This section explains how Detector Checker actually works, how it is evaluated, and where it fails. It is written from the deployed system rather than from marketing copy, and it is updated when the production detector changes.
What the tool produces
Detector Checker returns an AI-likeness score for a submitted text, a descriptive band, and — where the document scores highly enough — sentence-level highlights that show which passages carry the strongest signal.
The score is not a probability that a particular person or tool wrote the text. It is a measure of how closely the writing matches the patterns the detector has learned to associate with machine generation. No detector can inspect a private writing history or recover the tool that produced a passage, so the result is evidence for review, never proof of authorship.
How the score is produced
The score comes from a single document-level classifier: a calibrated ensemble that combines several independent, publicly available transformer text classifiers with word-predictability measurements and a set of statistical, structural and semantic features. Those components are weighted together by a calibrated model into one number.
English and non-English text follow different paths. When the text is unlikely to be English, the document is routed to a dedicated multilingual model rather than being scored by English-only components.
The three visible indicators are not the score
The tool also shows three auxiliary indicators — predictability, semantic and statistical. These exist to give a reader some texture about why a passage reads the way it does. They are not averaged, summed or voted into the final score, and a result can legitimately show a high score beside low auxiliary indicators, or the reverse.
The value labelled auxiliary signal agreement measures only how closely those three indicators agree with each other. It is not confidence, not accuracy, and not the probability that the result is correct.
Safeguards applied after scoring
A raw classifier score is not served directly. Several guarded adjustments sit between the model and the number you see, and each one is deliberately one-directional so its effect is predictable:
- A false-positive correction that can only lower a score. Formal, technical and literary human writing is the register most often misread as machine-written, and this correction exists to protect it.
- A short-document exception, because the correction above was found to misfire on short texts that the classifier was already confident about.
- An evasion adjustment that can only raise a score, for English text that shows the signature of deliberately humanised machine output.
- A recovery step that can only raise a score, for the narrow case where the false-positive correction demonstrably suppressed a document the underlying classifier was maximally certain about.
Input is also normalised before scoring so that invisible-character and look-alike-character tricks cannot be used to suppress a score.
Reading the result
| Band | Displayed score | What it means |
|---|---|---|
| Lower signal | below 40 | The text does not strongly match the patterns the tool associates with machine generation. |
| Indeterminate | 40 to 49 | The evidence is mixed. Treat this as “no useful answer”, not as a weak accusation. |
| Higher signal | 50 and above | The text matches those patterns more strongly. Sentence-level highlights are shown. |
The displayed percentage is deliberately capped at 1 and 99. The tool will never display 0% or 100%, because neither certainty is available to it.
Known limitations
- Short text is weak evidence. A sentence, heading, quotation or list fragment carries very little signal. The tool requires at least 50 characters and is far more useful on a complete passage.
- Formal and formulaic human writing is the main false-positive risk. Academic prose, technical documentation, legal and administrative text, and writing by non-native speakers are all more likely to be flagged.
- Editing and paraphrasing degrade detection. Machine text that a person has rewritten, translated or heavily edited becomes progressively harder to identify, and can become genuinely undetectable.
- Coverage is uneven across sources. Detection is stronger for some text generators than others, and a newly released model may not be well covered until the detector is evaluated against it.
- Non-English support is real but less validated. Multilingual input is accepted and scored by a dedicated model, but per-language performance has not been publicly validated.
How the detector was built and tested
The engine behind these numbers was developed against an evaluation corpus of 305,234 documents — 205,748 human-written and 99,486 machine-generated — spanning 51 text generators, 61 writing registers, 14 languages and 12 families of deliberate evasion technique, with duplicates and previously-seen material removed before anything was measured.
Changes are held to a written, hash-stamped standard agreed before the result is produced, and most proposals fail it: across the eight most recent evaluation rounds, one produced a change that reached production.
Read the full training and validation methodology →
Read next
- Training and validation methodology — how the detector is built and tested, and the split discipline used to keep results honest.
- Model card — the technical summary of the engine currently deployed.
- Model update history — what has changed in production, and when.
- Benchmarks and evaluation — why this site does not publish a single headline accuracy number.
- Limitations and false positives — practical guidance for acting on a result.