Is this text written by AI?

Paste any block of writing. We estimate AI-likelihood across a three-way verdict — human, mixed, or AI-assisted — and show the evidence behind every signal. No black-box score, no upload required, runs entirely in your browser.

Honest about the limits: the 2025 DetectRL benchmark put DetectGPT at AUROC 22.15 on academic writing and 12.21 on news under realistic conditions. Binoculars on Claude: 55.15 AUROC. Mixed human/AI text: ~52.51 average AUROC across classifiers. We treat results as triage and evidence — never as grounds for accusation. (The FTC's 2025 action against Workado is a reminder why overstated accuracy claims invite regulatory risk.)

runs locally — nothing uploaded

What you'll see

Anatomy of a text result

Six signals run in parallel, each independent of the others. The point is the breakdown — three signals firing on the same passage is the call, not any single number.

  • Burstiness

    σ / μ

    Variance of sentence length. Humans alternate long and short; AI converges on a flat rhythm.

  • Lexical tics

    Density

    Density of LLM-favored words and filler phrases (delve, tapestry, navigate, leverage, in conclusion…).

  • N-gram repetition

    Trigram

    How often the same 3-word sequence repeats. AI text recycles phrasing across paragraphs.

  • TheItSheWeButNow

    Sentence-start variety

    Unique / N

    Distinct opening words divided by sentence count. AI reuses openers; humans cycle through.

  • — ; — ;density per 1k chars

    Punctuation pattern

    Em-dash · ;

    Em-dash and semicolon density. Recent LLMs over-use both; most casual writers use neither.

  • Per-token perplexity

    Heatmap

    Each word colored by predictability. Bright = AI loves this token. Dim = surprise.

Three-way verdict, not binary

The honest 2026 framing is human / mixed / AI-assisted. Most real-world writing is co-authored — generic AI detectors that force a yes/no choice mislead users. We show evidence per sentence with explicit sample-quality flags.

Robustness under paraphrase

StealthGPT and humanizers still defeat most detectors. We re-test under light paraphrase variants and flag the verdict as fragile when small rewrites flip it — turning a known weakness into transparent UX, not a hidden failure.

Author-baseline mode (coming)

For journalists, legal teams, and publishers: upload prior writing from a claimed author. We then score whether new text is consistent with that author's historical style — forensic authorship, not generic AI-vs-human. Read the May 2026 research roundup.