HumanTalky.

Free · Sentence by sentence · Not stored by default

Claude AI Detector

Check whether text reads like Claude wrote it. You get a score, a confidence level and every sentence highlighted, so you can judge for yourself.
|Try:0 words · 0 / 20,000

Your text is only sent when you press Scan and is not stored, unless you choose to share a report or tick the feedback box. Quick scans are unlimited; deep scans use your daily allowance.

What Anthropic says about Claude text marks

Anthropic has said Claude text carries a statistical mark in its word choices, with a detector only in private preview so far. Short passages and code carry little of it.

Status: Statistical mark in word choices. Checked 12 October 2026. This changes as providers announce things; tell us if it is out of date.

How Claude writing tends to read

Claude output often shows careful, balanced sentences, hedged claims, tidy lists of three and a fondness for em dashes. These habits are what the detector scores. Human writers share some of them, which is why a score is a signal and never proof.

Claude detector questions

Can this detect Claude text?+

It detects the style of machine-written text, which includes Claude's typical habits: careful, balanced sentences, hedged claims, tidy lists of three and a fondness for em dashes. Anthropic has said Claude text carries a statistical mark in its word choices, with a detector only in private preview so far. Short passages and code carry little of it.

How does the Claude detector work?+

It reads every sentence for the habits of chat-model writing: even rhythm, stock phrases, formal wording with few contractions, tidy lists of three and repeated openers. A deep scan adds a language-model review on top of the statistics.

How accurate is it?+

It gives a signal, not proof. Short text, translated text, heavily edited text and very formal human writing can be misread either way, so every result shows a confidence level. Never use one score to accuse someone of using AI.

Can it tell which AI model wrote the text?+

No. It estimates whether text reads as machine-written, not which company's model produced it. Model-specific marks, where they exist, need the provider's own key.

Is my text stored?+

Not by default. Text is only sent when you press Scan and is not stored, unless you choose to share a report (kept 30 days) or tick the box that shares it with us as feedback.

Why does it say low confidence?+

Under about 80 words there is too little to judge, and scores in the middle range are uncertain by nature. Paste more text, or read the highlighted sentences yourself.