How AI Content Detectors Work (and Why They Get It Wrong)

AI detectors have become gatekeepers: teachers run essays through them, editors screen freelance submissions, platforms flag posts. Yet almost nobody using them can explain what the score actually measures — which leads to real harm when a false positive costs a student a grade or a writer a client. Understanding how these tools work makes you a much smarter consumer of their output.

Perplexity: how predictable is each word?

The core signal is perplexity — a measure of how surprising a text is to a language model. Large language models are prediction engines: given the words so far, they assign probabilities to what comes next. AI-generated text tends to follow the model's own highest-probability paths, because that's literally how it was produced. Human writing wanders: odd metaphors, abrupt topic shifts, oddly specific references. Low perplexity = predictable = AI-ish. High perplexity = surprising = human-ish.

An example helps. After "The chef seasoned the soup with...", a model strongly expects "salt" or "pepper". After "The chef seasoned the soup with..." a human might write "a childhood memory of her grandmother's kitchen." Statistically weird, vividly human. Detectors average these surprises across every sentence of your text.

Burstiness: the rhythm of human writing

The second signal is burstiness — variation in sentence length and structure. Humans write in bursts: a long, winding sentence full of clauses, then a short one. Like this. AI models, unless heavily prompted otherwise, produce eerily uniform paragraphs: sentence lengths cluster around the same range, the same connective phrases recur ("furthermore", "in addition", "it's important to note"), and each paragraph marches along at the same tempo. Flat burstiness is a statistical fingerprint even when the content itself is fine.

Why false positives happen

Here's the uncomfortable part: some legitimately human writing is low-perplexity and low-burstiness. Technical documentation, legal boilerplate, academic writing in a second language, formulaic corporate reports — all constrained, all predictable, all frequently flagged. The US Constitution famously scores as AI-generated on some detectors. Non-native English speakers, who naturally choose common words and regular structures, get flagged at higher rates — a documented bias with real consequences in classrooms.

• Writing in a formal, templated genre (legal, technical, academic)
• Writing in a non-native language with a simpler vocabulary
• Editing your own draft heavily until it reads 'smooth'
• Texts that are heavily summarized or paraphrased — summarization strips burstiness

And false negatives too

It cuts both ways. AI text that's been edited by a human, generated with high randomness settings, or prompted to vary its rhythm can sail past detectors completely. The arms race is asymmetric: generators improve monthly, while detectors chase statistical shadows of the previous year's models. Independent evaluations routinely find accuracy numbers far below the marketing claims, especially on shorter texts.

How to use detectors responsibly

Treat a detector score like a smoke alarm, not a verdict: it tells you where to look, never what happened. Run longer texts (100+ words — statistical signals need data), look at the per-sentence breakdown rather than the single headline number, and never use a detector score as sole evidence in any high-stakes decision. Our AI Text Detector shows exactly which signals fired — repetition, uniformity, predictable phrasing — so you can judge for yourself, and it runs entirely in your browser so unpublished drafts stay private.

If you're falsely flagged

Keep your draft history — versioned documents with timestamps are your strongest defense, since AI-generated text typically appears whole. Point out that detector vendors themselves warn against punitive use of scores. And if you habitually write formally, vary your rhythms: mix short punchy sentences with longer ones, choose the precise word over the expected one. Ironically, writing more distinctly is the best way to prove you wrote it — with or without a detector.