AI detectors are inconsistent and can unfairly flag human writing as AI-generated
a test by the authors guild found that tools from pangram and grammarly correctly identified all human-written texts, while sidekicker failed on every single article. this inconsistency poses risks to authors, as false positives can cost them contracts and damage their reputations.
the authors guild tested five ai detectors on articles from 2020-2022, finding that pangram and grammarly correctly identified all human writing, while sidekicker flagged every text as mostly ai-generated. despite the high performance of some tools, the guild warns that detectors are unstable black boxes that cannot reliably distinguish between a master human writer and a machine imitating that style. because professional writing shares statistical patterns with training data, skilled authors face the highest risk of false accusations that could void contracts. the results highlight a paradox where human expertise is penalized for resembling the output it helped train.