Epoch AI: AI Text Detectors Fooled by Mimicry

Epoch AI: AI Text Detectors Misled by Mimicry

AI text detectors struggle when language models imitate an author's style.

A research team from Epoch AI has demonstrated that popular AI text detectors such as Pangram, GPTZero, and Originality.ai can identify AI-generated texts with nearly perfect accuracy under standard conditions. However, when AI models are prompted to mimic the writing style of a specific author using text samples, the detectors' performance drops significantly, with an average of 13% of generated passages going unnoticed.

Scientific writing is a weak point: detectors failed to flag between 24% and 29% of AI-generated content imitating the style in this category, raising concerns about the reliability of these tools in educational environments.

Popular AI text detectors almost perfectly identify simple AI-generated texts. But when language models deliberately copy the writing style of a specific author, up to one in five AI texts escapes detection. Scientific writing is where detectors fail the most.

A research team from Epoch AI tested 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 test covered three categories: authentic human writing, AI-generated text from simple prompts, and AI text that deliberately imitates the style of a specific author.

The team compiled a corpus of 495 human passages from 99 authors, evenly distributed among blogs, fiction, and scientific writing. All texts were written before the release of ChatGPT in November 2022, effectively eliminating contamination from language models.

When it comes to simple AI-generated text, all three detectors performed almost flawlessly, with a maximum false negative rate of 0.7%. Human texts were also correctly classified in most cases. Pangram and GPTZero did not produce a single false positive. However, Originality.ai flagged 19 of the 495 human passages as AI-generated, a concerning false positive rate of 3.8%.

Style Imitation Lowers Detection Rates

This result changes when language models receive writing samples from an author as reference material. For this test, three state-of-the-art models (Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro) were each given five passages of real text from an author and asked to write new texts in the same style.

Among the 297 passages generated in this manner, an average of 38 went unnoticed, according to Epoch AI, corresponding to a false negative rate of about 13%. Pangram missed 10% of the texts imitating the style, GPTZero missed 11%, and Originality.ai missed 18%.

Once AI imitates human authors, the error rates of the detectors skyrocket.

For fiction, the false negative rate for all detectors ranged between 1% and 5%. Scientific writing revealed a very different story. Pangram failed to detect 25% of the academic texts generated by AI imitating the style, GPTZero missed 24%, and Originality.ai missed 29%.

The worst individual results were observed in specific model-genre combinations within scientific writing. Pangram missed 48% of the academic passages generated by Gemini, according to the published data. At Originality.ai, 39% of the academic texts from GPT-5.5 went unnoticed.

Different Methods, Same Blind Spots

Pangram uses a neural network trained on both human and machine-generated texts, although its founder has described the system as a "black box" since its verdicts cannot be traced. GPTZero measures how predictable word choices are and how this varies within a text, based on the idea that language models write more uniformly than humans. Originality.ai looks for statistical patterns it learned during training on human and AI-generated texts.

Despite these differences, the three detectors show the same pattern. They almost always detect texts from simple prompts but miss imitations much more frequently. Scientific writing, the genre where AI detection is likely most utilized in the real world, remains the hardest to flag correctly.

A previous test by the Authors Guild revealed that Pangram and Originality.ai reliably classified human texts as human. The Epoch AI study completes the other half of this picture: a low false alarm rate on human writing says little about the number of AI texts that actually slip through the cracks.