AI detection: what it is, how it works and what it cannot tell you
AI detection is the attempt to estimate whether a piece of text was written by a language model rather than a person. It is an estimate, never a proof. A detector reads the text, measures a set of statistical patterns and returns a probability or a score. Humanize 360 runs a free detector at /ai-detector that reports a Human Score out of 100 and lists the eleven signals behind it, so you can see why a text scored the way it did rather than just that it did.
There are two broad approaches. The statistical one measures things like word predictability, sentence-length variance and stock-phrase density, then combines them. The classifier approach shows a model a large pile of human and machine samples and lets it learn the difference on its own. Most commercial detectors blend the two. Neither approach reads meaning. They read texture, which is why a detector cannot tell you whether an essay is true, only whether it is shaped like something a model would produce.
What a detector actually measures
Vendors describe their methods in different words, but the underlying measurements overlap heavily. Ours are named openly:
- Sentence-length variance (burstiness) and word predictability (perplexity).
- Stock-phrase density, tell-word ratio and transition-word overuse.
- Opener repetition, paragraph uniformity and structural sameness.
- Punctuation habits such as em-dash and rule-of-three density.
- Contraction rate and vocabulary spread.
Why detectors disagree with each other
Each tool uses its own reference model, its own training samples and its own threshold for calling something "AI". Update any of those and the verdict on the same essay can move. This is why the same paragraph can read as human on one tool and machine on another in the same afternoon, and why no honest tool, ours included, promises what another one will say.
What AI detection cannot do
It cannot name the model that wrote a text. It cannot prove intent, because it cannot know whether a person wrote flat prose or a machine wrote polished prose. It is unreliable on short passages and on formulaic genres such as abstracts, lab methods and legal notices. It also flags careful non-native English more often than it should, because careful English is predictable English. Our page at /ai-detector-for-teachers goes into this in more detail.
How to read a detection result
Treat the score as an editor's note, not a verdict. Above 70 on our scale, leave the text alone. Between 45 and 70, open the breakdown and see which two or three signals are low. Below 45, the text carries the full set of tells and is worth a pass through the humanizer, after which you check again. If you are assessing someone else's writing, a score should never be the only evidence.
Common questions
How accurate is AI detection?
Nobody can give one number, because accuracy depends on the tool, the text length, the genre and the model that wrote the sample. All detectors produce false positives and false negatives. Treat any published accuracy figure as the vendor's own test on the vendor's own data.
Can AI detection be wrong about my own writing?
Yes. Formal, careful or heavily edited human writing is regularly flagged. That is a false positive, and it is the main reason we show signals instead of a bare percentage.
Does Humanize 360's detector predict Turnitin or GPTZero?
No, and no tool can. We measure the statistical patterns those tools also look for, so a high score means the common tells are gone. It is not a promise about another vendor's verdict.
Is the detector free?
Yes. 300 words per check, three checks a day, no sign-up. A free account adds 3,000 detector words a month.