AI detection with evidence for educators
The problem
Teachers face AI-assisted assignments daily. An opaque ’85% AI’ score gives no evidence to discuss with a student, and a false accusation costs a student far more than a missed draft.
Most detectors also go quiet on short answers, which is exactly what homework mostly is: two paragraphs, not two thousand words.
How Stipple helps
Stipple shows the specific phrases and stylistic markers behind the score, states a confidence band, and abstains when there is too little prose to judge — an honest ’not enough text’ on a two-sentence answer instead of a confident number on nothing.
Coverage indicators say how much of the submission the analysis actually read. For a teacher that means talking points you can walk a student through, rather than a verdict you cannot explain.
The call
View Full Recipe →curl -X POST https://www.stipple.sh/v1/detect-ai-text \
-H "Content-Type: application/json" \
-d '{"text": "Student essay draft..."}'Foster academic integrity with verifiable evidence and student-friendly dialogue.
What the result looks like
A probability with a confidence band, the specific tells that fired highlighted in the text, a coverage indicator, and honest abstention when the sample is too small.
A clean result means the analysis found no AI-style signals in the prose it could read. It is not a certificate of human authorship and never claims to be — that asymmetry is deliberate, because a false accusation is the expensive error here.
Since August 2026 some AI text also carries a provider watermark. Detecting that requires the provider’s own detection API, which is announced and not yet public, so style analysis remains the working tool for student work today.
How the corpus was built, what was caught, and what the method does not prove.
Questions
- How do I check an assignment for AI writing?
- Paste the submission into the browser tool, or call the API. You get a probability, the specific signals behind it, and how much of the text was readable. Anything consequential — an academic-integrity case — should combine that with process evidence: drafts, version history, and a conversation. Process evidence outranks any detector score.
- Does it work on short answers and single paragraphs?
- Style detection needs prose to read, so very short answers get an honest abstention rather than a number. That is deliberate: a confident score on two sentences would be the least reliable output we could give you. The result states what could and could not be judged.
- Can a student be accused on the basis of a score?
- No, and the result is built to say so. Treat it as the start of a conversation with the evidence shown — the highlighted tells a reviewer can walk through. A false accusation costs more than a missed draft, and the tool is designed around that asymmetry.
- What about text a student rewrote from an AI draft?
- Heavy rewriting lowers any style-based score. That is a documented effect of paraphrasing rather than a fault in the tool, and a rewrite also strips provider watermarks, so an unmarked rewrite proves nothing either way. Process evidence remains the strongest material in academic cases.
- Does this detect provider watermarks?
- Not yet — nobody outside the provider can, because the detection key is not public. When those APIs ship, watermark verification will slot in as a first check in front of style analysis. Until then, style analysis with shown evidence is the honest tool available.