RankDude

How I test AI detectors

Ten tests, the same for every detector, each graded from 1 to 5 on a fixed scale. The overall grade is their average plus my verdict, which can move it by up to one point. Every test is run by hand on an account I use, and the passages, the verdicts, the counts, screenshots and timings behind each grade go in the full test results published with the review.

Tested and written by Miles Foster, Owner and tester, RankDude.

Each detector gets the same corpus, built before any tool is opened: 50 AI-written passages of about 300 words each, five from each of ten prompts and spread across six models (GPT-5, GPT-3.5, Claude, Gemini, Llama and DeepSeek) at default settings; 50 human-written passages of the same length and genres, every one either written before 2022 so no AI could have made it or written by me, so a flag on them can only be a false positive; 20 of the AI passages put through three named humanisers and paraphrasers; 20 mixed passages, half an AI draft with human edits and half a human paragraph joined to an AI one; and 24 non-English passages, twelve AI and twelve human, across Spanish, French, German and Chinese. Every passage is labelled in advance as AI or human, so each verdict can be counted rather than admired.

Each detector is run on the best plan the account offers, on a free or trial account where one exists, at the default settings, with each passage pasted or uploaded once and the first verdict taken. Where a tool returns only a percentage rather than an AI-or-human call, a result counts as 'AI' at 50% or above and the threshold used is recorded; where it returns a label, the label counts. Nothing is graded from a vendor's accuracy claim, a benchmark table or a marketing page. Each review states how the account was obtained: used on a free tier, bought at the public price, or provided by the vendor. A vendor-provided account never changes a grade, and a vendor never sees a review before it is published.

How I rate

Every tool gets one score out of 5.

First, I run the same 10 tests on every tool and give each one a grade from 1 to 5. The average of the graded tests is the starting point; a test that cannot be graded is left out of it, and the review says how many tests the score is based on.

Then I add my verdict: what the tests cannot measure, such as how the tool holds up after the test weeks, how billing and cancelling work in practice, and how it compares with the others in the category. My verdict can move the score up or down by one point at most.

Every review shows how much my verdict changed the score, and why, in its scores table.

What a grade means

5Excellent
4Good
3Average
2Below average
1Poor

The final score is the average of the graded tests (ten when every test can be graded), to one decimal place, plus my verdict, which can move it by up to one point. If a test couldn't be finished properly it gets no grade, and the review says how many tests the score is based on. A detector with no file upload, no non-English support or no bulk mode is not penalised on a capability it does not claim; that test is scored on what it can attempt, with the count stated, or left out and noted. Some tests on a free plan use fewer passages than usual because of daily or word limits; I run them again on a paid plan when I can.

The 10 tests

1.Accuracy on AI-Written Text

The 50 passages are generated before any detector is opened, five from each of ten prompts, at the default settings of six models (GPT-5, GPT-3.5, Claude, Gemini, Llama and DeepSeek), and are not edited or humanised. Each is submitted once and the first verdict taken: a passage counts as caught when the detector labels it AI, or, where it returns only a probability, when that probability is 50% or more, with the threshold recorded. A hedge that gives no usable call is a miss. Confident wrong verdicts and the models the detector is weakest on are reported, and the raw catch count out of 50 is published.

  • 5: 47 to 50 of 50 caught
  • 4: 42 to 46 caught
  • 3: 35 to 41 caught
  • 2: 25 to 34 caught
  • 1: fewer than 25 caught

2.False Positives on Human Text

This is the test that matters most, because a human wrongly accused is the real-world harm a detector does. The 50 passages are every one either written before 2022, when no public AI could have produced them (older essays, letters, book passages, forum posts), or written by me, so any AI flag is a false positive by definition. Each is submitted once and marked human or AI on the first verdict, with the same 50% threshold where only a probability is returned. The false-positive rate out of 50 is the grade, and each flagged passage is screenshotted with its score. Any false-positive rate above 10% scores 2 or lower on this test however well the detector does elsewhere, because a tool that accuses real writers cannot be trusted on the evidence it gives.

  • 5: at most 1 of 50 flagged (2% or lower)
  • 4: 2 or 3 flagged (4% to 6%)
  • 3: 4 or 5 flagged (8% to 10%)
  • 2: 6 to 10 flagged (12% to 20%)
  • 1: more than 10 flagged (over 20%)

3.Resistance to Humanisers & Paraphrasers

Twenty of the AI-written passages are rewritten by three tools that exist to defeat detectors (a humaniser, a paraphraser and a word-spinner, each named in the review on the test date). The best rewrite of the three per passage is kept, so the detector faces the hardest version, and each is submitted once. A passage counts as caught only when the detector still labels the rewrite AI at the 50% threshold. The catch rate out of 20 is the grade; which humaniser beat the detector most often, and whether the rewrite read naturally, are reported. A detector that catches clean AI but is beaten by a one-click humaniser is only as good as the easiest way round it.

  • 5: 16 to 20 of 20 still caught
  • 4: 13 to 15 caught
  • 3: 9 to 12 caught
  • 2: 5 to 8 caught
  • 1: 4 or fewer caught

4.Mixed & Lightly-Edited Text

Ten passages are an AI draft with real human edits (sentences reworded, facts added, order changed); ten are a human paragraph joined to an AI one in a single text. Each is labelled in advance as containing AI, and is submitted once. A passage counts when the detector flags it as containing AI rather than passing it clean; where the detector highlights sentences or gives a percentage mix, whether the highlight lands on the AI part is checked and reported. This is the everyday case a teacher or editor meets, and a detector that only works on fully AI or fully human text is marked down here.

