A detector cannot prove ChatGPT use. It can highlight passages that resemble generated prose and give you a place to start a conversation about process, drafts and sources.
Compare flagged work to earlier writing from the same student. Ask them to explain a paragraph in their own words. Keep the extract and the report if you escalate. Never write up "Too close to call" or "Humanized AI" as if they meant "AI-generated".
Clear course policy beats surprise enforcement. Say what is allowed before you mark.
