How we index AI detectors
AI Detector Index records 72 AI detectors on 14 identical fields and 35 published evaluations of them. Every price, limit and accuracy figure on the site names its source and the date it was read, and the ranking pages are computed from that data, not written by hand.
What this index is, and what it is not
The index is a registry. Each detector gets one spec card with the same fields in the same order under the same source rules. It is not a review site: there are no star ratings, no "best overall" verdict and no editor score.
Rankings are derived. Each ranking sorts the cards on one field that has a published value: study count, published false positive rate, free tier size or language count. A tool with no published value for that field is left out, and the page says how many of the 72 tools qualified. All 72 cards sit in one hub, the five rankings in another and the thirty-five study records in a third. The AI Detector Index home table shows every card in one view.
The 14 fields on every card
| Field | What it records | Accepted source | Next recheck |
|---|---|---|---|
| 1. Name, vendor, official domain | Domain confirmed first | Vendor site | 8 December 2026 |
| 2. Modalities detected | Text, code, image, voice, video, music | Vendor product page | 8 December 2026 |
| 3. Generators covered | As stated, with date read | Vendor product page | 8 December 2026 |
| 4. Languages supported | Vendor list, with count | Vendor product page | 8 December 2026 |
| 5. Input limits | Words or characters per check, file types | Vendor help page | 8 December 2026 |
| 6. Free tier | What, how much, total or per period | Vendor pricing page | 8 December 2026 |
| 7. Paid plans | Currency, billing period, date checked | Vendor pricing page | 8 December 2026 |
| 8. API | Yes or no, docs link | Vendor docs | 8 December 2026 |
| 9. LMS and platform integrations | Canvas, Moodle, Blackboard, Google Docs, Chrome | Vendor integrations page | 8 December 2026 |
| 10. Vendor accuracy and false positive claims | Exact sentence, URL, date | Vendor page | 8 December 2026 |
| 11. Independent studies | Each result as reported, with date | The study itself | When a study is added |
| 12. Data retention and privacy statement | Policy URL, date | Vendor policy page | 8 December 2026 |
| 13. Launch year and ownership | Year and owner | Vendor about page | 8 December 2026 |
| 14. Rankings the card appears in | Computed from fields 4, 6, 10, 11 | The index data file | Every build |
When a vendor publishes no value for a field, the card leaves that field out. No field is ever filled with an estimate.
What counts as a source
Two source classes appear on every page and are kept apart. Vendor statements come from product, pricing, help and policy pages and are attributed in the sentence ("GPTZero says"). Independent studies are peer-reviewed papers, published benchmarks and lab reports with a method and a dataset. A news write-up is never the only citation: the editors read the study itself and quote the figure with the base it uses.
Every number sits in the same paragraph as its link, and prices and plan limits carry the date they were checked. The index data file is canonical: a tool's figure is identical on its card, in every ranking and in the page schema. If two studies disagree, both appear with their sample sizes and dates; the site does not average them.
The index uses published figures only. A logged test added later would appear as its own study page with inputs, outputs, dates and limits.
How the Studies Index works
The Studies Index holds one page per published evaluation: who ran it, when, which detectors, sample sizes and text types, results exactly as reported in the metric the study used (accuracy, false positive rate, false negative rate or AUROC), stated limits, and the DOI or publisher link. The inclusion rule is a method and a dataset described well enough to say what was measured. Vendor-run reports are indexed and labelled as vendor reports.
Each study page lists the cards it feeds, and each card lists the studies that tested it. At version 2026.11.2, GPTZero appeared in 7 of the 9 study records, ZeroGPT in 5, Turnitin and Copyleaks in 4 each, Originality.ai in 3, Winston AI and Sapling in 2 each, Scribbr in 1, and eight cards in none. At version 2026.11.3, after the eight records added on 9 November 2026, GPTZero appears in 11 of 35, the GPT-2 Output Detector in 9, Copyleaks in 6, Turnitin, Originality.ai and ZeroGPT in 5 each, GLTR in 3, Pangram, Winston AI and Sapling in 2 each, Hive and Scribbr in 1, and fourteen cards in none. At version 2026.11.4, after the six cards of the third wave on 9 November 2026, Writer appears in 4 records and AI or Not in 1 (as Optic), and 37 cards have none. Those counts are the rows of the independently tested ranking.
