Text

Sentence-by-sentence AI text detection, with the math.

Paste a passage of 80+ characters and Detecto returns a calibrated probability, sentence highlights, and the perplexity / burstiness numbers behind both. Sign-in optional for the demo.

See a sample report How calibration works

Try a scan

Paste any passage of 80+ characters.

0 / 50,000Demo, not saved

0 words · ≈ 1 credit · 1 credit per 1,000 words

No account needed. Anonymous scans are rate-limited per IP.

Built for the people who review other people’s content

Editors at independent magazines
University writing centers
Newsroom verification desks
PR review teams
Content QA leads at agencies

Accepted inputs

What you can paste.

Designed for the kinds of submissions a reviewer actually sees: long-form pieces, contributor drafts, student essays, batched copy.

Long-form prose

Essays, op-eds, analysis pieces, articles. Up to 50,000 characters per submission.

Contributor drafts

First-pass drafts from freelancers and contributors before the edit cycle.

Student work

Essays, take-home exam responses, reflective writing, lab discussion sections.

Batched copy

PR pitch batches, SEO content batches, agency QA samples across a billing cycle.

What we measure

Signals exposed in every text report.

Vendor-agnostic feature contributions; the label is the same whether the model behind it shifts or not.

Statistical

Average perplexity

Surprise per token relative to a base language model. Lower values are more model-typical.

Statistical

Burstiness

Variance in sentence-level perplexity across the passage. Human prose tends to be bursty.

Statistical

Sentence-length variance

Distribution of sentence lengths. Generation-typical output flattens this distribution.

Lexical

Lexical diversity (TTR)

Type-token ratio; how often unique words appear per total word count.

Lexical

Connector frequency

"First", "Second", "Ultimately", "In today’s world" and the rest of the giveaway scaffolding.

Calibration

Confidence variance

Detector’s own variance across resampled detection passes, how stable the score is on this input.

Honest limitations on text detection

The features that drive the score are well-understood and exposed in the report. The places they fail are exposed too.

Very short input (under ~120 words) widens the confidence band.
Translated copy can flag higher than human-original text in the same language.
Heavy genre conventions (academic, legal, regulatory) compress lexical diversity and shift the baseline.
A single high score on a passage you trust deserves a careful second look, not an automatic verdict.
The detector’s confidence variance is shown alongside the point estimate; respect it.

Run a scan

The full report fidelity, on the free tier.

No card to start. Your scans save privately to your dashboard, shared only when you choose.