AI Text Detectors Fail to Identify Mimicked Author Styles

A recent study by Epoch AI assessed the efficacy of three AI text detectors: Pangram, GPTZero, and Originality.ai.

The study found that up to 18 percent of AI-generated text imitating an author's style remained undetected by these tools.

The detectors showed a significant miss rate of up to 48 percent specifically in the genre of scientific writing.

The results highlight challenges faced by text detection technology when confronted with style-mimicking language models.

The findings raise concerns about the reliability of current AI text detectors in identifying AI-generated content accurately.

All coverage (1)