Frequently asked questions

A DeBERTa-v3 model running on our own servers reads the text in overlapping windows and scores the whole document and each sentence.

We calibrated it on 583 human-written academic texts so that about 1% are labelled “Likely AI”; on held-out texts the false-alarm rate was 1% and 0.7%. Unedited text from 2026 AI models was labelled “Likely AI” 96.9% of the time, but text rewritten by humanizer tools only 59.5%.

No. It is a statistical estimate. Human writing is sometimes flagged, and AI writing is sometimes missed. Use it as a signal to review your draft, not as evidence.

Paste at least 50 words, or upload a .docx or .pdf file (up to 4 MB and 100 pages). Without an account you can check 1,000 words at a time, 3 times a day.

No. It checks English text only. The website is available in 17 languages, so you can read the result in yours.

Our detector marks the sentences that read as machine-written. Only those are rewritten: several alternatives are scored, and the most natural one that scores lower is kept. Everything else stays as you wrote it.

No. Every revised sentence is checked: citations, numbers and abbreviations must match the original exactly, or the sentence is kept as you wrote it.

It rewrites only the sentences our detector flags, and keeps a rewrite only if our detector scores the text lower and the sentence still reads naturally. You see our detector's verdict before and after. Other detectors, including your school's, may score differently — we don't guarantee any result.

Yes. Polishing the revised text again usually lowers the flagged share further. If sentences are still flagged after two passes, rewriting the grey ones in your own words works best.

Checking text costs 1 credit per word and polishing costs 4 credits per word. Headings and reference lists are not charged when polishing.

Yes. Without an account you get 3 AI checks a day, up to 1,000 words each. A free verified account adds 10,000 words of AI checks every month.

Credit packs are valid for 12 months. Subscription credits are granted every month and do not roll over.

Credits for failed jobs are refunded automatically.

Your text is deleted from our servers within 24 hours, and you can delete it sooner. It is never used to train any model.

Polishing sends the sentences to Anthropic (Claude) for editing. Payments are processed by Stripe. Emails are sent through Resend. Hosting is on Vercel and Railway. Anthropic does not use API data for training and keeps it for up to 30 days under its standard policy. Site usage statistics are processed by Google (Google Analytics).

Only to run the detection or polishing you asked for, enforce limits, prevent abuse and handle payments. We do not sell your data and we do not use your texts to train models.