AI detector features compared

AI detector features compared on this hub are five fields of the same seventy-two cards, cut into five slices: which detectors document an API, which name an integration with a learning platform, which content types each one states it detects, which publish their code or weights, and which publish a false positive rate of any kind. A slice lists every card that states the feature, quotes the vendor's wording with the source and the date it was read, and links the card. Nothing in a slice is inferred: "works with your LMS" does not count as an integration, and a modality counts only when the vendor states it on a product or documentation page. The fields are defined on the methodology page, every card is in the detector index, and the home board lets a reader compare them side by side.

The API slice held eleven cards on 8 November 2026: Winston AI, Pangram, GPTZero, Copyleaks, Phrasly, Sapling, Originality.ai, ZeroGPT, Proofademic, Walter Writes and Hive each publish API documentation, and the slice links each set of docs and records the input limits, the free quota where one is stated and the plan prices as checked. Five vendors, QuillBot, Grammarly, Scribbr, JustDone and Turnitin, publish no API for AI detection, and the slice says so rather than guessing. The LMS slice held seven cards: Winston AI, Pangram, GPTZero, Copyleaks, Originality.ai, Walter Writes and Turnitin each name at least one of Canvas, Moodle, Blackboard, D2L Brightspace, Schoology, Sakai, Edsby or Google Classroom, and the slice records how the vendor says the integration is delivered, because a native plugin, a browser extension inside Google Classroom and a report delivered through the platform are three different things.

The modalities slice is the widest. Twenty-six of the 72 cards state text detection. Fourteen state image detection, seven state video, six state voice, three state code and two state music, as each vendor words it on 9 November 2026; where a vendor qualifies a modality, for example image detection through an API only or a feature described as coming soon, the slice carries the qualification. Studies in the index tested text almost exclusively, so a stated image, voice or video capability on a card has no study record behind it, and the slice says that plainly rather than treating a vendor statement as evidence.

Three slices were added in the eighth wave on 9 November 2026, bringing the hub to eight: the no-account slice lists the 15 cards whose vendor states a check with no sign-up or login, with the limit that applies to it; the bulk and batch slice lists the 15 cards that state a bulk upload, batch API, ZIP or folder import, from GPTZero's 250 files to Sapling's 5; and the compliance slice lists the 10 cards whose vendor names a security or privacy standard on a page the index has read, each a vendor statement rather than a certificate the index has seen.

Two slices were added in the ninth wave on 10 November 2026, bringing the hub to ten: the model changelogs slice lists the 10 cards whose vendor publishes a dated release note, model card or version post, from Turnitin's model release notes to Compilatio's what's-new entry of 5 November 2026, with the latest dated entry read per row; and the institution-only slice lists the 5 cards sold only through a licence, demo or sales team (Turnitin, Turnitin Clarity, Compilatio Magister+, Pindrop Pulse, Sensity), how each grants access, and the 14 cards that sell both a self-serve plan and an institutional tier.

Features are not evidence of accuracy. A detector with an API, three platform integrations and six modalities may have no study record, and a text-only tool may have six. The ranking of independently tested detectors and the false positive rate ranking carry the evidence; the slices carry the fit. Weber-Wulff and colleagues tested fourteen detectors on English text only and left code out of their tests (DOI 10.1007/s40979-023-00146-z), and the RAID benchmark covers eleven text generators and eleven adversarial attacks on text (DOI 10.18653/v1/2024.acl-long.674); neither tested an image, voice or video detector, which is why those columns of the modalities matrix carry vendor statements alone.

The table below lists the five slices with their register IDs, how many cards each held and the fields that decide membership. The use-case views apply the same fields to a reader's job; the slices apply them to one feature at a time.

Three slices were added in the twelfth wave on 10 November 2026, bringing the hub to thirteen: the team seats and SSO view lists the 20 cards that state a seat count, a per-seat price or single sign-on (from a team field written to every card that day), the handwriting and OCR view the 3 cards that state OCR of handwriting or image files and the 8 that take an image without reading it, and the mobile apps view the 5 cards whose integrations name an iOS or Android app.