Free AI Content Integrity Checker

No. Exact character checks and writing-pattern prompts can identify things worth reviewing, but they cannot prove human or machine authorship. The receipt records what ran, what did not run and the limits that apply.

For the AI model check, by default, yes, and the page says so where you paste rather than in the small print. The whole document is sent in one request over HTTPS to our own server in the EU, Google Cloud Run europe-west1 in Belgium, scored in memory and then discarded. Your text is not stored or logged, and the request leaves no per-visitor record, only an anonymous daily count. Every result prints how many of your words were sent and repeats the server’s own answer for what it did with them, so the claim is checkable on each run. Choosing “Process entirely in my browser” moves the same model onto your device after a one-off 34.5 MB download and sends nothing at all. Every other check, the hidden characters, the lookalikes, the protected facts, the writing rules and the watermark scan, has always run in this browser and still does. On either route the draft is never placed in the page URL, in browser storage, or in the analytics event the form fires.

Below 50 characters, the checker runs exact character checks but suppresses writing-pattern judgement. A short sample does not provide enough context for responsible editorial pattern prompts.

No. It previews only selected, allowlisted treatments. Joiners, combining marks, ambiguous bidirectional controls, links, code, quotations and protected spans are not changed automatically.

Not an Anthropic production watermark. Those keys are private and no public verifier exists, so that check is reported as not assessed rather than guessed. What the checker does run, on every assessment, is the published SynthID-Text detection mathematics against this tool's own three public demo keys, in your browser, and it shows the score for each key. A result near the 0.5 no-watermark value under our demo keys says nothing about any provider's production watermark, in either direction.

Measured on a 5,558-document long-form corpus; these rules have no training set, so none of it was fitted to them, the 116 writing-signal rules reached mixed signals or above on 45.1% of the AI writing while flagging 24.8% of the human writing. That is worse than the trained model on both counts, so on 28 August 2026 the rules stopped contributing to the AI verdict and became editing suggestions. They read register and chat-export formatting rather than authorship, which is why text pasted through an editor that strips bold, headings and bullets loses most of that signal, and why short question-and-answer text carries almost none of it whatever produced it.

Uploaded JPEG, PNG, WebP and PDF files are read locally for C2PA Content Credentials using the official Content Authenticity Initiative library. A file with no credentials is reported as exactly that, certificate trust is not judged and the page says so, and pasted text is never given a provenance verdict.

A small trained classifier, and the only check on this page that gives an AI reading. It runs one of two ways and you choose which: on our own EU server by default, which needs no download and answers in under a second, or entirely in your browser after a one-off 34.5 MB download. Both routes are the same trained file and both read the whole document in consecutive sections cut to fit the model's 512-token window — typically about 340 words, fewer where the writing is dense — reporting the strongest section. They run at different numerical precisions but flag at the same point, which every result prints beside the score rather than stating here, because that point is calibrated and can move; each route's accuracy is measured on the runtime that produced it and is shown with its denominators alongside the result. Its measured accuracy is disclosed with every result, it misses text a person rewrote far more often than text a model wrote in one pass, and it can flag human writing, so it is shown as one named check among several and never as proof. The trained classifier detects 883/922 (95.8%) of AI-written long-form documents on the EU server route and 889/922 (96.4%) in the browser, and wrongly flags 45/4,636 (0.97%) of human-written long-form documents on the server and 90/4,636 (1.94%) in the browser. Both figures are measured over the whole corpus at the operating point that ships. Every one of those figures is long-form: short marketing, SEO and social copy has never been measured on independent data. If the server route is unreachable or rate-limited, the tool says so and offers the browser route; it never substitutes a number from the writing rules.