Evidence: the sealed test and the model card

Published by AI Detector Checker; sealed test scored once on 17 October 2026; engine V4 release rel-v4-engine-2026-10-17-a194a62bf42a. Every number below carries its count and its Wilson 95% interval, and every table was generated from the sealed result file, never typed.

The sealed test: V4 against V3, by length band

What was measured. 12,685 texts were scored once by the frozen V4 engine and, on the same texts, by V3, the previous engine, with its released thresholds: 4,000 human texts from 8 source families the model never trained on, all dated 2021 or earlier; 6,113 AI texts from 9 model families plus two families never trained on; and two confirm halves (non-native English and modern formal speech). A human text counts as wrongly flagged when its score reaches 65 or more; an AI text counts as caught by the same rule.

What it means. At every length V4 flags fewer human texts than V3 and catches far more AI text. The shortest band remains the weakest: about 1 in 100 human texts of 100 to 149 words reached 65 or more, against about 1 in 700 at 150 words and above. The result card states the figure for the length of your text on every check.

AI texts caught, by model family

Texts of 150 words or more. The family rows combine every tested version; the version table below lists each one. The detector never identifies which model wrote a text; these rows describe how often texts known to come from each family were caught.

By model version

The exact versions in the sealed test, with the number of texts at 150 words or more and at 100 to 149 words, and the share caught by V4 wherever a version has at least 30 texts in that band. Texts were generated between 2026-10-10 and 2026-10-13 from 35 task types in 6 styles through each provider’s official API, except the ChatGPT texts and part of the Claude texts, which the publisher produced in the providers’ own apps and uploaded.

Models the engine never trained on

Two open-weights families were held out of training entirely and scored in the sealed test, so that the published result includes a genuine unseen-model measurement. The share caught across all lengths: gpt-oss 98.50% [97.53, 99.09] of 998 texts (V3: 16.53%); Gemma 98.55% [92.24, 99.74] of 69 texts (V3: 27.54%). The Gemma sample is small, which its wide interval shows; it is reported, not generalised. These two families are not a measurement of every model the engine has not seen.

Human writing the engine never trained on

The 8 sealed human families, with the share wrongly flagged (score 65 or more) at 150 words or more and across all lengths. All are dated 2021 or earlier by provenance.

Non-native English and modern formal writing

Two confirm halves, exempt from the date rule because they test exactly the writing a detector is most likely to misread: learner English, and formal set-piece speech from 2024 to 2025. Both were scored once with the sealed sets.

Development versus sealed

The thresholds and the score map were chosen on the development data. A sealed result that is worse than the development result is the honest measure of how much that choice flattered the engine. At 150 words or more, development recall was 99.39% (n=3279) against 98.97% sealed; at 100 to 149 words it was 98.83% (n=766) against 96.94% sealed, the largest gap in the test (recall 100-149: -1.89 pp). Human false flags at 150 words or more went the other way: 0.30% (n=40797) in development against 0.14% sealed. The complete side-by-side table is in the test record.

Text attacks (development data)

Ten deterministic attacks were applied to the development texts to measure evasion: the share of AI texts still caught after the attack, and the share of human texts wrongly flagged. These were measured on development data by the frozen model, not in the sealed test, and are not broken down by model family.

Student-genre diagnostic (not part of the sealed test)

Diagnostic, not part of the sealed test. Before the release, three further sets of human writing were scored to probe the genres students and teachers submit most: learner essays from two further corpora (2,543 texts, of which only 80 unique essays came from one of them), forum posts by students asking for feedback on their essays (2,000 texts, from a dataset without a licence, so reported for information only) and book summaries (1,500 texts, overlapping a source family in training). Wrongly flagged at 150 words or more: learner essays 1.80% [1.33, 2.44] (V3 1.06%); student forum posts 0.05% (V3 1.08%); book summaries 0.89% (V3 1.93%). No Arabic-first-language learner data could be found for any of these.

