You paste your text into an AI detector. It gives you a score. You have no idea what that score actually measures or why it landed where it did. So you try another detector — and get a completely different number.
This is not random. Every major AI detector is measuring the same two underlying properties of text: perplexity and burstiness. They just weight them differently, train on different datasets, and express the results on different scales. Once you understand what these two signals actually are, the whole confusing landscape of AI detection starts to make sense.
What perplexity actually means
Perplexity measures how surprised a language model would be by each word choice. Low perplexity = the words were predictable. High perplexity = the words were surprising. AI models write with low perplexity because they are literally trained to predict the most likely next word. Humans write with higher perplexity because we reach for unexpected words, unusual analogies, and idiosyncratic phrasing.
The term comes from information theory. In that context, perplexity quantifies how well a probability model predicts a sample. A model that perfectly predicts every word has a perplexity of 1. A model that is completely lost has perplexity approaching infinity.
For AI detection purposes, the relevant insight is simpler: when a language model like GPT-4 generates text, it picks the statistically most likely token at each step. The result is writing where almost every word choice was the obvious choice — which produces a very low perplexity score when you run the text back through a detection model.
Human writing does not work this way. We pick words based on aesthetic preference, emphasis, rhythm, and personal habit. A food writer might say "the sauce had a mineral quality" instead of "the sauce tasted slightly metallic." A tech journalist might write "the numbers tell a stranger story than the headlines" instead of "the data shows something different." These choices are less statistically predictable — they raise the perplexity score.
High vs low perplexity — what it looks like
Low perplexity
AI signalEvery word is the most likely choice. The text reads smoothly but feels generic. Nothing surprises you. You could have predicted the next word before reading it.
High perplexity
Human signalWord choices are less predictable. The writer uses an unexpected analogy, a slightly unusual phrasing, or a word that is technically correct but not the obvious pick. Feels more distinctive.
What burstiness actually means
Burstiness measures variation in sentence length throughout a piece of text. High burstiness = sentence lengths vary dramatically. Low burstiness = all sentences are roughly the same length. Human writers naturally oscillate between long and short sentences. AI models tend to produce sentences of similar lengths throughout, giving their text a flat, metronomic rhythm.
The term originates in network science, where it describes traffic patterns that come in irregular bursts rather than a steady stream. Applied to writing, a "bursty" text has long sentences followed by very short ones, then medium ones, then long ones again — no predictable pattern.
Think about how a skilled writer punches a point home. They build a long, complex sentence that lays out the context, adds the complication, includes the qualification — and then drops a single short sentence. Like this. That contrast is deliberate. It creates emphasis. It forces the reader to stop.
AI writing almost never does this organically. When an AI model generates a response, each sentence tends to carry roughly the same information load and therefore roughly the same length. The result is text that, while grammatically correct, has a uniform rhythm that experienced readers find oddly flat — even if they cannot explain exactly why.
Measuring burstiness in practice
You do not need a tool to spot low burstiness. Count the words in five consecutive sentences. If they are all between 18 and 26 words, you are looking at low burstiness. A human-written passage over the same length might have sentences of 8, 34, 11, 42, and 6 words. That variation is burstiness.
Pick any paragraph from a piece of text. Count the words in each sentence. If no sentence is shorter than 12 words and none is longer than 28 words, that paragraph has almost certainly been AI-generated or heavily AI-edited. Human paragraphs almost always contain at least one outlier on either end.
How detectors combine both signals
No production AI detector uses only one signal. The reliable ones layer perplexity and burstiness together, then run the result through a classifier trained on thousands of known AI and human text samples.
The basic logic works like this:
- Low perplexity + low burstiness → strong AI signal. Both indicators point the same direction. High confidence flag.
- Low perplexity + high burstiness → weaker signal. The word choices look AI-like but the rhythm variation is human-like. Could be AI with edited rhythm, or a formal human writer.
- High perplexity + low burstiness → another mixed case. Surprising word choices but uniform sentence length. Could be a deliberate writing style.
- High perplexity + high burstiness → weak AI signal. Both indicators look human. Most detectors will score this as likely human-written.
The reason different detectors give different scores on the same text is that they use different training data and weight these signals differently. A detector trained mostly on ChatGPT output will have very specific ideas about what "low perplexity" looks like for that model — and may miss the signature of Gemini or Claude output, which has slightly different statistical patterns.
Perplexity in real text — with examples
This is where the theory becomes practical. Here are two versions of the same sentence — one AI-generated, one rewritten with higher perplexity.
Content marketing is an effective strategy for businesses looking to increase their online visibility and attract potential customers.
Every word is the obvious choice. "Effective strategy," "online visibility," "potential customers" — all high-frequency phrases in marketing writing. A model trained on this domain would predict each word easily.
Content marketing is less a strategy than a bet — you invest months of writing before knowing whether any of it will find the audience it was meant for.
"Less a strategy than a bet" is an unexpected framing. "Invest months of writing" is less common than "create content." "Find the audience it was meant for" is an unusual construction. Perplexity rises.
Using high-quality images on your website can significantly improve user engagement and help communicate your brand message more effectively.
A sentence that could appear in almost any marketing article without modification. No distinctive word choice. Fully predictable from any model trained on web content.
Bad stock photos do more damage than most site owners realise — they signal that the company could not be bothered to show you what they actually look like.
"Do more damage than most site owners realise" is an opinionated, less predictable construction. "Could not be bothered to show you" is conversational and specific. Much harder for a model to predict word by word.
