AI detector rankings

Nineteen AI detector rankings are derived from the index, and each one sorts the seventy-two cards on a single field that has a published value: the number of study records that tested the tool, the lowest false positive rate a study reports on human-written text, the words a free tier allows per month, the number of languages the vendor states, the largest input per check, and, since 9 November 2026, the stated retention period for pasted text, the study records on image, voice, video, code or music, the share of paraphrased AI text still caught, the lowest published false negative rate, the gap between a vendor's accuracy claim and the best published result, and, since the fourth wave of cards later on 9 November 2026, the number of named integration surfaces (Chrome extension, Google Docs, Word, LMS, Zapier, MCP, API, mobile app) and the number of upload formats the vendor states. There is no overall ranking, no editor score and no "best" or "most accurate" verdict anywhere on this site. A ranking page states the field it sorts, the direction, the tie rule, the qualification rule and how many of the twenty-two cards qualified, and the methodology page holds the rules in one table. The home board shows the AI detector index the rankings are cut from.

A card enters a ranking only when the field has a published value. On 9 November 2026 ten of twenty-two cards qualified for the ranking by independent testing, because ten have at least one study record in the studies index; the same ten qualified for the false positive rate ranking, because each has at least one study that reports a rate on human-written text. Only three cards qualified for the free tier ranking, because only three vendors state a monthly figure: ZeroGPT at 1,250,000 words, GPTZero at 10,000 words and Turnitin at zero, since its detection is reached by institutional licence alone. The other nineteen free tiers are stated per scan, per day, per image, per check or as a one-time total, and the index lists them in the vendor's own words rather than converting them. 31 cards qualified for the languages ranking; the ten whose vendors publish no natural language count are named on the page as unranked. The five rankings added on 9 November 2026 qualify 5 cards (data retention), 14 (studies by content type), 8 (adversarial robustness), 10 (false negative rate) and 7 (claim vs evidence gap), each page stating why the rest fall out.

Position in a ranking is a count or a rate, never a judgement. GPTZero ranks first by independent testing because eleven of the 35 study records name it (seven of nine on the morning of 9 November 2026, before eight records were added), and the same records include the highest false accusation count in Weber-Wulff 2023, nine of eighteen human-written documents (DOI 10.1007/s40979-023-00146-z), and a 0.00% false positive rate at threshold 0.5 in the RAID benchmark (DOI 10.18653/v1/2024.acl-long.674). Both figures sit on the GPTZero card with their base. A reader who takes the rank without the record will misread it, so every ranking page repeats the base and the study year beside each value and links the record.

Ties are broken by a rule stated on each page. In the ranking by independent testing, cards with the same count are ordered by the year of their most recent study, newest first, then by name. In the other three, cards with the same value are ordered by how many study records name them, then by name. Rank, value, date and chart direction are computed from the index data file at build time, and each metric in that file declares whether higher or lower is better, so a ranking cannot drift from the cards it is cut from. When a card's field changes, the ranking changes with it and the change is written to the corrections log.

The table below lists the twenty rankings with the field each sorts, the direction and how many cards qualified. Vendor-stated figures never enter a sort beside study results: Proofademic's and Codequiry's false positive claims, for example, are shown on the false positive rate page in their own table, attributed to the vendor, and are not ranked, and Pangram's sat there until a study-sourced rate replaced it on 9 November 2026. The one ranking that reads vendor claims, the claim vs evidence gap, sorts the distance between the claim and a study figure, never the claim alone. The use-case views apply the same fields to a reader's situation rather than ordering the cards.