AI Text Detector

What this tool is used for

• Getting a heuristic signal about a passage's likely origin.
• Comparing two passages on the same signals.
• Understanding which features these detectors actually measure.
• Checking your own writing against the heuristics out of curiosity.
• Demonstrating in a discussion how weak such signals are.

Frequently Asked Questions

How reliable is AI detection?

Not reliable enough to accuse anyone. Every detector, including this one, measures statistical regularity — and human writing that is careful, edited or non-native scores as "AI-like" for exactly the same reasons. Treat a high score as a prompt to read the text, never as evidence.

What is burstiness?

Variation in sentence length. Human writing mixes a four-word sentence with a thirty-word one; generated text tends to hover around a mean. Low variance is the signal — but a technical manual, an abstract or a translated text is naturally low-variance too.

What else is measured?

Vocabulary diversity, sentence-length variation, and the frequency of phrases that appear disproportionately in generated prose. Each is weak on its own; the combined score is still a heuristic, not a classifier trained on labelled data.

Does it work on languages other than English?

The sentence splitting does — it understands CJK terminators (。！？), so a Chinese paragraph is no longer read as a single sentence and scored as structurally zero-burstiness. The phrase list is English, so that signal simply does not contribute.

Can text be edited to defeat it?

Trivially, and that is the deeper problem. Varying sentence lengths and replacing a few stock phrases moves any score substantially, so the tools are easiest to fool for anyone deliberately trying — leaving honest writers as the ones most often flagged.

Why do false positives fall hardest on non-native writers?

Because the signals — predictable word choice, uniform sentence length, limited vocabulary — describe careful second-language writing as accurately as they describe generated text. Published studies have found exactly this bias, which is why a score must never stand alone as evidence.

What would actually prove authorship?

Process, not statistics: draft history, version control, or a conversation about the work. A detector estimates how predictable text is, and predictable writing is not the same thing as generated writing.

Common errors and gotchas

• Treating the output as evidence, when these detectors have high false-positive rates and no reliable ground truth.
• Using it to accuse someone, which has caused documented harm to non-native writers in particular.
• Assuming a low score means human, since light editing defeats every heuristic.
• Applying it to short passages, where there is far too little signal.
• Reading a percentage as a probability, when it is a heuristic score with no calibration.