How the rankings are derived
| Ranking | Field sorted | Direction | Qualification | Tie rule |
|---|---|---|---|---|
| Independently tested | Count of indexed studies that tested the tool | Higher first | One indexed study | Most recent study date |
| False positive rate | Lowest published rate on human-written text from an indexed study, sample size shown | Lower first | A study reporting the rate | Larger sample size |
| Free tiers | Words or characters free per month, as the vendor states | Higher first | A stated free allowance | Higher per-check limit |
| Languages | Count of languages the vendor lists | Higher first | A stated language list | Earlier launch year |
| Input limits | Largest words or characters per check the vendor states for any plan, characters shown beside a conversion at 5.5 characters a word, the one rate every page of the index uses since 9 November 2026 | Higher first | A stated numeric maximum per check; an unlimited tier is listed apart | Web tool before API, then a figure stated in words before one converted from characters |
| Data retention | Days the vendor's policy says submitted text or images are kept, 0 where the policy says they are not stored or are deleted after analysis | Lower first | A policy that states a period or states that submissions are not stored; "as long as necessary" or an open-ended account history is listed apart | More study records first, then name |
| Studies by content type | Count of indexed studies that scored the tool on image, voice, video, code or music | Higher first | A card that states at least one content type other than text | More study records in all first, then name |
| Adversarial robustness | Highest share of paraphrased, humanised or adversarially edited AI text the tool still classified as AI in an indexed study, as the paper prints it | Higher first | A study that prints a detection or accuracy figure on an attacked AI set; false negative rates after an attack are not inverted | More study records first, then name |
| False negative rate | Lowest share of AI-generated text an indexed study prints as missed, undetected or labelled human, with the false positive rate from the same study beside it | Lower first | A study that prints the rate or a count of missed AI texts with its base; a complement of a sensitivity or "percent correct" figure is not used | More study records first, then name |
| Claim vs evidence gap | The vendor's headline accuracy percentage minus the highest accuracy an indexed study prints for the tool on AI or mixed text, in points; where a card's only record scored images and the claim covers images, the image accuracy is paired and the row says so | Lower first (a negative gap means the claim sits below the best result) | A numeric accuracy claim about the vendor's own tool and at least one indexed study that prints an accuracy figure for it | More study records first, then name |
| Platform integrations | Count of the eight named integration surfaces the card's integrations list states: Chrome or browser extension, Google Docs, Microsoft Word, an LMS (Canvas, Moodle, Blackboard, D2L, Schoology, Google Classroom, Sakai, Edsby), Zapier, MCP, a documented REST API, a mobile app | Higher first | At least one of the eight surfaces named on a vendor page; an entry the vendor's own words withhold from the detector does not count | More study records first, then name |
| File types | Count of distinct upload formats the card's input limits text states, spelled as the vendor spells them (jpg and jpeg count as two where both are listed; Word counts once) | Higher first | A vendor page that names at least one accepted format; a card whose pages name none is listed apart | More study records first, then name |
| Minimum input | Smallest text length the vendor states as a requirement for a check ("minimum", "at least", "must", or a result returned only above the figure), normalised to characters at 5.5 characters a word for ordering only, the vendor's unit printed beside it; a figure the vendor calls recommended, guidance or a tip is tabled apart | Lower first | A vendor page that states a numeric minimum as a requirement; a card that states only a recommendation or no floor is listed apart | More study records first, then name |
| API pricing | Published price of a text detection API per 1,000 words in US dollars, converted from the vendor's unit: a per-character rate at 5.5 characters a word, a credit price with a stated words-per-credit ratio, or a plan price divided by its stated monthly word quota; the lowest-priced tier that includes text detection is sorted | Lower first | A vendor page that prints a price and a unit for text detection that convert to words; per-image, per-second or per-result prices and "contact sales" are tabled apart | More study records first, then name |
| LMS coverage | Count of named LMS platforms in the card's integrations list, of nine: Canvas, Moodle, Blackboard, D2L or Brightspace (one platform), Schoology, Google Classroom, Open LMS, Microsoft Teams where listed as an assignment integration, Cineca; "LTI" and "LMS via API" name no platform | Higher first | At least one of the nine named on a vendor page; a card that states an LMS route with no named platform is listed apart | More study records first, then name |
| Free checks per day | The free allowance the vendor states with a per-day unit, counted in checks, scans, credits or operations where one unit is one submission; several daily quotas on one page are summed and each printed | Higher first | A vendor page that states a per-day free figure in checks; a per-day figure in characters, a per-month figure, a per-check limit, a one-time trial or an unlimited tier with no figure is tabled apart | More study records first, then name |
| Generators covered | Count of distinct generator models, families or products the vendor names as detected or trained against, each listed item once; a company counts once only where none of its models is named; catch-all phrases and headline counts ("35+ LLM families", "500+ generators") are printed and not counted | Higher first | A generators field that names at least one model; a card whose field names none is tabled apart | More study records first, then name |
