Lecture 08 — The arms race

Why detectors keep failing, and what it means that an entire industry now exists to defeat them

12 April 2027

One detector vendor’s own benchmark reports 99.3% accuracy across 3,000 samples, with a false-positive rate of 0.24%. Independent reviewers testing the same class of tool put real-world accuracy closer to 80–92%, falling further — to around 96.5% for that same vendor — the moment the text has been paraphrased or mixed with human editing rather than generated cleanly. Neither number is wrong, exactly; they’re measuring different things, and the gap between a vendor’s self-reported figure and an independent test is itself the first thing a forensic reviewer should learn to distrust. That is not a story about bad engineering. It is a story about Goodhart’s Law: the moment a detection signal becomes a target — an “AI content” flag that costs a creator a grade, a payment, a job — someone builds a product whose only job is to remove that signal.

This lecture is the theoretical spine of the course: every method we’ve covered is a proxy for the thing we actually care about (was this made honestly, and does it deserve trust), and every proxy becomes gameable exactly in proportion to how much weight gets put on it. That doesn’t make forensic work pointless. It means the discipline is knowing how much weight any one signal can bear, and never resting a serious accusation on a single one.

Outline#

• Goodhart’s Law, applied to content detection instead of economic indicators
• the evasion-tools economy, and why it is structurally guaranteed to exist
• multi-signal review as the only defensible practice, and why it’s still not a guarantee

Further reading#

GPTZero’s own accuracy benchmarking page is worth reading precisely because it’s a vendor describing its own test — notice what the methodology does and doesn’t control for, and compare it against the independent figures cited above.