Evidence: the sealed test and the release decision

Published by AI Detector Checker; sealed test scored once on 5 September 2026; engine V3 release rel-v3-stage1-2026-09-05-308f3a03fb05.

Sealed test and release override

What we set out to meet. Before the sealed test, the release target for human writing was at most 2 in every 100 wrong AI signals on texts of 100 to 149 words, and at most 3 in every 100 for the hardest type of human writing in the test (the worst binding family). For longer texts the target was at most 1 in every 100 with an upper Wilson bound of 1.5 in 100.

What we measured. On human texts of 150 words or more the system gave a wrong AI signal for 4 of 2,343 texts, about 1.7 in every 1,000 (Wilson interval 0.7–4.4 in 1,000). On texts of 100 to 149 words it was 14 of 586, about 2.4 in every 100 (Wilson interval 1.4–4.0 in 100) — above the target of 2 in 100. The hardest human type, software issue reports from an issue tracker, received a wrong AI signal on 11 of 229 texts, about 4.8 in every 100 (Wilson interval 2.7–8.4 in 100) — above the target of 3 in 100. 245 of 2,929 human texts (8.4 in 100) fell in the Uncertain band.

The decision. Under the frozen test protocol the result was a fail on those two bars, and the protocol’s own record says so. The publisher decided to release the engine anyway, with the two shortfalls published here, on every result card for short texts, and in the technical documentation. This page never describes the release as having passed its final confirmation.

Per-type figures: every human type in the sealed test

2,929 human-written texts from 10 sources the system had never seen, all dated 2021 or earlier. Counts are texts that received a wrong AI signal (the Likely AI-written band).

By text length. 100–149 words: 14 of 586 (about 2.4 in 100). 150–299 words: 3 of 891. 300–599 words: 1 of 536. 600 words or more: 0 of 916.

Two human types excluded from the published rate (reported, not binding). Volunteer replies written in the style of an AI assistant (the OpenAssistant corpus, 482 texts on the development surface): about 12.9 in every 100 received a wrong AI signal and about 31 in every 100 fell in the Uncertain band. Contractor-written step-by-step maths solutions (48 texts): about 2.1 in every 100. These are excluded because the assistant register itself is what the system measures; if you check text written to sound like an assistant’s answer, expect a high score that says nothing about who typed it.

AI texts: the three-way split

Of 100 AI texts from the tested models: 55 strong signal, 26 uncertain, 19 no signal.

Current models tested: 400 texts, 200 from gpt-5.6-terra and 200 from gpt-5.6-sol, generated under the same recipe as the human prompts. Result: 221 strong signal (about 55 in 100), 103 uncertain (about 26 in 100), 76 no signal (about 19 in 100). By model: gpt-5.6-terra 126 of 200 strong; gpt-5.6-sol 95 of 200 strong.

One model lineage the system had never seen: 602 texts from Muse-Glimmer-30B. Result: 251 strong signal (about 42 in 100), 164 uncertain (about 27 in 100), 187 no signal (about 31 in 100). This is one lineage only and must not be generalised to all unseen models.

In our sealed test of texts from the tested current AI models, about 55 in every 100 received a strong AI-writing signal. A text with no signal is not evidence that a person wrote it.

Model identity

Method summary

Thresholds, mapping and bars were frozen and hash-stamped before the sealed surfaces were opened; each surface was scored exactly once; nothing was retuned afterwards. The human surface came from ten sources never used in development (health reference answers, book reviews, a professional Q&A site, parliamentary debate, presidential speeches, an open-source issue tracker, news, broadcast transcripts and the Congressional Record). Full protocol: methodology and training and validation.

History

V2 (4–5 September 2026), dated history. 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 (about 1.5 in every 1,000); 77 of 775 AI texts from model families seen in training (about 10 in every 100) and 293 of 600 texts from one unseen lineage, Inkling-Small (about 49 in every 100), were flagged. These figures describe V2 only.

Previous-generation detector (August 2026), dated history. 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; the complete package, including sample-level outcomes, the manifest, charts and checksums, remains downloadable: Download the historical benchmark package (ZIP). Reference: Dugan et al. (2024), RAID.

Model card: V2 (4 September 2026), dated history

The card of the engine V3 replaced. It is kept here so that a result seen on 4 or 5 September can be matched to the engine that produced it; nothing in it describes the current product.

Version history

• 27 August 2026 — previous-generation detector, 880-text RAID evaluation published.
• 4 September 2026 — V2 replaced the previous engine; V2 sealed confirmation published.
• 5 September 2026 — V3 released under an owner release override (see above); this page replaced the benchmark page.