| Languages evidence | Count of indexed study records that print a result for the card on non-English text, machine-translated text or English by non-native writers; vendor multilingual figures excluded | Higher first | At least one such record; a record that names a language without scoring the card on it does not count | More study records in all first, then name |
| Code languages | Count of programming languages the vendor names for code detection on a page recorded on the card, or in a vendor test it reports; headline counts and names on archived pages the live page dropped are printed and not counted | Higher first | A card whose modalities list code and whose pages name at least one language; a code card naming none is listed apart, and cards with no code modality are listed with the modalities they state | More study records first, then name |
| Launch year | The year the vendor page, press release or paper the card cites states the detector launched; a company founding year stands in only where the vendor dates nothing else, and a Wayback Machine capture date is never entered | Earlier first | A launch year field that is not empty; a card with none is listed apart with what its pages say and any withdrawal or change of hands the card records | More study records first, then name |
Vendor-stated false positive rates sit in a separate column on that ranking and never enter the sort. The five rankings added on 9 November 2026 (data retention, studies by content type, adversarial robustness, false negative rate and the claim vs evidence gap) read fields written into the data file on 9 November 2026 from the eight study records added the same day; each field carries the study slug, the base and the threshold beside its value, and the ranking page quotes all three. Total rather than monthly allowances are flagged in the row, and the languages ranking links a study on non-English text where one exists. Ranks, dates and arrows are computed from the index data file at build time, so a corrected figure moves a row the same day. The two rankings added in the fourth wave on 9 November 2026 (platform integrations and file types) read two fields the merge script derives from every card's integrations list and input limits text under the rules in the table, so a card enters either ranking the day its integrations or input limits field is read. The two rankings added in the sixth wave on 9 November 2026 (minimum input and API pricing) read two fields the wave 6 merge script writes from every card's input limits text and plans and API fields under the rules in the table; both sort lower first, both print the vendor's own unit beside the converted figure, and the API pricing page describes its order as the lowest published price and never names a cheapest tool. The ranking added in the eleventh wave on 10 November 2026 (launch year) reads the launch year field every card has carried since launch, so no new field was written; the wave 11 merge script added the field's sort direction to the metrics table.
Walter Writes and Proofademic on this site
Walter Writes and Proofademic appear as cards with the same 14 fields and the same source rules as every other tool, and in a ranking only where a published value qualifies them. Links to their product pages are tagged like any outbound product link.
Dates, updates and the corrections log
Each page shows its own "Updated" date, and the date moves only when a figure or sentence changes. The index version number changes when a card or study is added; version 2026.11 was created on 8 November 2026. Recheck plans are stated with a date in the revision log; the current plan is prices and free tiers on all cards rechecked on 8 December 2026.
The corrections log records every change to a published number: what changed, from what to what, why, and the source. The contact page explains how to file one.
Limits of this index
A detector with no indexed study has no evidence row and sits outside the evidence rankings. Vendors change plans and models without notice; the date beside each figure is the reader's guard. Studies test specific versions on specific texts, so a 2023 result describes that version; the study pages name the version where the paper does.
No detector score should be the only evidence in a misconduct, hiring or publishing decision. The results as evidence question, added 9 November 2026, lists the seven primary documents in the index that say the same and the 24 vendor caveats that say it in the vendor's words.
Questions about the method
How many AI detectors are in the index?
The index holds seventy-two cards at version 2026.11.13: sixteen were added on 8 November 2026 and twelve on 9 November 2026, in two waves of six. A card is added once the editors have confirmed the vendor's domain and filled every field the vendor publishes; each addition raises the index version and is logged.
Does AI Detector Index test detectors itself?
The index uses published figures only, from vendor pages and the 9 indexed studies. A logged test added later would appear as its own study page with inputs, outputs, dates and limits, and feed the cards like any other study.
Why does a detector have no accuracy figure on its card?
Because neither the vendor nor an indexed study has published one. 37 of the 72 cards have no study record. A field with no published value is left out, never estimated.
How are the rankings calculated?
Each ranking sorts one field from the index data file: study count, published false positive rate, free tier size or language count. Only tools with a published value qualify, and the page states how many of the 72 did.
How do I report an error or a new study?
Email the editors through the contact page with the page, the field and your source URL. Every accepted change is written to the corrections log with the old and new value and the source.
Are Walter Writes and Proofademic treated differently?
No. Both cards carry the same 14 fields, cite the same source classes and enter a ranking only where a published value qualifies them, the rule applied to the other 20 cards.
The editorial policy sets the sourcing rules; the corrections log records every change; the Studies Index holds the thirty-five records; the independently tested ranking is computed from them; and all 72 cards sit in the detectors hub.