How AI detectors work, and where their scores get unreliable

How AI detectors work, at a high level

Most detectors combine three signals: perplexity, burstiness, and n-gram fingerprints. None of these are exotic. They are basic measurements applied to text, and each has been used in computational linguistics for decades.

Perplexity

Perplexity asks: how surprising is the next word given the previous ones? AI output tends to pick high-probability next words because that is literally what the model is optimizing for. Human writers occasionally pick a less likely word, which raises perplexity. Low average perplexity is a signal that text might be AI-generated. See our perplexity glossary entry for the longer version.

Burstiness

Burstiness measures variation in sentence length and structure. Human writing tends to be bursty: short fragments next to long flowing sentences, with occasional one-word lines for effect. AI writing tends to be smooth, with sentences clustering around a similar length. Low burstiness is another signal. The burstiness glossary entry has examples.

N-gram fingerprints

Some detectors look for specific multi-word phrases that appear unusually often in AI output. Phrases like "it is important to note" or "navigate the complex landscape" appear in AI writing at rates many times higher than in pre-2023 human writing. Detecting their density is part of the signal too. The AI detector glossary entry covers the broader category.

Why detector scores can mislead

The signals above are real, but the conclusions drawn from them often are not. Three failure modes show up consistently.

False positives on edited human writing

Professional editors strip variability. They cut unusual word choices, regularize sentence length, and impose a house style. The output is the kind of clean, controlled prose detectors are trained to flag as AI. A newspaper op-ed that went through three editors can score higher for AI likelihood than a first-draft AI output. Our learn post on AI detector false positives has concrete examples.

ESL bias

Writers whose first language is not English are flagged at higher rates than native speakers. The 2023 Stanford study found false-positive rates above 60 percent for some ESL essay sets, against near-zero rates for matched native-speaker essays. The reason appears to be that careful ESL writing uses a more restricted vocabulary and more regular sentence structure, both of which look AI-like to a detector.

Base-rate fallacy

Even a detector that is 95 percent accurate produces a lot of wrong calls in absolute terms. If one in twenty papers in a class is genuinely AI-generated, and the detector is 95 percent accurate on both classes, the number of false positives roughly equals the number of true positives. The score on a single paper does not tell you which group it falls into.

How to write less robotically

Detector scores are unreliable. The underlying signals they measure are not random though. If your writing reads flat or generic, that is real, and worth fixing for its own sake. Here is what actually moves the needle.

• •Vary sentence length. Mix four-word fragments with twenty-word sentences. Burstiness rises and the prose reads more alive.
• •Pick one specific fact per paragraph. One number, one place name, one named person. Specificity is the fastest path away from AI prose.
• •Cut signposting words. Furthermore, additionally, moreover. If the connection needs a bridge word, the argument needs work.
• •Drop the rule of three. AI writing forces ideas into three-item lists. Pick two. Or four. Or one.
• •Allow some mess. A side comment, a half-finished thought, a sentence that runs on. Perfectly even prose reads algorithmic.
• •Use I when it fits. First person reads honest. "I keep coming back to" beats "It is worth considering."
• •Read it out loud. If you would not say it that way to a person at coffee, do not write it that way for one.

Where Humanize Me fits

Humanize Me is a tool for cleaner rewriting, not a tool for fooling detectors. We rewrite the patterns that make AI output recognizable to careful human readers: adjective stacks, rule-of-three lists, signposting, abstract framing, polished closers. The result reads more like a person wrote it, which is the same direction detector signals point. We do not promise any specific detector score, and we will not.

We are also not a complete substitute for editing. The rewrite is a first pass. Your judgment about what a sentence should claim, what tone fits the audience, and what claims need a citation is still on you. The tool gives you a cleaner draft to work from.

If the writing-improvement framing fits your job, the use-case pages at /use-cases cover specific applications: email, essay, LinkedIn, product copy, and more. The comparison page at /compare covers how Humanize Me stacks up against other tools in the category.

Frequently asked questions