Every release since is listed in the model update history, with its release id. Future versions must remain separately accessible or be clearly archived. A result should never be silently replaced after a detector change.

Frequently asked questions

How accurate is AI Detector Checker’s AI detector?

It is not summarised by one accuracy figure. In a sealed test on human-written texts of 150 words or more, from sources the system had never seen, it gave a wrong AI signal for about 2 in every 1,000 texts. For texts of 100 to 149 words it was about 2.4 in every 100. In our sealed test of texts from the tested current AI models, about 55 in every 100 received a strong AI-writing signal. A text with no signal is not evidence that a person wrote it.

Which texts are hardest?

Short texts, technical writing such as software bug reports, and volunteer replies written in the style of an AI assistant are harder for the system; the figures for each type are on the Evidence page.

Is the score the chance that AI wrote the text?

No. It is an AI-writing signal score, a frozen mapping of the detector’s internal value, shown with a score stability range. The band is decided by the raw thresholds only.

Does this test cover ChatGPT, Claude or Gemini?

The sealed test covered two current OpenAI models and one unseen lineage. It establishes no rate for any other model, and the detector never identifies which model wrote a text.

Can this result prove who wrote a text?

No. Results should be combined with context, provenance, revision history and human review.

Try the AI detector · Read the methodology

Model card (detail)

The full model card of the deployed engine, merged here from the former model-card page.

A technical summary of the detection engine currently running in production. This card describes the deployed system only — not experimental work, and not earlier architectures.

Current engine: V3 (released 5 September 2026)

V3 is the frozen V2 fusion core, new frozen length-banded decision thresholds, a frozen monotonic 0-100 display mapping, and a new product contract and result card. No classifier was newly trained. The sealed test of this engine and the publisher’s release decision are at the top of this page; the thresholds are published in full, in internal and displayed units, on the methodology page.

Architecture

A compact CPU model. Three classical text classifiers — a word n-gram model, a character n-gram model and a stylometric-feature model — each produce a calibrated score, and a small logistic model fuses them into one internal document-level decision score. No neural encoder is used in the current product: a programme of high-capacity encoder experiments was closed after each candidate failed its pre-registered safety gates on unseen human sources.

One scoring path. Every submission of at least 100 words is scored by the same English-trained components; there is no separate model for other languages. Text in other languages is processed but its results are unsupported.

No post-hoc adjustments. The internal decision score is compared with the two cut points frozen for its length band before the final sealed confirmation and published on the methodology page; nothing else moves the result.

Inputs and limits

Result semantics

The internal decision score is rank-like, not the chance of AI authorship; the displayed score is a frozen monotonic mapping of it and the band is decided by the frozen raw thresholds alone. A higher value means a stronger match to the patterns the detector associates with machine generation, but the same value does not carry the same meaning across every kind of human writing — which is why the product shows the stability range and publishes the false-flag rate per text type. The two cut points depend on the length of the text and are listed, in internal and displayed units, on the methodology page.

Order: 100–149 / 150–299 / 300–599 / 600+ words.

Every result carries the disclaimer that it does not prove AI or human authorship. No auxiliary indicators or certainty values are shown, and no individual sentences are marked.

Intended use

AI Detector Checker is intended only for low-stakes review of your own text, or text submitted with the writer’s informed consent. The output is a conservative AI-writing signal, not proof of authorship, misconduct, intent, or use of any specific AI system.

• Self-review of your own drafts to find passages that may need clearer evidence, specificity, or editorial attention.
• Low-stakes editorial review of consented text by writers, editors, publishers, and content teams.
• Detector research, quality analysis, and testing that do not evaluate, rank, accuse, penalize, or determine eligibility for any person.

Out of scope — do not use this for

• Any academic, admissions, employment, disciplinary, legal, healthcare, credit, insurance, housing, immigration, eligibility, or compliance decision about a person.
• Accusing, grading, admitting, rejecting, hiring, firing, diagnosing, ranking, penalizing, or determining eligibility for any person.
• Establishing who wrote a text or identifying which AI tool produced it. The engine performs no authorship or model attribution.
• Scoring text that you do not own and that the writer did not give informed consent to have analysed.

