How Our AI Detection Works — and Where It Can Be Wrong
We don’t build or train detection models. Each tool sends your content to a third-party provider and shows you what comes back. This page lists those providers, the real limits, what a score means, what we store, and the ways results can be wrong.
Which provider checks what
| Tool | Provider | What you get | Limits |
|---|---|---|---|
| AI Text Detector | ZeroGPT.com API (third-party) | AI-likelihood percentage, highlighted sentences, a short verdict and the detected language | 300 words minimum, 150,000 characters maximum |
| AI Image Detector | Sightengine (third-party) | AI-generation probability and a separate face-manipulation probability | One image up to 15 MB (PNG, JPG or WEBP) |
| AI Video Detector | ZeroGPT.com API (third-party) | AI-generation probability with ML, OCR and metadata sub-signals; no face-manipulation score | One clip up to 50 MB (MP4, MOV or WEBM) |
| Plagiarism Checker | ZeroGPT.com API, which runs the search through Copyscape; Winston AI as a backup if that fails | Plagiarism and originality percentages and a list of matching source URLs with match percentages | Up to 150,000 characters |
| Grammar Checker | LanguageTool (open-source engine, public API) | Spelling, grammar and style issues with up to five suggestions each, plus our own simple style scores | Up to 20,000 characters |
Uploads for the text tools can be .txt, .md, .docx or .pdf files up to 8 MB. The site supports English only.
What a score means
Every detection result is a probability from the provider’s model, not a fact about who or what made the content. We turn the percentage into a label using fixed thresholds: 55% or higher is shown as an AI-generated signal, 30% to 54% as mixed or inconclusive, and below 30% as a human or authentic signal.
Those labels are a reading aid. A high score is a reason to look more closely; it is never proof, and a low score never proves that a person made something. Don’t use a score on its own to accuse anyone or to make an academic, hiring or legal decision.
How results can be wrong
False positives on human writing
Text detectors can flag genuine human writing as AI. It happens more with short passages, formulaic writing (abstracts, legal boilerplate, templates), very polished text and writing by non-native English speakers.
Missed AI content
AI-generated text that has been edited, paraphrased, translated or mixed with human writing can score as human. A low score is not proof that a person wrote something.
Compressed or re-encoded media
Compression, resizing, screenshots and re-encoding remove detail the image and video models rely on, so results on copies are less reliable in both directions. Re-encoded video often loses its original metadata, and the OCR signal is empty when a clip has no readable text.
Video is not a deepfake detector
The video result estimates whether a clip as a whole looks AI-generated. It returns no separate face-manipulation score, so it can't confirm or rule out a face swap.
Plagiarism only covers searchable sources
Matches come from sources the provider can search, mostly public web pages. Paywalled papers, private documents and offline sources can be missed, and properly quoted or cited passages still count as matches.
Grammar suggestions are automated
Suggestions can be wrong or miss errors. The style and readability scores shown with grammar results are our own simple calculations, not a certified measurement.
More detail: why AI detectors flag human writing and why there’s no single accuracy number.
What leaves our servers and what we store
- Your text, image or video is sent to the provider listed above for that tool, so it leaves our servers. Each provider’s own privacy policy covers what it does next.
- If you’re signed in, your scan history keeps the first 200 characters of scanned text (or the file name for images and video), the verdict and the score. It never keeps the image or video itself.
- Grammar checks aren’t saved to your scan history.
- Signed-out visitors get a cookie with a random ID, used only to count free scans; they have no server-side scan history.
Full details are in the privacy policy.
Free scans and plans
3 free scans in total, shared across all five tools (not per day or per month). No signup needed to start. Paid plans remove the free scan limit on AI detection and add plagiarism and grammar word allowances; see pricing.