AI detectors ranked by the upload formats the vendor states

Field sorted: the number of distinct upload formats each card's input limits field states, spelled as the vendor spells them, compiled on 9 November 2026 from the vendor pages recorded on the cards (read 8 November 2026 for the first sixteen cards and 9 November 2026 for the twenty-two added that day). A token counts when it is written as an extension (.docx), in capitals (PDF) or inside a vendor's comma list (gif, jpg, png), so "upload doc (file types not published)" and "1 credit = 1 word" count nothing; jpg and jpeg count as two where both are listed, because the vendor listed both; "Word" counts once as a format when stated as one; zip and a CSV list of documents are containers, not formats, and programming languages are not formats. Direction: more ranks higher. Tie rule: more study records on the card first, then product name alphabetically, which is why Hive (1 record) sits above Reality Defender (0) at 15 and Turnitin (5) above AIDetectPlus, Quetext and Smodin at 4. Qualification rule: a vendor page that names at least one accepted format; a card whose pages name none is listed apart with its own wording. Grammarly Authorship's "Microsoft Word document" is where the record is kept, not an upload, so that card states no format. 19 of 38 qualified on 9 November 2026; 20 of 72 qualify after the fifth wave, Resemble Detect at 17 formats (19 from 10 November 2026, when the amr and 3gp tokens were counted under the stated rule; Ircam Amplify moved from 7 to 9 the same day for AIF and AIFF), with Codeleaks, Illuminarty and Pindrop Pulse stating none. Full rules: methodology page; no record in the studies index tests file handling, and the input limits ranking reads the same field for size ceilings.

The same vendor can state different formats on different surfaces. AI or Not's web tool takes MP3, MP4, PNG, JPG, WEBP or GIF and documents in DOC, DOCX or TXT, while its API takes images in jpg, jpeg, png, webp, heic, heif or tiff and audio "in common formats such as MP3 and WAV" (docs.aiornot.com, read 9 November 2026); the count takes every format either surface names once. Sapling's web tool takes "PDF or Word document with selectable text" and its API file endpoint ".pdf, .docx, .pptx, .odt, .txt, .md, .html or .htm" (sapling.ai, read 8 November 2026), so PDF counts once though it appears twice and Word and docx count separately because the vendor wrote both. Copyleaks lists "OCR files: gif, png, bmp, jpg, jpeg" and "Image files: JPG, JPEG, PNG, WEBP, BMP, HEIC, HEIF" (copyleaks.com, read 8 November 2026); the overlap is counted once, giving 8. A stated format says the upload is accepted, not that the detector reads it well: Turnitin excludes lists, tables, references and code from scoring inside the four formats it accepts (guides.turnitin.com, read 8 November 2026).

A Google Doc reaches most of these detectors by download and upload rather than by link: no card states that a docs.google.com share link can be pasted in, as the Google Docs question, added 9 November 2026, records.

A PDF in this table is a container: 18 cards accept one, four say the file needs a text layer or typed text, and none states OCR of a scanned page, by the scanned PDF question read 10 November 2026; the handwriting and OCR view lists the three cards that state OCR of handwriting or image files instead.