AI text tends to be more predictable, because models are literally optimised to produce likely continuations.
Burstiness — variation in sentence length and complexity. Human writing varies more; it has short sentences, long ones, fragments. AI output is more even.
So a detector is really asking: does this read as unusually smooth and predictable?
That is a proxy, and the gap between the proxy and the thing you actually want to know is where all the problems live.
Why they get it wrong
False positives on clear, careful writing. Anyone taught to write plainly — short sentences, consistent structure, no flourishes — scores as machine-like. Good technical writing is the most flagged human writing there is.
Systematic bias against non-native English speakers. This is the most serious problem. Writers using a smaller, more careful vocabulary produce exactly the low-perplexity signature detectors look for. Multiple studies have found substantial false-positive rates on non-native writing.
Trivially defeated. Light editing, asking the model to vary sentence length, or running text through a paraphraser drops scores sharply. The people deliberately evading detection are the least likely to be caught.
No ground truth. Detectors cannot be properly validated at scale, because nobody can build a large, verified corpus of human text that is guaranteed AI-free post-2022.
Short text is guesswork. Under a few hundred words there is not enough signal for the statistics to mean anything.
The percentage is not a probability. "87% AI" does not mean an 87% chance it was AI-written. It is a score on an internal scale, presented in a format that implies far more precision than exists.
What this means in practice
If you run a school or a business: a detector score is not evidence. Treat it as a prompt to have a conversation, never as a finding. Ask the person to discuss their work, show drafts, or explain their reasoning — those reveal far more than any score.
If you are accused: keep your drafts and version history. Document editing history in Google Docs or Word shows the work being built over time and is far more persuasive than arguing with a percentage. Ask what the detector's false positive rate is on writing like yours — most people using them cannot answer.
If you publish content: Google does not use these detectors, and does not penalise AI-assisted content as such. It penalises unhelpful content. A useful page is fine regardless of how it was drafted; a thin, mass-produced page is a problem whether a human or a model wrote it.
Does Google detect AI content?
Worth stating separately because it drives so much anxiety.
Google's stated position is that it rewards helpful content regardless of how it is produced. What it targets is scaled content abuse — mass-produced pages made primarily to rank rather than to help.
The practical distinction is not human-versus-AI. It is whether the page is worth reading. A well-researched AI-assisted article is fine. Two hundred templated pages published in a week are not, whoever wrote them.
Making AI-assisted writing genuinely yours
Not to defeat detectors — to make the writing better, which happens to have the same effect:
- Write the opening and closing yourself. They carry most of the voice.
- Add specifics only you have — your data, your example, your customer.
- Vary sentence length deliberately. Read it aloud; anything you would not say, cut.
- Cut the stock constructions. Tricolons, "in today's landscape", "it's not just X, it's Y", and closing paragraphs that restate the opening.
- Take a position. Generic text hedges. Say what you actually think.
Common questions
Are AI detectors accurate? Not reliably. They produce false positives on clear writing and on non-native English, and they miss lightly edited AI text.
Can AI detectors be wrong? Frequently, in both directions. The percentage score implies a precision the underlying method does not have.
Do AI detectors discriminate against non-native speakers? Studies have found substantially higher false-positive rates on non-native English writing. It is the most serious known problem with them.
How do detectors work? By measuring how predictable and how uniform the text is, then inferring. They do not detect AI directly.
Can you beat an AI detector? Easily, with light editing or paraphrasing — which is why they catch careless users rather than deliberate ones.
Does Google penalise AI content? Google targets unhelpful, mass-produced content, not AI assistance as such.
What should I do if I'm falsely accused? Produce your drafts and document version history, and ask what the tool's false positive rate is on writing like yours.