---
name: "ai_writing_detection"
description: "Interpret, audit, and respond to AI-writing detection: how detectors (Turnitin, GPTZero, Originality.ai) score text, what their signals and accuracy limits really are, and how to rewrite flagged prose or handle a false positive fairly. Use when someone's writing got flagged, when you need to assess detector output, or when editing AI-assisted drafts to read like a person's work."
---

# AI Writing Detection

## Purpose
Turn a detector score into a grounded read. Commercial AI detectors do not verify authorship; they score statistical patterns in text. This skill covers what those patterns are, how the major tools differ, where their accuracy claims break down, how to revise prose that reads machine-like, and where the ethics lines sit. It does not help anyone submit machine-written work as their own where that is against the rules.

## Workflow

1. **Name the question.** Detection work falls into three cases. Route first:
   - *Was I flagged unfairly?* Go to the false-positive defense playbook.
   - *Is this text AI-written?* Read the signals and their limits, then give a judgment with confidence bounds.
   - *How do I make this read like a person wrote it?* Apply the rewrite patterns to the writer's own draft, not to disguise authorship.

2. **Read the signals.** Every major detector (Turnitin, GPTZero, Originality.ai) measures a handful of statistical properties. See `references/detector-signals.md` for the mechanics and tool-by-tool notes.

3. **Apply the accuracy facts.** A score is a probability, not proof. Key bounds: vendor claims (~98%) hold only for raw, unedited AI output; independent tests put real-world detection at 60-85% on edited text, and false positives hit non-native English writers and formal academic prose hardest. See `references/accuracy-and-limits.md` for numbers, sources, and which universities disabled the tools and why.

4. **Revise or respond.**
   - For rewriting: run `bin/text_signals.py` on the draft to see which signals push it into the machine-like zone, apply the rewrite patterns in `references/rewrite-patterns.md`, and re-run to confirm the shift. Keep the writer's meaning and voice; do not strip them to dodge a detector.
   - For a false accusation: use the defense playbook in `references/false-positive-defense.md` (process evidence first, then the score's known failure modes).

## Output Contract
- Say what kind of answer you are giving: a signal reading, a revision, or a defense plan.
- When judging authorship, always state a confidence band and the reasons for it; never present a detector score as conclusive.
- Cite the specific signal (perplexity, burstiness, sentence-level highlight) behind each observation, not the score alone.
- For rewrites, show before/after on at least one sentence and explain which signal the change moves.

## Operating Rules
1. Never invent accuracy numbers, study results, or detector behaviors. Every factual claim about a tool traces to a named source; if you cannot name one, say you do not know.
2. Never help someone pass off machine-written work as their own in a setting that forbids it. Revising the writer's own draft is fine; anonymizing authorship to beat an integrity check is not.
3. Do not present any detector result as proof of AI authorship. Scores fail on edited text, formal prose, and non-native English writing; treat every result as one signal among several.
4. Do not name or discuss the person who commissioned this skill or their preferences anywhere in skill output. Talk about the work, not who ordered it.
5. Keep SKILL.md lean. Deep detector mechanics, the accuracy evidence, and the rewrite playbook live in `references/`; do not duplicate them here.