Burstiness in real text — with examples
Email marketing remains one of the most cost-effective channels available to small businesses today. When executed correctly, it allows companies to communicate directly with their customers. Building a strong email list takes time and consistent effort. The return on investment can be substantial when campaigns are well-targeted.
Sentence lengths: 19, 18, 13, 15 words. Variation is minimal. Rhythm is flat throughout. No sentence stands out.
Email marketing still beats nearly every other channel on ROI — not because it is flashy, but because the audience opted in. They asked to hear from you. That changes everything about how they read what you send, and most companies completely waste that permission by sending content designed for strangers.
Sentence lengths: 24, 8, 4, 26 words. The four-word sentence "That changes everything" creates deliberate emphasis. Rhythm is varied and purposeful.
Why different detectors give different scores
This is the question that frustrates most people who use these tools: how can three detectors give scores of 15%, 62%, and 88% on the same paragraph?
Three reasons explain almost all of the variation:
- Different training corpora. A detector trained heavily on GPT-3.5 output will have calibrated its perplexity thresholds for that model's specific vocabulary patterns. Run text generated by Claude 3.5 Sonnet through it and the patterns may be subtly different enough to lower the confidence score.
- Different signal weighting. Some tools weight burstiness very heavily and treat low sentence-length variation as near-definitive. Others focus primarily on perplexity at the token level. The same text can score very differently depending on which signal a tool leans on.
- Model version updates. AI models improve continuously. Detectors trained six months ago are calibrated to older model outputs. Newer models like GPT-4o produce text with slightly higher natural perplexity than their predecessors — meaning older detectors may score newer AI text as more human-like.
Never rely on a single detector score. Run text through at least two or three tools and look for consistent patterns — which sentences are flagged across all of them. Those are the real problem areas. Where detectors disagree, the text is genuinely borderline.
Why human writers get false-flagged
This is one of the most important and underreported problems in AI detection. Perplexity and burstiness are useful signals, but they are not infallible — and certain categories of human writers are systematically penalised by them.
Non-native English speakers are the most affected group. When writing in a second language, people tend to stick to familiar, safe vocabulary — which is exactly the kind of low-perplexity word choice that detectors flag as AI. A Ukrainian software engineer writing a technical blog post in English may score 85% AI without having used any AI tools at all.
Technical and academic writers face similar problems. Formal writing conventions demand specific vocabulary and consistent sentence structures. A legal document, a scientific paper, or a product specification sheet will often have low burstiness by design — variation in sentence length would actually be a quality problem in those contexts.
Writers trained in certain journalistic styles — particularly those schooled in plain-language communication — learn to write short, direct sentences of similar lengths. This produces low burstiness scores that look AI-like to detectors.
A detector score is not evidence of AI authorship. It is a probabilistic signal. Using a high detector score as grounds for an accusation — in academic, employment, or editorial contexts — without substantial corroborating evidence is not appropriate and, in many institutional contexts, explicitly prohibited.
How to improve your perplexity and burstiness
If your own writing — or AI-assisted writing you have edited — is getting flagged, the fix is not mysterious. These are the changes that actually move the signals.
To raise perplexity
- Replace category words with specific words. Not "effective marketing strategy" but "the campaign that doubled pipeline in Q3." Specific is always less predictable than categorical.
- Use your own vocabulary, not the topic's vocabulary. Every subject area has a set of high-frequency terms that detectors recognise as statistically common. Reaching for a less conventional but equally accurate word raises perplexity.
- Add genuine opinion. Opinionated phrasings ("this approach is overrated," "most guides get this wrong") are statistically unusual in their construction and therefore raise perplexity meaningfully.
- Include acknowledged uncertainty. "I'm not sure this applies at scale" is not a phrase any AI model generates spontaneously. It is highly unpredictable — and therefore high perplexity.
To raise burstiness
- Add short sentences deliberately. After a long, complex sentence, stop. One word or one clause. Let it land.
- Cut the connective tissue. Many AI sentences are long because they join two thoughts with "which," "that," or "and." Split them. Let the gap do the work.
- Vary paragraph length too. A single-sentence paragraph between two longer ones creates a burstiness effect at the paragraph level, which some detectors also measure.
Tools like Humanify's AI Humanizer handle the mechanical parts of this — varying rhythm and removing flat transitions at scale. But the perplexity work — adding genuine specificity and personal perspective — still requires a human decision. No tool can fabricate a real opinion.
Want to see how your text scores right now? Humanify's AI Detector is free and shows sentence-level results.
Check Your Text FreeThe real limits of these signals
Perplexity and burstiness are genuinely useful detection signals. They are not sufficient ones. A few things worth keeping in mind:
- Both signals shift as AI models improve. Modern models like GPT-4o and Claude 3.5 Sonnet produce text with higher natural burstiness and slightly higher perplexity than earlier versions. The detection gap is narrowing. Detectors trained a year ago are already partially outdated against current model outputs.
- Both signals can be gamed. Once you understand what detectors measure, you can deliberately edit text to score better without the content being meaningfully more human. This is why content quality — whether the piece adds genuine value, demonstrates real expertise, contains first-hand knowledge — remains the standard that matters beyond detection scores.
- Neither signal detects factual accuracy. AI text can score high on both signals while containing confident errors. Perplexity and burstiness say nothing about whether the content is true. That evaluation still requires a human reader.
For a deeper look at how to use these signals practically when evaluating text, the guide on how to detect AI-written content covers the full multi-method approach. And if you are working on making your own writing less detectable — meaning more genuinely human — the techniques in making AI writing sound human address both signals directly.