Use case
An AI detector for teachers that shows its evidence
The problem
Every marker now reads work that might be AI-written, and most detection tools hand back a bare percentage with no reasoning. A number alone can’t be discussed with a student, defended to a parent, or weighed against what you know about the writer.
The stakes of a wrong call are asymmetric: a missed AI essay is a grading problem, but a false accusation follows a student around. Detectors are known to false-flag formulaic, coached, and non-native-English writing — the exact styles many honest students produce.
How Stipple helps
Stipple reads the prose and returns a probability, a confidence band, and up to six specific tells — the phrases and patterns that moved the score, highlighted in the pasted text. When there’s too little text to judge, or the submission isn’t prose, it abstains instead of guessing.
Detection accuracy is measured and published — 94.7% across text from 12 language models on a 38-sample benchmark (95% interval 83–99%) — alongside an honest account of false positives. You see the same evidence the score is standing on.
The call
curl -X POST https://www.stipple.sh/v1/detect-ai-text \
-H "Content-Type: application/json" \
-d '{"text": "<the submission text>"}'
# → { "probability": 0.91, "lean": "ai", "tells": [ "uniform sentence rhythm", ... ],
# "limitations": "Triage, not proof." }Use it the way you’d use a colleague’s second opinion: run the passage, read the tells, then talk to the student — and ask for drafts or version history before any consequence.
Questions
- What AI detector do teachers actually use?
- Most institutions rely on whatever is built into their LMS or plagiarism suite, and many teachers run an independent checker as a second opinion. Stipple is built for that second-opinion role: free with a Google sign-in, and it shows the tells behind the score so you can weigh the evidence rather than trust a bare percentage.
- Can I fail a student because the detector says AI?
- No — and this tool is deliberately built to discourage that. The score is a confidence signal from writing style, not proof of authorship, and detectors are documented to false-flag formulaic, coached, and non-native-English writing. Treat a high score as a reason to have a conversation and ask for drafts or version history, never as a verdict on its own.
- How accurate is it?
- Measured and published, not claimed: 94.7% detection across text from 12 language models on a 38-sample benchmark, with a 95% interval of 83–99%. On any single essay the honest unit is the probability and its confidence band — and when there is too little prose to judge, the tool abstains rather than guessing.
- Is student work stored?
- No. Checks are processed and not retained — the text is never stored. If you sign in, your history keeps only the verdict summary (score, lean, how many tells fired), never the submission itself.
- How long an essay can it check?
- Most essays are scored in a single read. Longer pieces are split on paragraph boundaries and scored section by section — up to four sections of roughly 2,300 words each per check, each with its own score and tells, plus the weighted overall. A piece longer than that reports exactly how much was read instead of pretending it saw everything.
- Can students use this to check their own essay before submitting?
- Yes, and it is a legitimate use: a student sees the same kind of evidence a marker sees — which passages carry AI-style signal and how strong it is — so nothing surprises them at accusation time. No tool can promise an essay will pass a different detector; what this gives a student is foreknowledge, and a reason to keep drafts and version history, which are far stronger evidence of authorship than any score.
- A student’s own writing keeps getting flagged. What should they do?
- Treat it as the documented failure mode it is — formulaic structure, coached academic phrasing, and non-native-English styles all raise false-positive risk. The durable protection is process, not style surgery: keep drafts and version history, and never let a detector score stand as the whole case in either direction.