A research team at Epoch AI has demonstrated that widely used AI text detectors, including Originality.ai and Pangram, face significant challenges when language models are trained to replicate the writing style of specific authors. These detectors typically perform well under standard conditions but show a marked decline in accuracy when confronted with AI-generated text that mimics human authorship.

The study involved testing three of the most widely used AI text detectors: Pangram (version 3.3.2), GPTZero (model 2026-05-11-base), and Originality.ai (Turbo 3.0.2). The researchers found that when AI models are prompted to imitate a specific author's writing style using text samples, the detectors' performance drops significantly.

Under these conditions, the detectors failed to identify AI-generated text in approximately 13 percent of cases. This figure highlights a critical vulnerability in current detection systems, which are otherwise capable of identifying AI-generated text with near-perfect accuracy under standard conditions.

This flaw has significant implications for the reliability of AI text detection tools. It raises concerns about the ability of these systems to effectively distinguish between human and AI-generated content in scenarios where AI models are trained to replicate specific writing styles. This could lead to increased instances of undetected AI-generated text being used in academic, journalistic, and professional settings.

The findings suggest that current AI text detection systems may not be robust enough to handle the evolving capabilities of language models. As AI continues to advance, the need for more sophisticated detection mechanisms becomes increasingly apparent, particularly in environments where the authenticity of written content is critical.