How this AI text detector actually works
Type "AI detector" into a search bar and you'll find dozens of tools that hand back a single confident-looking percentage with no explanation of where it came from. That opacity is a problem: a percentage you can't interrogate isn't much more useful than a coin flip dressed up in a nice interface. This tool takes a different approach. It's built entirely on two measurable, well-documented signals from computational linguistics — burstiness and a perplexity proxy — and it shows you both numbers, plus which sentences pushed the score in which direction, instead of hiding the math behind a black box.
Everything below explains exactly what those two signals measure, why they correlate with AI-generated text, where the correlation breaks down, and how to read a result responsibly. If you only take one thing from this page, take this: a detection score is evidence, not a verdict.
Burstiness, explained
Burstiness measures how much sentence length varies across a passage. Human writers are naturally "bursty" — we write a short, punchy sentence, then a long, winding one with three clauses, then another short one for emphasis. That variation isn't a stylistic choice most people make consciously; it's a side effect of how humans plan language in real time, influenced by breath, emphasis, and changing trains of thought.
Large language models, by contrast, are trained to predict the statistically most likely next token given everything before it. Left to their own devices (without an explicit instruction to vary rhythm), that process tends to produce sentences that cluster around a similar length and cadence. The result reads smoothly — arguably too smoothly — with a kind of metronomic evenness that's unusual in unedited human prose.
This tool calculates burstiness as the coefficient of variation of sentence length across your text: the standard deviation of word-counts-per-sentence, divided by the mean. A low coefficient (sentences clustering tightly around the same length) pulls the score toward "likely AI." A high coefficient (a genuine mix of short and long sentences) pulls it toward "likely human."
Perplexity, explained
Perplexity is a term borrowed directly from how language models are evaluated during training. Formally, it measures how "surprised" a language model is by a given sequence of words — low perplexity means the text was highly predictable to the model; high perplexity means it contained more unexpected word choices and phrasing.
Running true perplexity requires the same kind of neural network used to generate text in the first place, which isn't something that can happen inside a browser tab without shipping megabytes of model weights to your device. So this tool uses a lightweight, transparent proxy built from two components you can see in the results panel: lexical diversity (the ratio of unique words to total words) and the density of common AI-associated transition phrasing — words and phrases like "furthermore," "it is important to note that," "delve," "robust," and "leverage," which show up disproportionately often in AI-generated text compared to typical human writing on the same topics.
Text with a wide vocabulary and a low density of these stock phrases scores higher on the perplexity proxy (more human-like); text that leans heavily on generic connective tissue scores lower (more AI-typical).
How accurate is AI detection, really?
Independent research on commercial AI detectors — including tools used by universities — has found accuracy rates that vary widely depending on the text's length, topic, and the writer's first language, and none report anything close to 100% in adversarial conditions (i.e., text that's been lightly edited specifically to evade detection). A few patterns show up consistently across studies:
- Short passages are harder to score reliably. Under roughly 50 words, sentence-length variation and vocabulary diversity don't have enough room to show a clear pattern either way.
- Lightly edited AI text is the hardest case. Once a person rewrites even 20-30% of an AI draft in their own words, burstiness and lexical diversity shift meaningfully toward "human," even though the ideas and structure may still be AI-originated.
- Detectors disagree with each other. The same passage run through different detection tools frequently produces different scores, because each tool weighs signals differently and trains on different reference text.
None of this means detection scores are worthless — they're a genuinely useful signal, especially at the extremes (very uniform, very generic text vs. clearly idiosyncratic writing). It means a score in the ambiguous middle range, or a score on a short passage, deserves real skepticism.
Why human writing sometimes gets flagged
False positives are the most consequential failure mode of AI detection, because the cost of wrongly accusing a human writer is much higher than the cost of a detector occasionally missing real AI text. A few groups of writers are disproportionately likely to see "likely AI" readings on genuinely original work:
- Non-native English speakers who were taught formal grammar rules often write with more consistent sentence structure and rely more on textbook transition words — both signals this tool (and most others) associate with AI text.
- Technical and legal writers whose genres reward precision and consistency over stylistic variation.
- Writers on unfamiliar topics, where a narrower working vocabulary naturally reduces lexical diversity.
