AI detectors alone are the wrong tool. They're noisy, biased against ESL students, and they reduce a complex teaching question to a binary. Here's how to think about it instead.
Layer 1: design assignments AI can't trivially do
- Anchor writing to in-class events, conversations, and observations
- Require process artifacts: outline, draft, revision, reflection
- Use timed in-class writing for high-stakes assessment
Layer 2: make the writing process visible
TeacherAI's Research Paper Writer logs every AI interaction, every paste, and every major revision. When you can see process, you don't need to guess at the final product.
Layer 3: detection as a conversation-starter
Detection scores should never be the basis for an academic-integrity charge alone. Use them to start a conversation ('walk me through how you wrote this'), not to end one.
What good detection looks like
- Sentence-level highlights, not a single page-wide percent
- Confidence ranges, not fake-precise scores
- Calibrated against multilingual student writing samples