AI detectors do not read your text the way a teacher does. They measure two statistical properties — perplexity and burstiness — and use those numbers to decide whether a human or a machine wrote the words.
If you have ever passed a draft through a detector and been surprised by a high AI score, these two signals are almost certainly why.
What perplexity means in AI detection
Perplexity measures how surprising a passage of text is to a language model. As the model reads each word, it predicts what word comes next. If the actual word is almost always what the model expected — high probability, low surprise — perplexity is low. AI text scores low because language models are trained to choose the most likely continuation. That makes their output statistically smooth.
Human writers do something different. They reach for precise or unexpected words, draw on personal knowledge, and take small stylistic detours. That unpredictability raises perplexity, and detectors read that as a human signal.
Low perplexity is the single most consistent indicator that text was machine-generated, which is why swapping a handful of synonyms rarely moves the needle — the surrounding sentences are still too predictable.
What burstiness means in writing
Burstiness describes the rhythm created by sentence-length variation across a passage. Read almost any well-written human essay and you will notice a natural mix: a one-liner. Then two longer sentences that build on it. Then maybe something quite short again. That uneven cadence is high burstiness.
AI models tend to produce sentences of similar length, creating a flat, metronomic rhythm. Detectors measure the variance in sentence length across a passage. Low variance — the "flat" pattern — is a strong secondary indicator of AI authorship.
Burstiness is the signal most writers forget about. You can rewrite every sentence and still fail a detection check if the rhythm stays uniform.
Why both signals matter together
Perplexity operates at the micro level: individual word choices. Burstiness operates at the macro level: sentence and paragraph rhythm. A passage can have higher-than-average perplexity (unpredictable words) but still score as AI if the sentence lengths never vary. Detectors combine both signals — and often add their own proprietary models on top.
Fixing only one side of the problem is rarely enough, which is why structural rewriting outperforms vocabulary-only edits.
How AI text compares to human writing
| Property | AI-generated text | Human writing |
|---|---|---|
| Perplexity | Low — predictable word choices | Higher — specific or surprising words |
| Burstiness | Low — sentences roughly the same length | High — short and long sentences mixed |
| Transitions | Formulaic ("Furthermore," "It is worth noting") | Varied and context-specific |
| Detail level | General and abstract | Concrete, particular, sometimes personal |
Why common quick fixes don't work
Swapping synonyms changes the vocabulary but not the prediction probability of surrounding words. Adding a single short sentence breaks burstiness locally but leaves the rest of the passage flat. Invisible Unicode characters are stripped by modern detectors and can flag your work as tampering.
These surface edits leave the underlying statistical fingerprint almost unchanged. Detectors score the full passage, not individual sentences, so isolated fixes rarely produce a meaningful drop.
How humanizers address perplexity and burstiness
A structural humanizer works on the level that actually matters. It restructures sentences so rhythm varies naturally, replaces predictable phrasing with more specific human choices, and breaks up uniform paragraph flow. The result is a different statistical profile, not just different words.
Understanding what an AI humanizer actually does helps set realistic expectations: the goal is to change the patterns detectors measure, not to deceive anyone.
UnMarkedAI rewrites cadence and structure while preserving your facts and intent, then highlights which sentences still show strong AI patterns so you can focus your edits. Always run the result through a detector before you publish or submit — no tool can guarantee a clean score on every check, because detectors update their models regularly.
One thing worth knowing: even genuine human writing can score poorly on burstiness when the style is intentionally formal or measured — academic prose, legal briefs, technical documentation. That is a known limitation of detection tools, and it is one reason AI detection affects more than just AI-generated content.
Interactive FAQ
What is perplexity in AI detection?
Perplexity measures how predictable a piece of text is to a language model. AI-generated text scores low on perplexity because language models favor high-probability word sequences, making the output statistically smooth and expected. Detectors read that predictability as strong evidence of machine authorship.
What is burstiness in writing?
Burstiness describes the variance in sentence lengths across a passage. Human writers naturally mix short punchy sentences with longer ones, producing high burstiness. AI models tend to generate sentences of similar length, creating a flat rhythm that detectors flag as machine-generated.
Can I improve perplexity and burstiness manually?
Yes — vary sentence lengths deliberately, replace generic phrases with specific detail, and break predictable paragraph structure. A humanizer like UnMarkedAI automates most of this work and shows you exactly which sentences still read as AI, so you know where to focus your attention.
Do all AI detectors use these signals?
Most major detectors — GPTZero, Copyleaks, Originality.ai, Winston AI — use variations of perplexity and burstiness alongside their own proprietary models. Improving both properties tends to reduce scores across multiple tools, not just one, because the underlying patterns they measure overlap significantly.
Make your AI text sound human.
Paste your draft into UnMarkedAI, see which sentences look AI-generated, humanize them, and verify the result before you publish.
Knowing which signals drive your detection score is the first step — addressing both perplexity and burstiness together is what actually moves the number.