How AI Detection Actually Works in 2026: Perplexity, Burstiness, and What Detectors Really Measure

Most advice about bypassing AI detectors is wrong. Not slightly off — completely wrong. "Vary your sentence length" and "add transition words" are the kind of tips that sound reasonable but don't actually move the needle on real detection tools.

If you want to understand why your text keeps getting flagged, you need to understand what detectors are actually measuring. Not the marketing copy. The actual math.

The Two Numbers That Matter

Every major AI detector — GPTZero, Turnitin's AI writing indicator, Originality.ai, Copyleaks — is fundamentally measuring two things:

Perplexity. How surprising is each word choice, given the words before it?

Burstiness. How much does sentence length and complexity vary throughout the piece?

That's it. Everything else is a variation on these two signals.

When a language model like GPT-4 or Claude writes something, it picks the statistically most likely next word at each step. The result is text with very low perplexity — it's predictable at the word level in a way human writing never is.

Humans don't write like that. We reach for the slightly unusual word. We get distracted mid-sentence. We start a paragraph one way and end it another. That unpredictability is what perplexity measures, and it's what AI writing lacks.

Why "Vary Your Sentences" Doesn't Work

Here's the thing about burstiness: it's not just about sentence length. It's about the variance in complexity.

AI-generated text often has uniform grammatical structure even when sentence lengths differ. A short sentence and a long sentence can both be written at the same complexity level — same clause nesting, same passive/active ratio, same reliance on coordinating conjunctions. Detectors measure all of that, not just word count.

Real human writing has genuine unpredictability at the structural level. A person might write three short punchy sentences, then a long winding one that backtracks on itself, then a fragment. That pattern is hard to fake by just counting words.

What Turnitin Measures in 2026

Turnitin rolled out a major update to its AI detection in early 2026. The key change was adding what they call "writing style consistency" scoring alongside the base perplexity analysis.

What this means practically: Turnitin now looks at whether the style of your writing stays consistent across a document in a way that's characteristic of AI output. Human writers drift — they get tired, they shift formality levels, they have pet phrases that appear and then disappear. AI writers don't.

The update also specifically targets humanized text. Their training data now includes examples of AI text that has been run through humanizing tools, so the surface-level randomization that older tools produced gets flagged.

This is why tools that just swap synonyms or shuffle sentences don't work against Turnitin anymore. The underlying statistical signature survives those kinds of edits.

The Structures AI Loves (and Detectors Recognize)

Beyond the statistical measures, trained detectors — and increasingly human reviewers — look for specific structural patterns that AI produces at a much higher rate than humans:

The three-point structure. AI defaults to introducing a topic, making three parallel points, and summarizing. Humans rarely do this unless they're filling out a template.

Passive voice clusters. "It has been shown that..." / "This can be understood as..." / "The results were found to be..." — AI uses passive constructions at roughly 3x the rate of human writers in informal and semi-formal contexts.

Topic sentence loyalty. Every AI paragraph starts with a clear topic sentence. Human paragraphs frequently start mid-thought, return to a previous idea, or begin with a question.

The "Furthermore" family. Furthermore, Moreover, Additionally, Consequently, Subsequently. These transitional adverbs appear in AI text at rates that no human writing corpus matches.

Perfect parallel structure. When AI lists three things, they're always grammatically identical. When humans list three things, the third one usually breaks the pattern slightly.

What Actually Moves Your Score

If the above is the problem, here's what the solution actually looks like:

Attack the n-gram probability. The sequences of 3-5 words that appear in your text need to not match common AI output patterns. This means rephasing at the clause level, not just swapping words. "The implementation of this approach" needs to become something like "rolling this out" or "actually doing it."

Break the register consistency. Human writing shifts formality mid-document. A little more casual in one paragraph, a bit more precise in another. AI maintains one level throughout.

Use grammatical variety that AI avoids. Fragments. Sentences starting with conjunctions. Parenthetical asides. Self-corrections ("or rather," / "well, more precisely"). Trailing thoughts.

Kill the passive voice clusters. Active voice doesn't just read better — it statistically distinguishes human writing from AI output in every major detector's training data.

The Practical Takeaway

The detectors are getting smarter. Turnitin's 2026 update specifically trained against humanized text, which means the old tricks — synonym replacement, sentence shuffling, basic paraphrasing — don't work.

What works is rewriting at the structural level. Not editing the words. Rebuilding the architecture of the prose so that the statistical fingerprint of the original AI output is gone.

That's what UnChat's two-pass humanization system does. The first pass destroys the perplexity signature. The second pass acts as a Turnitin auditor, hunting down the structural patterns that survived — passive constructions, parallel clauses, topic-sentence-led paragraphs — and eliminating them.

The result isn't AI text that's been lightly disguised. It's text that has a genuinely different statistical profile.