Method

Ranking criteria and test method

Five criteria anyone can check on a vendor's own site, a fixed formula, and text sets to test accuracy yourself.

01 · Criteria

The rubric

Each criterion scores 0, 1 or 2. Machine-readable: /rubric.json

CriterionWeight012
Published evidenceDoes the vendor publish how it tests accuracy, including its false-positive rate?×20No accuracy claim, or none you can trace1A headline accuracy claim without a method2A published method or paper with a false-positive rate
Explains its scoreDoes it show which passages drove the score, or only a single number?×10A single score1A score plus a general label2Sentence or passage highlighting
Free to testCan you run your own test before you pay?×10No free access for individuals1A trial or one-time credits2A free tier that renews
Public priceIs the price on a public page, or behind a sales call?×10Sales or institution only1Partly public2Plans and prices on a public page
Fits a workflowCan it run where the work is done: an API, an LMS, a docs plugin, bulk scans?×10Paste-in box only1One integration or an API2An API plus team, bulk or LMS options

02 · Text sets

What to test with

Same topics across sets, 300–800 words each.

SET H10Human, before 2022Written and published before November 2022, when ChatGPT launched.
SET N5Human, second languageBy writers whose first language is not English, shared with permission.
SET A10AI, uneditedFrom at least two models, same topics as set H, no edits.
SET E5AI, edited by a personSet A texts edited by a person for ten minutes or more.
SET P5AI, paraphrasedSet A texts run through a paraphrasing tool once.

03 · Steps

Reproduce the ranking

  1. Fix the list

    Detectors with a web product you can buy or access that returns an AI score for English text.

    Included on the review date: 8 detectors. Excluded: research demos and tools without a public product page.

  2. Score each criterion from public pages

    Read the vendor's own pricing, product and research pages. Record each score as 0, 1 or 2 with the page it came from.

    Source of truth: the vendor's site on the review date. Each profile lists the pages read.

  3. Weight and total

    Evidence counts double; the rest count once. The most a detector can score is 12.

    total = 2·evidence + explains + free + price + workflow

  4. Break ties the same way every time

    More evidence first, then freer access, then alphabetical order.

    sort by total desc, evidence desc, free desc, name asc

  5. Test accuracy on text sets, separately

    Accuracy is never folded into the rank: it shifts with each model update and each kind of writing. Run the sets below on your own text.

    Protocol: /test-your-detector

  6. Re-review

    Prices, plans and published claims are re-checked at every review, and the date at the top of the ranking changes with it.

    Last reviewed: 19 November 2026

04 · Scores on 19 November 2026

Every score, with its source

#DetectorEvidence×2ExplainsFree testPriceWorkflowTotalCost
01CopyleaksSchoolsPublished testing method with false-positive ratesHighlights AI-written phrases, sentence by sentenceFree checks, 25,000 characters without loginPersonal and Pro priced on copyleaks.com/pricingDocs add-on, extension, MCP; API and LMS by quote12/12Personal $16.99/month ($13.99 billed yearly). Pro $99.99/month ($74.99 billed yearly). Education and Enterprise by quote. Check
02GPTZeroTeachersPublishes benchmarks with its false-positive rateSentence-level highlightingFree plan, 10,000 words a monthPlans priced on gptzero.me/pricingAPI, Chrome extension, team plans, LMS12/12Free: 10,000 words a month. Premium $12.99/month, Professional $24.99/month (billed yearly). Check
03Originality.aiSEO publishersPublishes accuracy data and false-positive ratesSentence-level precision on every plan10,000 words, one timeAll plans priced on originality.ai/pricingBulk scan, team seats, API on Enterprise11/12Free: 10,000 words, one time. Pro $14.95/month ($12.95 billed yearly). Enterprise $179/month. Check
04PangramEditorsModel card and stated false-positive rateHighlights the sections it reads as AI7-day free trialPlans and API rates on pangram.com/pricingAPI, MCP, Docs and Gmail, team plans11/12Individual $20/month after a 7-day free trial. Professional $65/month. API $0.05 per 100 words. Check
05SaplingWritersStates false positives occur; no published rateScores each sentenceFree, 5,000 characters per checkPlans priced on sapling.ai/pricingMetered API, browser extension, team plan10/12Free: checks up to 5,000 characters each. Pro $25/month ($12 billed yearly). API metered, $5 minimum. Check
06ZeroGPTQuick checksAccuracy claims; no published false-positive rateSentence highlightingFree tier, $0Plans priced on zerogpt.com/pricingPay-as-you-go API9/12Free tier, $0. Personal plans from $14.99/month. API $0.042 per 1,000 words. Check
07Winston AIScans & OCRHeadline accuracy claimsPrediction map, sentence by sentence14-day trial, 2,000 creditsPlans on gowinston.ai/pricingAPI, team members, PDF reports9/12Free 14-day trial (2,000 credits). Essential listed at $18/month billed monthly; promotions change it. Check
08TurnitinUniversitiesPublishes its false-positive rate and how it holds itHighlights text in the reportNo individual accessNo public priceMoodle, Blackboard, Canvas and other LMSs8/12Sold to institutions. No price on its public pages: ask your institution or Turnitin. Check