Lecture 11 — The ethics of accusation

What forensic work owes the person on the other end of a wrong verdict

3 May 2027

Every method in this course can be pointed at an innocent person. A distinctive human writing style can trip a text detector. A real photograph, badly compressed, can trip an image one. A student’s unmedicated tremor can read, to an untrained reviewer, like a synthetic voice’s artefacts. A forensic practice that only worries about missing the guilty and never about convicting the innocent is not a forensic practice — it’s a witch-hunt with better instruments.

This lecture borrows the presumption of innocence and the burden of proof directly from actual criminal procedure, not as a metaphor but as a working standard: what would it take for you to be confident in an accusation, not just suspicious, and what do you owe someone when you’re wrong. It sets up the week’s casework, which asks you to argue both sides of a real provenance dispute before you’re allowed to reach a verdict.

This isn’t hypothetical. A Johns Hopkins professor had to hand-review a student’s drafts after a detector flagged over 90% of her own paper as machine-written. Two University of California, Davis students were flagged and separately cleared on appeal. In early 2026, a court ordered an Adelphi University student’s “100% AI” finding expunged, calling it without valid basis. More than two dozen universities, including MIT, Yale and UC Berkeley, have since restricted or dropped AI-detection tools outright — not because detection doesn’t matter, but because a tool wrong 5 to 20% of the time was being treated as a verdict rather than a lead.

Outline#

• false positives have a face: what a wrongful “this is AI” accusation costs a real person
• burden of proof, applied to content forensics
• why you argue the side you don’t believe before you’re allowed to conclude anything

Further reading#

The Washington Post — proving your innocence after a false positive covers exactly what a wrongly accused student actually has to do next, which is a more useful thing to read this week than another explanation of why detectors fail.