Why This Is Harder Than It Looks
Every few months a new "AI detector" launches claiming near-perfect accuracy, and every few months researchers show it can be fooled by a light edit pass or flagged human writing as AI with total confidence. The honest answer is that reliably proving authorship of a block of text is an unsolved problem, and anyone selling you 99% certainty is overselling. What does work is looking at the statistical fingerprints AI writing tends to leave — not as proof, but as a useful signal.
Signal One: Burstiness (Sentence-Length Variance)
Human writers are inconsistent in a specific way: we mix short punchy sentences with long, winding ones, often without meaning to. Linguists call this burstiness. Large language models, trained to produce smooth, readable prose, tend to even this out — sentence after sentence lands in a similar length range. You can measure this yourself: split a passage into sentences, count the words in each, then compute the standard deviation ÷ mean (the coefficient of variation). A low number means suspiciously uniform pacing; a high number is a good sign of human variance.
Signal Two: The LLM "Tell" Phrases
Certain phrases show up in AI-generated text far more often than in ordinary human writing, because they're statistically common completions the model has learned to reach for. Watch for clusters of phrases like:
- "It's important to note that..."
- "In today's fast-paced world..."
- "Delve into" / "rich tapestry" / "boasts a"
- "Furthermore" / "moreover" / "in conclusion"
- "Navigate the complexities" / "unlock the potential"
One or two of these mean nothing — plenty of humans write "furthermore." A dense cluster of five or six of them in a short passage is a much stronger signal.
Signal Three: Transition-Word Repetition
AI-generated essays often over-structure themselves, opening sentence after sentence with formal transitions — "Additionally," "Moreover," "On the other hand." Human writing tends to connect ideas more implicitly. If more than a quarter of a passage's sentences open with a formal transition word, that's worth noting alongside the other two signals.
What This Can and Can't Tell You
Combining these three signals into a single score gives you a reasonable heuristic estimate — good enough to flag a suspicious product review, a possibly-fabricated testimonial, or a school essay worth a closer look. It is not good enough to:
- Prove authorship in an academic integrity case
- Justify firing or rejecting someone over a cover letter
- Override a human's own denial without other evidence
Formal, careful human writers — technical writers, non-native English speakers who learned formal writing patterns, lawyers — can trigger "AI-like" scores on burstiness and transitions alone. And AI text run through even a light manual edit pass can dodge the phrase list entirely. Treat the score as one data point, not a verdict.
Where This Heuristic Is Actually Useful
Despite its limits, a fast local heuristic has real, practical uses. Content moderators triaging a queue of product reviews can use it to prioritize which ones deserve a closer human look. Editors reviewing freelance submissions can flag pieces worth a second read before publishing. Teachers building a portfolio of a student's writing style over a semester can spot a single essay that reads very differently from everything else that student has turned in — not as proof, but as a prompt for a conversation. In every one of these cases, the score is a triage signal, not a verdict, and it works best combined with context the tool itself can't see: does this sound like how this person usually writes?
Using the Detector Responsibly
The AI-Generated Text Detector runs all three checks locally in your browser — nothing you paste is uploaded anywhere — and shows its work: the exact burstiness number, the exact phrases matched, and the exact transition ratio, so you can judge for yourself whether the flags actually make sense for the text in front of you. Paste in 100+ words for the most reliable read, and always pair the score with your own judgment about context.