Transparent Policies, Integrity, and the Equity/Ethics Audit

Why detection cannot carry an integrity process: a detector score is a probabilistic signal with unstable error rates, documented false-positive bias against non-native English writers, and a chilling effect on the help-seeking that integrity actually depends on. The lecture replaces it with an evidence stack you design — process evidence in the assignment, a low-stakes conversation, comparison to demonstrated capability, and required documentation — ordered restoratively: notice, converse, then respond according to what the evidence shows. Policy language shifts from deterrence by threat to definition and invitation.

The center of a defensible policy is core intellectual substance: the thinking, judgment, and final claims students must own in your course. The lecture shows why blanket bans, anything-goes permissiveness, and vague 'use AI responsibly' language all fail students, then builds a 150-to-250-word policy from seven components, including documentation expectations, a privacy warning, and an explicit help-seeking invitation. A read-aloud test checks the tone, a decision tree gives students a safe next action at every point (including after a mistake), and the policy must agree with each assignment's stated AIAS level.

An audit across eight dimensions — access, language and disability, privacy, truth and verification, copyright, labor and environment, bias, and detection — with one standard throughout: ethics is a design constraint, not a disclaimer. Each dimension comes with operational checks, like requiring a no-cost equivalent path, permitting translation when language production is not the outcome, treating AI output as never-a-source, and declining to authorize detection. The audit's discipline is that it must change one concrete sentence in your policy or assignment before the policy is finalized.