The Arms Race

Week 8

99.3% accuracy. Also: 80–92% accuracy.

Same class of tool. Different measurement.

Two numbers, one detector

  • vendor’s own benchmark: 99.3% accuracy across 3,000 samples, 0.24% false-positive rate
  • independent reviewers, same class of tool: 80–92% in real-world use, falling to roughly 96.5% even for a strong performer once the text is paraphrased or lightly edited rather than generated clean

Neither number is “wrong”

They’re measuring different things: a clean benchmark versus content that’s been through exactly the kind of editing real writing goes through. The gap between them is itself the lesson.

Goodhart’s Law: when a measure becomes a target, it stops being a good measure.

Applied here

The moment a detector’s verdict starts costing someone a grade, a payment or a job, someone builds a product whose only purpose is to remove the signal that verdict depends on. That’s not a bug in any one detector. It’s what happens to any single signal once enough weight rests on it.

Further reading

gptzero.me/news/ai-accuracy-benchmarking/

Read a vendor’s own benchmark methodology — critically

The practice, not the tool

No detector deserves to be the whole case. Multi-signal review — text, image, provenance, context — is the only defensible practice, and even that isn’t a guarantee.