Measured limitations

• V2 sealed confirmation (4 September 2026, dated history): 4 of 2,604 human-written texts of 100 words or more received the strong label (about 1.5 in every 1,000); no source with 200 or more texts had a strong flag. The current V3 result is on the Evidence page: In a sealed test on human-written texts of 150 words or more, from sources the system had never seen, it gave a wrong AI signal for about 2 in every 1,000 texts. For texts of 100 to 149 words it was about 2.4 in every 100. Volunteer replies written in the style of an AI assistant are excluded from the published false-flag rate and reported separately: on the development surface about 13 in every 100 such texts received a wrong AI signal.
• Input below 100 words is refused rather than scored. Paraphrased, translated or human-edited machine text becomes harder to identify and is not separately validated.
• Recall is deliberately low and varies by model family. V2 sealed confirmation (dated history): 77 of 775 texts from model families seen in training were flagged (about 10 in every 100); 293 of 600 texts from one unseen lineage (Thinking Machines Inkling-Small) were flagged (about 49 in every 100). The current V3 result: In our sealed test of texts from the tested current AI models, about 55 in every 100 received a strong AI-writing signal. A text with no signal is not evidence that a person wrote it. The second figure describes one family and is not a general unseen-model rate; some families are almost never flagged.
• Non-English text, source code, contemporary fiction and heavily edited mixed text are not validated; the current validation is strongest for English.
• No single accuracy figure is published, because a number without its dataset, threshold and generation details cannot be independently checked. See benchmarks.

Privacy

Submitted content is sent to the service and analysis backend for scoring. For current analyses, the reviewed WordPress application does not create a per-analysis database record; historical non-content records remain, and infrastructure may create technical logs. Do not submit sensitive content. See the privacy policy for the current data flow and retention details.

Integrity

Every release records a hash of each artefact the scoring path loads and verifies them at start-up, refusing to serve if anything differs. A fixed reference set is replayed against the live service after every deployment. See training and validation.

Limitations and false positives (detail)

Merged here from the former limitations page.

Understanding the limits of AI detection matters because the result of a scan is never the full story. A detector can help you review text more carefully, but it cannot reconstruct the writing process, prove intent, or be used to make, support, influence, or trigger a high-impact decision about any person. This page explains where uncertainty comes from, why mistakes can happen, and how to use AI Detector Checker more responsibly when a result looks clear, ambiguous, or unexpectedly wrong.

Honest tools acknowledge uncertainty. AI Detector Checker is designed to provide pattern-based signals, not absolute claims. That makes the output more useful in real review workflows, because it encourages closer reading, context, and human judgment instead of over-reliance.

Try the free AI detector when you need a first-pass review, then use the guidance below to understand what a flagged or mixed result can and cannot tell you.

What This Page Covers

This page explains the practical limitations of AI detectors, including false positives, false negatives, mixed authorship, and the kinds of text that are harder to classify reliably. It is meant for users who want a clearer view of what a detector can do well, where it can struggle, and why one result should be interpreted in context rather than treated as final proof.

If you want to understand how to read the three result bands more directly, see how to interpret AI Detector Checker results. This page focuses on the part that comes after that: uncertainty, edge cases, and responsible use.

Why AI Detection Has Limits

AI detection is pattern-based. It does not read minds, observe authorship history, or watch how a document was created. It looks at characteristics of the text itself and infers how likely those patterns are to resemble AI-generated writing. That can be useful, but it also means the result is an estimate, not direct proof.

That limitation is not unique to one tool or one result. It is part of the problem AI detection is trying to solve. Human writing and model-generated writing can overlap in style, especially when a text is formal, standardized, short, heavily revised, or written for a narrow purpose. The detector sees the finished output, not the full drafting process behind it.

If you want more background on the scan process itself, review how AI Detector Checker analyzes text. The important point here is that any detector is making a structured inference from observable writing patterns, not delivering a direct record of how the text was produced.

False Positives and False Negatives: What They Mean

A false positive happens when human-written text is flagged as AI-like. A false negative happens when AI-assisted or AI-generated text is not clearly identified by the detector. Both matter because both can distort how a result is interpreted.

False positives matter because they can lead people to overreact to writing that is genuinely human but happens to look more uniform, formal, or standardized. False negatives matter because they can create too much trust in text that was shaped more heavily by AI than the result suggests.

A useful detector tries to reduce both types of error, but no detector can eliminate either one completely. That is why AI Detector Checker presents signal-based results rather than pretending every scan can be reduced to a perfect yes-or-no answer.

Why Human Writing Can Be Flagged

Human writing can be flagged when it shares patterns that detectors often associate with AI output. That does not mean the writer did anything wrong. It means the final text looks more machine-like in certain ways.

