Authors Guild test finds some AI detectors perfectly identify human writing while others fail on every single text

The Authors Guild is not a neutral testing body. It is an active litigant and policy advocate against AI companies, which means this benchmark serves a rhetorical purpose alongside any empirical one. The study's methodology, sample size, and text selection criteria are not described in the coverage, which makes the perfect-accuracy claims for Pangram and Grammarly as hard to trust as the catastrophic failure claims for ZeroGPT.

This is largely disconnected from recent activity in our archive, as we have no prior coverage of AI detection benchmarks or the Authors Guild's technical work. The story belongs to a broader, slower-moving debate about authentication infrastructure for text, one that has been building since large language models began producing fluent prose indistinguishable from professional writing. That debate spans academic integrity vendors, publishing platforms, and legal proceedings, none of which we have tracked yet.

Watch whether Pangram or Grammarly publish their own methodology responses to this benchmark within the next 60 days. If neither engages with the Authors Guild's test design publicly, that silence will tell you more about the study's rigor than the scores themselves.