Different formats keep different layers, so what you can expect differs too. Alongside what you can put in, this lists what tends to come out of it.
PNG, JPEG, WEBP, TIFF, HEIC, GIF and BMP are read. PNG tells you most through the text chunks a generator leaves; JPEG through its quantisation table and thumbnail structure. Regions that look edited are marked on a heatmap.
Reads PDF, DOCX, HWP, HWPX, PPTX and XLSX alongside TXT, MD, RTF and CSV. Korean HWP was in from the start because it is what Korean offices actually use. In documents the questions are the authoring tool, the edit history, the revision sessions and whether the cited sources exist. In source code (Python, JavaScript, TypeScript, Java, C/C++, Go, Rust, HTML, CSS and more) style is not examined — comment density and naming are decided by linters and coding standards, so they cannot serve as evidence. What is counted instead is the generation marker a file declares about itself, and residue carried over from a chat window.
MP3, WAV, FLAC, OGG, OPUS, M4A and AAC, plus MP4, MOV, WEBM, MKV and AVI. Container information, encoding history, waveforms and frames are analysed. No deep detection model is included yet; when one is, its limits will be published with it.
A ZIP is opened and each file inside scanned separately, with the results grouped. You can also paste text straight into the scan box, where the limits of stylometry apply in full.
This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.