AI detector and writing-pattern checker

AI detectors try to answer a question that has no perfectly observable answer: who or what wrote this text? Unlike plagiarism checking, which compares a document against a known database, detection looks for statistical fingerprints in the writing itself. That distinction matters because the fingerprints are probabilistic, not definitive.

The accuracy problem

Published evaluations of commercial detectors show a wide spread of reliability. Studies have reported error rates that spike dramatically for two groups in particular: non-native English writers, whose more constrained vocabulary mirrors the predictability of model output, and highly edited or formally structured human writing, which can also look uniform. One widely cited evaluation found that classifiers built on older GPT generations failed to generalise to newer models entirely after an update, dropping toward chance performance overnight.

That failure mode is inherent: models change, prompting styles evolve, and detectors that are not continuously retrained drift out of date. When you use any score — including ours — you are seeing a snapshot of pattern-matching against the training data the detector was built on.

What the signals actually measure

Most detectors, including this one, combine several cheap, interpretable features rather than relying on a single opaque model. Perplexity — roughly, how “surprised” a language model is by a word choice — is a common one: AI text tends to pick high-probability words, so it reads as low-perplexity. Burstiness, the variation in sentence structure and length, is another; human writing tends to alternate between long, complex sentences and short ones, while model output is more evenly paced.

These are useful proxies, but they are proxies. A careful human editor working from a house style guide can produce low-perplexity, low-burstiness text. A model prompted to write in fragments can produce the opposite. The score reflects the text, not the author.

How to use a score responsibly

• Treat it as a prompt, not a verdict. A high score is a reason to read a passage more carefully — nothing more.
• Look at the evidence, not just the number. Which sentences were flagged and why tells you more than the aggregate score.
• Consider the context. Form emails, technical documentation and exam answers all have legitimate reasons to sound uniform.
• Never act on a score alone. Academic and employment decisions require human review, drafts, and conversation.

For a deeper walkthrough of the underlying mechanics, see how AI detection works. If you have been wrongly flagged, AI detector false positives explains what to do next.