• Highly formal academic prose: polished, impersonal, and highly structured writing can sometimes look statistically predictable.
• Repetitive or formulaic structure: when paragraphs follow similar patterns, the text may appear more uniform than natural drafting usually does.
• Generic low-specificity business copy: language that sounds polished but vague can trigger stronger AI-like signals.
• Template-heavy writing: standard document structures, policy language, and routine summaries often reduce stylistic variation.
• Polished but predictable transitions: overly smooth connections between ideas can make the prose feel more synthetic.
• Non-native English patterns: predictable constructions may reflect language proficiency, not machine authorship.
• Translated text: translation can standardize phrasing and flatten local voice.
• Technical or domain-specific summaries: specialized writing can be concise and formulaic even when entirely human-written.

These cases are one reason a flagged result should always lead to closer review, not immediate certainty.

Why AI-Generated Text Can Slip Through

AI-generated text can sometimes avoid a strong detection signal for reasons that have little to do with the detector being careless. The text may have been heavily revised by a human, blended with original writing, or shortened to the point where the detector has fewer signals to work with.

In other cases, the output may come from workflows that do not look like obvious raw generation. A writer might start with AI, rewrite major sections manually, add original examples, or restructure the draft heavily. Language models also evolve, and not every prompt produces the same style. Some passages are simply more ambiguous than others.

That is why “Likely human-written” does not guarantee human authorship. It means the text showed fewer clear AI-like patterns according to the signals available in that scan; at its conservative setting the detector misses much AI-written text.

Mixed Authorship and Edited Drafts

Many modern drafts are neither purely human nor purely AI. A writer may use AI to brainstorm, generate an outline, rewrite a paragraph, or smooth transitions, then continue editing in their own voice. Another writer may begin with a human draft and use AI only for restructuring or phrasing support. These blended workflows are increasingly common.

Mixed authorship is one of the main reasons some results fall into a midrange or produce uneven signals across the document. One section may feel natural and highly specific, while another may sound more standardized or over-smoothed. That does not always support a clean binary conclusion.

Users should avoid forcing a simple answer when the evidence is mixed. In many real workflows, the useful question is not “human or AI only?” but “which parts of this document deserve closer review, clarification, or revision?” The AI Detector Checker use cases show where that kind of review matters most.

What Affects Result Reliability

Some texts are easier to classify than others. Reliability depends on the quality and quantity of detectable signals in the draft, as well as the kind of writing being reviewed.

• Short text: brief passages provide less evidence and can be harder to classify reliably.
• Highly formal or template-like writing: standardized phrasing reduces stylistic variation.
• Non-English or mixed-language text: the current validation is strongest for English; other languages are unsupported.
• Translation effects: translated passages may read more standardized than original composition.
• Non-native writing: predictable phrasing can reflect language background rather than machine generation.
• Hybrid human + AI drafting: blended workflows can create mixed signals that do not settle cleanly.
• Heavily edited AI output: strong revision can weaken the original AI-like patterns.
• Technical or domain-specific writing: precise, formula-driven language can resemble high-structure model output.
• Uneven document sections: an introduction, body, and conclusion may not all carry the same signal strength.
• Very short openings or closings: short sections can look disproportionately simple or formulaic.

The current validation of AI Detector Checker is strongest for English text; other languages have not been validated for the current detector, so results on them are unsupported. For more detail, see how language and translation affect AI detection.

Common False Positive Scenarios

A student essay with highly formal language

A student may write in a cautious, structured, overly polished style because they are trying to sound academic. The result can look more machine-like even when the work is genuinely their own.

A translated marketing page

Marketing copy adapted from one language into another may lose natural local rhythm and begin to sound standardized. That can create stronger AI-like signals even without direct generation.

A technical abstract or policy summary

Highly condensed writing in technical or institutional settings often follows narrow conventions. That can reduce variation and make human-authored summaries look more synthetic.

A human draft with repetitive corporate phrasing

Internal business documents sometimes rely on repeated stock language, cautious wording, and uniform structure. The result may be fully human-written but still appear statistically predictable.

What AI Detector Checker Does to Reduce Error

AI Detector Checker is designed to make interpretation more responsible, not more absolute. Instead of presenting a yes-or-no verdict, it reports a 0-100 AI-writing signal score with a stability range from a single document-level classifier and one of three bands: Likely human-written, Uncertain, or Likely AI-written. In a sealed test on human-written texts of 150 words or more, from sources the system had never seen, it gave a wrong AI signal for about 2 in every 1,000 texts. For texts of 100 to 149 words it was about 2.4 in every 100. Every result states that rate for its length and carries the reminder that it does not prove AI or human authorship. Short texts, technical writing such as software bug reports, and volunteer replies written in the style of an AI assistant are harder for the system; the figures for each type are on the Evidence page. For a closer explanation of what each band shows, see understanding the three result bands.

