AI detectors

They return a number that looks like a verdict and is not one. This is what the number is calculated from, why the same paragraph scores differently in two detectors, why ordinary human writing gets flagged, and what to do if yours is.

What they measure

No detector can see where a passage came from. Every one of them measures properties of the text and infers backwards, and there are three families of measurement in use.

Predictability

Feed the text to a language model, word by word, and ask how surprised it is by each one. Text where every word was the obvious next word scores as highly predictable. This is usually reported as perplexity: low perplexity means the model found the text easy to predict.

The reasoning is that a model generating text tends to pick likely words, so its output is unusually predictable to another model. The reasoning holds on average and fails on individuals — clear, plain, conventional human writing is also predictable, which is the root of most false positives.

Burstiness

How much the text varies. Human writing lurches: a four-word sentence beside a forty-word one, a dense paragraph beside a thin one, predictable stretches beside surprising ones. Generated writing tends to hold a steadier rhythm. Burstiness is the variance of those measures rather than their average, and it is the signal that survives paraphrasing best.

Stylometry

Countable surface features: the density of em dashes and curly quotes, how often a paragraph opens with a connective, how often lists have exactly three items, the rate of particular stock phrases. Cheap to compute, easy to explain, and the weakest of the three on its own — every individual feature is something people legitimately do.

Why two detectors disagree

Because they are measuring different things, calibrated on different corpora, with thresholds chosen by different people. A predictability detector using one model as its yardstick will disagree with one using another. A stylometric detector trained mostly on student essays will read a technical document strangely. There is no shared definition of what score means what.

So the honest way to read any of these numbers is as one instrument's reading, not a verdict — the same way you would read a single blood-pressure measurement rather than a diagnosis.

Why human writing gets flagged

False positives are not rare and they are not random. They cluster:

• Writers using English as a second language. A smaller working vocabulary and more conventional sentence construction both read as predictable. This is the best-documented bias in the field and it is large.
• Formal and technical registers. A standards document, a legal summary or a lab report is supposed to be uniform, unsurprising and free of personality. All three read as machine-like.
• Anything heavily edited. Editing removes the lurching. A piece that has been through three drafts and a sub-editor has had exactly the irregularity a detector looks for smoothed out of it.
• Short passages. Every one of these measures needs enough text to be stable. Under a couple of hundred words the numbers move on noise.

If your own writing is flagged

Rewriting it to score better is the wrong first move — it makes the writing worse and concedes the premise. What actually helps:

• Keep the evidence you already have. Version history in a document, drafts, notes, commit timestamps. Process evidence is far stronger than any counter-score, and it exists whether or not anybody asks for it.
• Ask what the number was. "Flagged" is not a result. Which detector, what score, against what threshold, on how many words.
• Ask for the false-positive rate. A tool that flags one in twenty genuine pieces has flagged a great many innocent people in any large class, and its vendors generally publish a figure.
• Run a control. Something you certainly wrote, of similar length and register, through the same detector. If it flags that too, the reading is about your register rather than about this piece.

Where this site stands

humanaze reports its own reading of how patterned a passage is, and it is careful about the words it uses: it says a passage reads like model output, never that it was written by one. The score is a heuristic, it is labelled as experimental in the interface, and it is explicitly not evidence of authorship. Everything behind it is a countable property of the text — sentence-length variance, phrase densities, vocabulary spread — which is why the report lists what it found rather than only handing you a number.

This site is also not a way around any of the above, and does not present itself as one. Rewriting changes the patterns because the patterns are worse writing; that is the entire case for it. If a piece of work has to be yours, having a model rewrite it does not make it yours, and no software changes that.

Common questions