Detect AI-Written Text & Essays

This landing is the text check: paste an essay, an article, marketing copy, a review or a post — anything from about fifty words up to twenty-five thousand characters. The same studio also checks images, audio, video and face swaps on separate tabs; those upload a file and run on our servers, while this one runs in your browser.

No accuracy figure is published here, because none of the numbers this page used to print had a measurement behind them. Text detection is the least reliable kind: fluent, formal or heavily edited human writing reads as machine-like, and lightly rewritten machine text reads as human. What you get is a probability with a confidence attached and the specific signals that moved it, so you can judge whether the reasons look right for the passage in front of you.

The first 2 checks use no credits, and signing in adds 2 more. After that a check costs 10 credits, and any paid plan lifts the limit entirely while it is active.

The check never identifies a system, so the question does not really apply to it. It measures word-choice and rhythm patterns typical of machine writing — it reads the writing in front of it rather than matching it against any list — which means it has no concept of which assistant produced a passage, and cannot tell you.

Locally. The text check runs entirely inside your browser: your writing is never uploaded, never transmitted and never stored, and no copy of it reaches us or anyone else. Only the usage counter that meters your checks touches the network. The image, audio, video and face-swap tabs are the opposite — those upload the file and analyse it on our infrastructure.

One of three readings — likely human, possibly AI, or likely AI — with a probability from 0 to 100, a confidence level, and plain notes on what drove it: how much formulaic machine vocabulary appeared, and how uniform the sentence rhythm and word variety were. The middle reading is a real answer, not a failure: it means the signal was genuinely mixed.

There is no wait to speak of. The analysis is arithmetic over your text and runs on your own machine, so the result appears as soon as you press the button — the brief "analyzing" moment on screen is presentation, not processing.

As one screening signal, and never on its own. There is no detection endpoint in the public API today, so checks are run one at a time in the studio rather than automated over a queue. And any policy built on a text score needs a human step before it reaches a person: this is the modality where a false positive becomes an accusation.

Five checks — text, image, audio, video and face swap — sit in one studio on one balance, and the text one runs in your browser without uploading anything. What no detector offers, here or anywhere, is certainty; a tool that reports a flat percentage with no confidence figure and no reasons is hiding the same uncertainty this one shows you.

Treat it as supporting material, never as the basis of an accusation, and be especially careful in an academic setting. Text detectors are known to flag writing by non-native English speakers at higher rates, and a probability is not evidence of authorship. Combine it with at least one other method and human review, and use a result to open a conversation with a student or an author rather than to close one.