The detector combines several measured writing signals into one label rather than relying on any one pattern alone. That matters because writing can look AI-like for different reasons, and one pattern by itself is rarely enough to support a trustworthy conclusion. The features overview explains why the product deliberately shows a label instead of a headline number, and why an Uncertain result is a legitimate outcome.

For users who want more methodological context, the benchmarks and performance methodology provide additional transparency. These are safeguards that improve review quality. They are not proof of perfection.

What to Do If You Suspect a False Positive

If a result seems wrong, the most useful next step is a careful review of the text and its context. A flagged result may reflect style, genre, translation, or structure rather than actual AI generation.

• Re-read the text in context: check whether the writing is naturally formal, repetitive, or standardized for the document type.
• Compare only permitted drafts: in a low-stakes self-review or informed-consent editorial workflow, compare the current text with the writer’s own earlier drafts and source notes.
• Check whether the draft is translated, template-driven, or technical: those factors can affect how the detector reads the text.
• Gather more context before drawing conclusions: understand how the draft was prepared and what kind of writing it was meant to be.
• Revise for clarity and specificity where appropriate: strengthening voice, evidence, and context can make the document easier to interpret.
• Reanalyze after legitimate revision: use a second scan as a review aid, not as a search for a “perfect” number.
• Avoid accusations based on one scan alone: a single result should never be treated as conclusive proof.

If the material is unpublished, internal, or sensitive, review AI Detector Checker privacy and in-session text handling before making scanning part of a regular workflow.

When Not to Treat One Scan as Final Proof

AI Detector Checker must not be used in academic, admissions, employment, disciplinary, legal, healthcare, credit, insurance, housing, immigration, eligibility, compliance, or other high-impact decisions about a person.

A single automated result cannot capture how a draft was produced, how much revision took place, or whether formal or translated writing shaped the signal. A strong label is not proof of misconduct, and a no-signal label does not guarantee human authorship. Human review does not make the output suitable for decisions about a person; prohibited uses remain prohibited even alongside other evidence.

AI Detection, Plagiarism Checking, and Human Review

AI detection, plagiarism checking, and human review solve different problems. Plagiarism checking compares a draft against known or indexed sources to identify overlap. AI detection looks for authorship-style patterns that suggest the text may be machine-generated or strongly machine-shaped. Human review adds the context that neither automated method can fully provide.

That is why one does not replace the others. A document can be original but still feel heavily AI-generated. It can also be human-written and still borrow language too closely from outside sources. For a deeper comparison, see AI detection vs. plagiarism checking.

How to Use AI Detector Checker Responsibly

A responsible workflow is limited to your own text or text submitted with the writer’s informed consent. Read the label as a signal, then rely on your own reading of the text for revision, fact-checking, sourcing, or editorial quality review.

From there, choose a low-stakes next step such as revision, source-checking, factual verification, or further editing. Do not use the scan to approve, reject, escalate, accuse, or otherwise evaluate a person. For quick operational questions, the AI Detector Checker FAQ is a useful companion. For broader trust and product context, visit about AI Detector Checker.

FAQ

Can AI detectors be wrong?

Yes. AI detectors can produce both false positives and false negatives because they rely on pattern recognition rather than direct proof of how the text was written.

What is a false positive in AI detection?

A false positive happens when human-written text is flagged as AI-like by the detector.

What is a false negative in AI detection?

A false negative happens when AI-assisted or AI-generated text is not clearly identified by the detector.

Why can human writing look AI-generated?

Highly formal, repetitive, translated, technical, or template-driven writing can sometimes appear more uniform and predictable, which may increase AI-like signals.

Does a strong AI-writing signal prove misconduct?

No. A “Likely AI-written” label indicates a stronger match to AI-like patterns, but it does not prove misconduct, intent, or complete authorship history on its own.

What should I do when a result seems wrong?

Review the text in context, gather more information about the draft, compare with expected writing style when appropriate, and avoid treating one scan as the final word.

Use AI Detection With More Context

The most reliable way to use an AI detector is with clear expectations and responsible interpretation. When you understand where limits come from and how false positives can happen, the result becomes more useful, not less.

Run a scan in AI Detector Checker and use the output as a stronger starting point for informed human review.

The methodology section explains the safeguards that exist specifically to reduce these false positives, and the model card lists the measured limitations of the deployed engine.