Experimental AI text signal

Inspect style.
Never infer authorship.

Paste at least 60 words to receive a writing-style warning signal. It is intended for triage and cannot establish who wrote the text.

Open the free checker

Scope before score.

Every result is a warning signal. The evidence and limitations below define what this route can and cannot support.

Minimum sample

Samples below 60 words are blocked in the website and return unknown through the API.

Live policy

The deployed text-style-signals-v3-warning-only lane exposes score, model, source, reason and word count.

OOD evidence

On 1,562 MAGE GPT-4 OOD texts, the tested production hybrid reached 79.00% accuracy with 13.12% real false positives.

Do not infer more than the lane measured.

  • Short, Dutch, edited and mixed-language content can be unreliable.
  • Style overlap can create false positives for formulaic human writing.
  • Never accuse, penalize or reject a person based on this score.

Continue with evidence

False positivesRead more →MethodologyRead more →Open the API docsRead more →