Detect AI-generated text in documents and submissions

Prose Style Detection estimates whether a document's prose shows machine-generated writing patterns. It uses Stipple to return a confidence estimate together with specific linguistic evidence, reasoning, and limitations. The skill deliberately abstains when the document is not prose, avoiding a style judgment on forms, tables, scans, or spreadsheets. It is Apache-2.0 licensed.

The workflow emphasizes the distinction between style and authenticity. Generic human writing can resemble automated output, while generated writing can be plain and varied. The result therefore belongs in triage alongside other evidence and human judgment. It does not establish who wrote the document or whether the underlying document has been altered.

Use it for educators screening submissions, publishers or platforms triaging incoming content, HR reviewing drafted CVs, or researchers examining source provenance. In the HR workflow, the result is a flag for review rather than grounds for rejecting a candidate. The same evidence-aware approach applies to other consequential uses.

Templated, coached, translated, accessibility-assisted, and non-native-English writing can be disproportionately flagged. Do not use the estimate as proof of misconduct or as the sole basis for employment, academic, publishing, or disciplinary action. Preserve the original work and provide meaningful human review and an appeal path before a decision affects the person involved.

Submission sends document contents to a hosted third party. Obtain explicit approval, remove unnecessary personal or confidential material, and check current retention and deletion terms before uploading.

Inputs can be a URL, local PDF, DOCX, TXT, or Markdown file, or raw text. A document is submitted to the detection endpoint; raw text can be sent as a JSON object to the same endpoint. A literal endpoint example is:

https://www.stipple.sh/v1/detect-ai-text

Read applicability first. If applicable is false, report that the material is non-prose and stop rather than interpreting a guessed authorship score. For applicable text, probability is model confidence on a zero-to-one scale, not calibrated truth. The lean field distinguishes machine-generated, human, and uncertain orientations.

The tells array identifies phrases and patterns such as stock transitions, uniform sentence length, or low variation. Reasoning explains the assessment, and limitations states the noise profile. The suggested report presents the probability and lean, prose ratio, linguistic tells, reasoning, and limitations together so the number is accompanied by its practical limits.

Stipple provides a free anonymous tier that works without a key and a free-key option for authenticated use. For document genuineness or tampering, the workflow points to the separate verify-document skill. That forensic route answers an integrity question; this skill examines prose style and keeps its output within that narrower purpose.