Paste your text. See every AI tell, highlighted.

Yes. It's plain JavaScript running in your browser — the page has no backend, and your text is never uploaded, stored or sent anywhere. Reload the page and it's gone.

It scans against our maintained blacklist of 90+ phrases and constructions that current AI models overuse (delve, moreover, em-dash chains, 'in today's fast-paced world'…), and computes tell density per 1,000 words. It's a style scanner, not an 'AI probability' detector — which is why its results are explainable, phrase by phrase.

Detectors output a probability score you can't argue with. This shows you the exact phrases that make text read as machine-written — so you can fix them. Different job: they judge, this edits.

Fix the highlighted tells with the replacement tables on the blacklist page, run one of the humanize prompts in your own AI, or use a multi-model rewrite. Then re-scan. The before/after gallery shows the full workflow on real examples.

No honest tool can promise that — detectors are imperfect and change weekly. What removing the tells reliably does is make text read better to humans, which is the standard that actually matters.

Software that estimates the probability that a text was machine-generated. It does not “know” who wrote something — it measures statistical properties of the writing and compares them with patterns typical of AI output. The result is a probability score, not evidence of authorship.

Most rely on two measurements. Perplexity: how predictable each next word is to a language model — AI text is produced by systems choosing likely words, so it scores as highly predictable. Burstiness: how much sentence length and structure vary — human writing swings between short and long sentences, generated text tends to cruise at one altitude. Newer detectors add trained classifiers that learned from millions of labeled samples. All of them measure how conventional the style is, not who typed it.

Vendor claims like “99% accuracy” cannot be independently verified, and outside tests regularly find both missed AI text and flagged human text. Accuracy also decays: every new model generation writes differently, so a detector tuned on last year's output ages quickly. Treat any score as a hint, never a verdict.

Because they detect conventional writing, not machine writing. Non-native speakers who learned textbook English, writers in rigid genres (reports, legal text, academic abstracts) and anyone with careful, even prose all score as “predictable”. Research has repeatedly found elevated false-positive rates for non-native writers — a key reason several universities have disabled detector integrations.

Sometimes, temporarily, unreliably. Scores are so volatile that paraphrasing often flips a verdict — in both directions. But no method reliably beats all detectors today and still works next month, because detectors are retrained constantly. Any tool guaranteeing “undetectable” text is promising something it cannot deliver. And if the goal is hiding AI use where it's prohibited, the real risk was never the detector — it's the moment a human asks you to explain your own text.

That is not what a good one is for. A humanizer that synonym-swaps to dodge scores usually makes the text worse, and the result stays fragile against the next detector update. What legitimately changes detector behavior as a side effect is actual editing: adding real specifics, varying rhythm, cutting boilerplate. Better writing happens to be less statistically predictable — that is the honest overlap, and it is the only durable one.

Three things. Keep drafts and version history — process evidence beats any counter-score. Add concrete details only you would know: names, numbers, decisions, constraints. Vary your sentence lengths deliberately. And wherever AI assistance is allowed, disclose it — disclosure turns a detection question into a non-issue.

Poorly. Statistical measurement needs sample size; below a few hundred words the confidence collapses, which is why many vendors enforce minimum lengths. A score on a single paragraph is close to noise.

Some AI labs can embed statistical watermarks into generated text that matching tools can later detect. In practice, adoption is partial, watermarks weaken under editing and translation, and there is no universal standard — so watermark detection only ever covers a slice of AI text.

No — even the vendors say scores should not be sole evidence, and most institutional guidance treats a flag as a reason for a conversation, not a finding. If you are the one holding the score: ask about process, drafts and understanding. Ten minutes of questions settle what a percentage never can.