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.
| Field | Value |
|---|---|
| Engine version | engine-2026-08-26 |
| Release | rel-2026-08-26-00ad353f8f70 |
| Deployed | 26 August 2026 |
| Task | Binary document-level classification: machine-generated vs. human-written text |
| Output | AI-likeness score 0–100, a descriptive band, and sentence-level highlights |
| Languages | English (primary); other languages accepted and routed to a dedicated multilingual model |
Architecture
A calibrated ensemble. Several independent, publicly available transformer text classifiers are combined with word-predictability measurements and a set of statistical, structural and semantic features. A calibrated model weighs those inputs into a single document-level score.
Two scoring paths exist. Text assessed as unlikely to be English is routed to a separate multilingual model rather than being scored by English-only components. The English path is not affected by that routing.
A small number of guarded, one-directional adjustments are applied after scoring — described in the methodology. Each can move a score in one direction only, which is what makes their effect auditable.
Inputs and limits
| Property | Value |
|---|---|
| Minimum input | 50 characters |
| Maximum input | 50,000 characters |
| Input type | Plain text, pasted or extracted from an uploaded document |
| Displayed score range | 1–99 (never 0 or 100) |
Score semantics
The score is an AI-likeness measure, not a calibrated probability of authorship. A score of 80 does not mean “80% likely written by AI”. It means the text matched the patterns the detector associates with machine generation more strongly than a text scoring 60 did.
| Band | Score | Sentence highlights |
|---|---|---|
| Lower signal | below 40 | Not shown |
| Indeterminate | 40–49 | Not shown |
| Higher signal | 50 and above | Shown |
Auxiliary signal agreement, shown beside the score, measures only how closely the three visible indicators agree with one another. It is not confidence in the result.
Intended use
- A first-pass review signal that helps a person decide which text deserves a closer look.
- Prompting a conversation about how a piece of writing was produced.
- Screening at a stage where a wrong answer is cheap to correct.
Out of scope — do not use this for
- Any academic, employment, disciplinary, legal, healthcare, immigration or compliance decision, whether alone or as the deciding factor.
- Establishing who wrote a text. The engine performs no authorship attribution.
- Identifying which AI tool produced a text. The engine performs no model attribution, and any product that claims otherwise should be treated sceptically.
- Scoring text a person did not consent to have analysed.
Measured limitations
- False positives concentrate in formal, technical, formulaic and non-native English writing.
- Detection weakens sharply on short input, and on paraphrased, translated or human-edited machine text.
- Coverage is uneven across text generators; performance on any specific model should not be assumed from performance on another.
- Non-English performance has not been publicly validated per language.
- No single accuracy figure is published, because a number without its dataset, threshold and generation details cannot be independently checked. See benchmarks.
Privacy
Submitted text is scored in memory and is not retained after the response is returned. It is not stored, not used for training, and not shared. See the privacy policy.
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.