- Heavily edited human drafts, where a copyeditor has smoothed out the natural burstiness of an original draft in the name of "clean" prose.
If you're on the receiving end of a false positive, the burstiness and perplexity-proxy numbers on this page are useful precisely because they let you point to the actual mechanism — "my sentences are consistently 15-18 words because I write in short technical steps," for example — rather than arguing against an unexplained percentage.
What this tool is — and isn't — good for
Good uses: sanity-checking your own writing before submitting it somewhere that uses AI detection, understanding why a passage might read as "flat" or "generic" so you can revise it, spot-checking AI-assisted drafts to see how much still reads as machine-typical after your edits, and learning what actually distinguishes AI phrasing from human phrasing (the highlighted sentences below the gauge are a good teaching tool for this).
Poor uses: treating any single score as proof in an academic integrity case, running very short snippets (under 50 words) and trusting the result, or using this as the only check before publishing something where AI-generated content is prohibited — pair it with a human editorial read.
How to write text that reads as clearly human
Whether your goal is to pass detection because you're a human writer tired of false flags, or to genuinely reduce your reliance on unedited AI output, the underlying advice is the same, because it's about actual writing quality, not gaming a score:
- Vary your sentence length deliberately. After drafting, read your paragraph aloud. If every sentence takes about the same breath to say, break some up and combine others.
- Cut stock transition phrases. Replace "it is important to note that," "in today's world," and "furthermore" with a direct statement or a plainer connector like "but," "so," or nothing at all.
- Add a specific, concrete detail. A real number, a specific example, or a personal anecdote raises lexical diversity in a way generic phrasing can't.
- Read it against your own natural speaking voice. If you wouldn't say a sentence out loud that way, rewrite it until you would.
If you're starting from an AI draft and want to edit it into something that reads as your own voice, our AI Humanizer tool applies exactly this kind of sentence-length and phrasing variation automatically, as a starting point for your own edit — not a replacement for one.
How different detectors weigh signals
| Signal | What it measures | Strength | Weakness |
|---|---|---|---|
| Burstiness | Sentence-length variation | Fast, transparent, hard to fake accidentally | Genre-dependent (technical writing is naturally low-burst) |
| Perplexity (true) | Model's surprise at word choices | Strong signal when available | Requires a full language model to compute — can't run client-side |
| Lexical diversity (this tool's proxy) | Unique-word ratio | Lightweight, runs instantly in-browser | Penalizes short or single-topic passages |
| Watermarking | Hidden statistical pattern inserted by the model provider at generation time | Very high accuracy when present | Only works if the generating model actually applied a watermark, and only that provider can check it reliably |
Most commercial detectors combine several of these signals, plus a trained classifier on top, which is why scores vary between tools even on identical text. Understanding the individual signals — rather than trusting one opaque number — is the more durable skill.
Frequently asked questions
Can an AI text detector be 100% accurate?
No. Every public AI detector, including this one, produces a probability, not a verified fact. Large language models are explicitly trained to sound natural, and human writers sometimes write in a flat, consistent register too — especially in technical, legal, or academic contexts. Treat any single score as one data point, not a verdict.
Will this tool store or share the text I paste in?
No. The standard analysis on this page runs entirely in your browser using JavaScript — your text is never sent to a server, logged, or stored. You can verify this yourself by disconnecting from the internet after the page loads; the tool still works.
Why did my own original writing get flagged as "likely AI"?
The most common reasons are: very consistent sentence length (common in technical writing, legal drafting, or writing by non-native speakers who learned formal grammar rules), heavy use of transition words like "furthermore" and "moreover," and a narrow vocabulary range on a specific topic. None of these prove AI involvement on their own.
Can I use this to prove a student or employee used AI?
You shouldn't rely on any automated detector — this one included — as the sole basis for an accusation. Detection scores are probabilistic and have documented false-positive rates, particularly against non-native English writers. Use a detector as a prompt for a conversation, supported by other evidence like drafts, version history, or a discussion of the work, not as standalone proof.
Does this tool detect AI-generated images or code, too?
This specific tool analyzes written prose. For images, use our AI Image Detector, which looks at file metadata and compression signals instead of sentence-level text patterns.