Model card: V4

Intended use, and out of scope

AI Detector Checker is intended only for low-stakes review of your own text, or text submitted with the writer’s informed consent: self-review of drafts, editorial review of consented text, and detector research that does not evaluate, rank, accuse, penalise or determine eligibility for any person. It must not be used for any academic, admissions, employment, disciplinary, legal, healthcare, credit, insurance, housing, immigration, eligibility or compliance decision about a person, nor to establish who wrote a text or which AI tool produced it: the engine performs no authorship or model attribution.

Limitations and false positives

A detector can help you review text more carefully, but it cannot reconstruct the writing process, prove intent, or be used to make, support or trigger a high-impact decision about any person. The figures on this page describe how often the engine was wrong on the kinds of writing it covered; they say nothing about kinds it did not cover, and they cannot see who typed a text, whether a draft was edited, or whether a person wrote to a template.

• False positives. Human writing does reach 65 or more: 4 of 2,863 sealed human texts at 150 words or more, 11 of 1,136 at 100 to 149 words, and more often on modern formal speech and non-native English. Short, formal, templated and translated writing is where it happens.
• Missed detections. About 1 in 100 AI texts of 150 words or more scored below 65, and edited or paraphrased machine text, which was not measured, is likely to score lower still. A human-zone result is not a clearance.
• Mixed authorship. Text that a person drafted and a model polished, or that a model drafted and a person rewrote, has not been measured for V4. Expect the Uncertain zone, and expect it not to tell you which parts are which.
• Out of scope. Non-English text, source code, deliberately obfuscated text beyond the development attack battery, and the model families that were neither trained on nor measured.

If a result on your own writing looks wrong, the guide on what to do if your text was flagged explains what the flag does and does not mean and what evidence to keep. If a page here is wrong, write to us through the contact page; corrections are made on the page itself.

History: earlier engines, dated

V3 (5 September 2026 to 17 October 2026). Release rel-v3-stage1-2026-09-05-308f3a03fb05: the V2 fusion core (word, character and stylometric heads with a logistic fusion) with frozen length-banded thresholds and a frozen isotonic display map. Its own sealed confirmation, on a different sealed set: 4 of 2,343 human texts of 150 words or more wrongly flagged; 14 of 586 at 100 to 149 words (above its bar of 2 in 100); 11 of 229 issue-tracker texts (above its bar of 3 in 100); 221 of 400 texts from the two OpenAI models then tested caught; 251 of 602 from one unseen lineage. It was released under an owner override with both missed bars published. On the V4 sealed test, scored on the same texts as V4, its figures are in the tables above.

V2 (4 September 2026). Release rel-2026-09-04-c3caa5009358: a three-band flagger without a displayed score, using the same fusion core at a single threshold. Its sealed confirmation: 4 of 2,604 human texts of 100 words or more received the strong label; 77 of 775 AI texts from model families seen in training and 293 of 600 texts from one unseen lineage were flagged.

Previous-generation detector (27 August 2026). An earlier engine was evaluated on 880 English texts from the public RAID training subset (440 human, 440 AI) with a pre-specified protocol. Those figures describe a retired engine and are not comparable to the sealed test above. Reference: Dugan et al. (2024), RAID.

Every release is listed in the model update history with its release id. A result is never silently replaced after a detector change.

Frequently asked questions

How accurate is the detector?

It is not summarised by one accuracy figure. In the sealed test of 17 October 2026, human texts of 150 words or more from sources the engine had never seen reached 65 or more about 1 in 700 times (4 of 2,863); at 100 to 149 words about 1 in 100. About 99 in 100 texts from the tested current AI models were caught at 150 words or more. A text below 65 is not evidence that a person wrote it.

Which texts are hardest?

Short texts of 100 to 149 words, modern formal speeches, non-native English and plot summaries are the kinds of human writing most often wrongly flagged; the figures for each are in the tables on this page.

Is the score the chance that AI wrote the text?

No. It is an AI-writing signal score on a fixed map of the engine’s internal value, chosen so that 65 means the same thing at every length. It is not the share of the text written by AI, and it is not a probability.

Does this test cover ChatGPT, Claude or Gemini?

Yes, and six further families; the versions are listed above. The engine still never identifies which model wrote a text.

Can this result prove who wrote a text?

No. A result alone does not prove who wrote a text; combine it with context, provenance, revision history and human review.