Case 08 — Breaking a detector

Trying to make your own content evade a detector, and taking your success or failure seriously as evidence

12 April 2027

Playing adversary for a session is uncomfortable if you’ve spent seven weeks building up the other set of skills. It’s also the fastest way to understand exactly how much confidence any single detector actually deserves.

Before the session#

Nothing to prepare beyond having read the week’s lecture — you’ll be working with material provided in the session.

In the session#

Given a piece of AI-generated content and a detector, you have twenty minutes to make edits that push the detector’s confidence down without materially changing what a human reader would take away from the content. Log every change you tried, including the ones that didn’t work.

Afterwards#

The honest version of this exercise ends with most of the room succeeding, at least partially. Sit with what that means for any policy — including one you might draft in week 10 — that treats a detector’s verdict as sufficient grounds for a decision on its own.