Subject

How we test AI detectors

Samples

180

Latest run

RUN-2026-11-A

Method changed

9 November 2026

Every detector reads the same 180 samples: 100 written by people and 80 produced by AI, 20 of those edited by a person before scoring.

Method last changed: 9 November 2026. Page updated: November 9, 2026.

Sample pages sorted into five tabbed piles beside a chart recorder

What is in the sample set?

One square per sample. Green family: written by people. Blue family: produced by AI.

100 human samples · 80 AI samplesEvery detector reads all 180, in the same run.

The sample set by subset, with count, source and reason
Subset	Count	Source	Why it is in the set
Written by people			100
Student essays	60	PERSUADE 2.0 student essays, grades 6 to 12 (The Learning Agency, CC BY 4.0)	The writing teachers check most. Most readers are marking or submitting essays.
Non-native English essays	20	English-learner essays from licensed public corpora, credited in the manifest	Students writing in a second language are flagged in error more often; this subset counts how often.
Professional prose	20	Published articles under CC BY, written before 2022	Polished, formal writing by editors and writers, the kind a client sends with a screenshot.
Produced by AI			80
Raw AI text	60	Written for the same prompts by three current chat assistants, versions and dates in the manifest	AI output used as is. A detector that misses this is not doing its job.
Human-edited AI text	20	Raw AI texts revised by an editor, about a third of the sentences rewritten by hand, edit log kept	AI output a person reworded before scoring. Closer to what a detector meets in real use.
Total	180		The same 180 for every detector in a run.

What we record for each sample

Every reading becomes one row in the run file: one sample, one detector, one result. Nothing is averaged before it is written down.

The same row is what you download from test runs, so any count on the site can be rebuilt from the file.

Example values show the format only.
Field	Example	What it holds
run_id	RUN-2026-11-A	The run this reading belongs to
sample_id	H-002	Which of the 180 samples
subset	student_essay	One of the five subsets
words	000	Length of the sample as scored
detector	gptzero	The tool that read it
detector_version	2026-11-04-base	Version or the date the tool was used
label	human	The tool's own verdict, as shown
score	0.00	The tool's number, as shown
flagged	false	Our flag rule applied to the reading
read_at	2026-00-00	Date of the reading

How is a flag counted?

A sample is flagged when the detector labels it AI or mostly AI, or when it shows only a score and that score is at or above 50%.

A flag on a human sample is a false alarm. A flag on an AI sample is a catch. Mixed labels count as flagged on human samples and are also reported on their own.

The same rule applies to every detector in a run, and it is written into the run file next to each reading.

Flag line human, under the line: clear human, over the line: false alarm AI, over the line: a catch

Human sampleNot flagged

Human sampleFlagged: counted as a false alarm

AI sampleFlagged: counted as a catch

Metrics, as counts

Every metric is a count out of a fixed base, never a bare percentage, so you can always see how many samples it rests on.

False alarms on human writing

of 100human samples

Human samples the detector flagged.

Written as: 00 of 100

Detection on raw AI text

of 60raw AI samples

Raw AI samples the detector flagged.

Written as: 00 of 60

Detection on edited AI text

of 20human-edited AI samples

Edited AI samples the detector flagged.

Written as: 00 of 20

Access and price

of 8plan checks

Free plan, limits, sign up and price, each checked on a stated date.

Written as: 0 of 0

Scoring weights

Fixed for every detector and every page. Job pages change which samples count, never the weights.

Fixed weights. Together they make the overall score.
Factor	Weight	What it measures
False alarms on human writing	40%	Human samples scored as AI, out of 100
Detection on raw AI text	30%	Unedited AI samples scored as AI, out of 60
Detection on human-edited AI text	15%	AI samples a person revised, scored as AI, out of 20
Access and price	15%	Plan checks passed, each with a date

Run cadence and dating

A new run starts when a detector on the list ships a new model version, and every run carries its own ID and date. Earlier runs stay published.

Read the editorial policy for when a new run is started.

Every run has an ID and a dateRun IDs look like RUN-2026-11-A. The date is the day the last reading was taken.
Every page names its runThe chart header and the dates line on each page cite the run its numbers come from.
Prices carry their own datePrices and plan limits are checked separately and show a "checked on" date next to them.
Old runs stay upEarlier runs and their files stay on the test runs page, so a past citation still resolves.

What the sample is

The properties of the 180 samples as a set. Per sample details are in the manifest file below.

Samples: 180
Human: 100
AI: 80
Language: English
Years written: Before 2022 for human samples
Length: 300 to 900 words
Human-edited AI: 20 samples revised by a person
AI generators: Three current chat assistants, named with versions in the manifest
Consent: Human samples come from published corpora released for reuse, each credited in the manifest

Raw files

Download the files behind run RUN-2026-11-A. Every number on the site can be rebuilt from them.

ReadingsOne row per sample per detector, with the fields aboveCSVDownload the readings
Sample manifestThe 180 samples: ID, subset, length, source noteCSVDownload the manifest
ScoresThe weighted scores per detector, overall and per jobCSVDownload the scores
Plan checksPrice and plan limits with the date each was checkedCSVDownload the plan checks

Every run and how to cite it

Found an error?

Readers and detector companies use the same route.

Send it with the IDs
Use the contact page. Include the page, the run ID and the sample IDs, and what you saw.
We check it against the file
We rerun the reading or recheck the plan, and compare it with the row in the run file.
The outcome is logged
Every change, and every dispute we reject, goes on the corrections log with the date and the evidence.

Questions about the method

Why do false alarms carry the most weight?

A false alarm lands on a real person: a student, a writer, a client. Missing some AI text is a smaller harm than accusing someone of using it, so false alarms on human writing carry 40% of every score.

Why include essays by non-native English writers?

Because detectors flag them more. Liang and colleagues (2023) found seven detectors misclassified TOEFL essays far more often than essays by US eighth graders, so our set counts that subset on its own.

Why test AI text a person has edited?

Because that is how AI text usually arrives. In Weber-Wulff and colleagues (2023), accuracy fell from 74% on unedited AI text to 42% on AI text with human edits.

What counts as a flag?

A sample is flagged when the detector's own label says AI or mostly AI, or when it shows only a score and that score is 50% AI or more. Mixed labels count as flagged on human samples, because that is what a teacher sees, and are also reported separately.

Do you use each tool's paid plan?

Where the free plan cannot take all 180 samples, yes. The plan used for each detector is recorded in the run file next to its readings.

Do you test AI image detectors?

Yes, with a separate sample set of real photos, real artwork and AI images, scored with the same flag rule and weights: 40 real photos, 20 real artworks, 30 raw AI images and 10 AI images edited after generation. That run is in progress; the pick is on the best AI image detector page.

Can I check your numbers?

Yes. Every run publishes its readings, sample manifest, scores and plan checks as CSV files on the test runs page, and corrections are logged with their evidence.