  • 5: 17 to 20 of 20 flagged, highlights accurate where offered
  • 4: 14 to 16 flagged
  • 3: 10 to 13 flagged
  • 2: 6 to 9 flagged
  • 1: 5 or fewer flagged

5.AI Model Coverage

The AI set is built so that each of the six models (GPT-5, GPT-3.5, Claude, Gemini, Llama and DeepSeek) contributes the same number of passages. The catch rate from the accuracy test is split by model, and the grade is set by the weakest model, not the average, because a detector that is blind to one popular model will miss exactly the text a user pastes from it. The per-model table is published, the newest and the oldest models are called out, and any model the detector openly says it does not support is noted and left out of its own line rather than counted as a miss.

  • 5: every model caught at 90% or more
  • 4: every model at 80% or more
  • 3: every model at 65% or more, or one model below with the rest strong
  • 2: one model caught under half the time
  • 1: two or more models caught under half the time

6.Non-English Text

Three AI and three human passages in each of Spanish, French, German and Chinese make up the 24, every one labelled in advance. Each is submitted once and marked correct or wrong on the first verdict. Both error kinds are reported: AI passages missed, and human passages falsely flagged, which matters because detectors are known to accuse writers whose first language is not English. A detector that states it is English-only is tested in English alone, marked as not offering non-English, and this test is left out of its score with the fact stated rather than counted against it.

  • 5: 22 to 24 correct, at most one human passage flagged
  • 4: 19 to 21 correct
  • 3: 15 to 18 correct
  • 2: 10 to 14 correct, or human passages flagged across languages
  • 1: 9 or fewer correct, or English-only when it claims otherwise

7.Document & File Support

Each point is checked by doing it on the test date: paste a passage; upload a PDF; upload a DOCX; upload a plain-text or other common file; scan several documents in one batch; check a text by URL; a per-scan word or character cap generous enough for a full essay (2,000 words or more) on the plan tested; a browser extension or word-processor add-in; a documented API; and an export of the results to a file. A point counts only when it works end to end on the account used, not when it is merely listed on a plan I am not on, and each is screenshotted.

  • 5: all ten points
  • 4: 8 or 9
  • 3: 6 or 7
  • 2: 4 or 5
  • 1: 3 or fewer, or paste-only with a tight cap

8.Score Clarity & Evidence

Checked on the passages already run, each point is pass or fail: a plain AI-or-human call rather than only a bare number; a percentage or probability alongside it; a stated confidence band or margin; sentence-level highlighting of the parts it thinks are AI; a readable explanation a non-expert can follow; a shareable or downloadable report; a scan history kept on the account; the same text scored the same way twice in a row (repeatability, tested by scanning one passage three times); a clear statement of the detector's own limits; and no dark-pattern upsell blocking the result. Clear evidence is what separates a screening tool from an accusation, so vague percentages with no highlighting and no explanation score low here.

  • 5: all ten points
  • 4: 8 or 9
  • 3: 6 or 7
  • 2: 4 or 5
  • 1: 3 or fewer, or a bare number with no evidence

9.Scan Speed

All 50 scans in the AI-accuracy test are timed on the same connection, from pressing check until the verdict and score are shown; progress bars and queue messages do not count. The median is the grade; the average, fastest, slowest, any failed or timed-out scans, and the free-tier queue or slow-down when it differs from the paid tier are reported. Batch and file scans are timed separately and noted, since a long document naturally takes longer than a pasted paragraph.

  • 5: median under 5 seconds
  • 4: 5 to 9.9 seconds
  • 3: 10 to 19.9 seconds
  • 2: 20 to 44.9 seconds
  • 1: 45 seconds or longer, or scans time out

10.Pricing & Free-Tier Limits

On the free tier I check text until the detector stops me and record the real daily or per-scan word limit, whether an account is needed, whether the free tier runs the same model as the paid one, and whether file upload, bulk mode, highlighting and the API are free or paid. Every plan is then read from the checkout, not the marketing page, with the date: the monthly cost billed yearly, what the paid plan lifts or adds, any credit or per-word metering, and the cost of checking 1,000 words worked out in dollars on the cheapest plan that removes the free cap. Any pay-as-you-go or credit pricing is worked out as the cost of a typical day's checking.

  • 5: free tier of 1,000 words a day or more, or a paid plan under $10 a month that lifts limits and adds files and bulk
  • 4: a usable free tier and a paid plan $10 to $19.99
  • 3: a tight free tier, or a paid plan $20 to $29.99
  • 2: $30 or more, or no usable free tier
  • 1: poor value or unclear credit pricing

Documented but not graded

Plagiarism checking, a built-in humaniser or rewriter, a grammar and writing assistant, team and classroom dashboards, LMS and Google Classroom integrations, a learning-management or gradebook tie-in, writing-process or authorship replay, provenance and watermark reading, image and code detection, SSO and admin controls, and a mobile app are checked and described in every review. When they work well they count in the written assessment, but they never add points to the overall grade.

Rules that apply to every test

  • Real use of the product on a real account, never vendor accuracy claims, benchmark tables or promotional descriptions.
  • Identical passages, humanisers, languages, thresholds and sample sizes for every detector in the category, with the first verdict counted every time and the best model the plan offers.
  • Every human passage is one I wrote or one written before 2022, so a flag on it can only be a false positive, and no real person's work is submitted to accuse them.
  • Every grade has evidence: the passage, the verdict, a count, a measurement, a screenshot or a table, in the full test results.
  • A grade changes only when a retest produces new evidence; the review records the date, the model version and the reason.

Questions about a specific grade go to the contact page. Makers can submit an AI detector for testing. The tested ranking is on the AI detectors page.