AI text detectors: imitation of writing styles becomes a blind spot

AI text detectors are designed to identify machine-generated content. However, a recent study by Epoch AI shows how strongly their reliability depends on how a text was produced. Pangram, GPTZero and Originality.ai were tested on 495 human-written passages known to have been published before the launch of ChatGPT, as well as 297 AI-generated texts produced from simple prompts and 297 AI-generated texts deliberately written to imitate the style of real authors.

For directly generated AI texts, the detectors performed almost flawlessly, correctly identifying them in 99% to 100% of cases. However, when the AI models were given five writing samples from an author and asked to imitate their style, detection rates fell significantly. Pangram missed 10%, GPTZero 11% and Originality.ai 18% of the style-imitating texts. Academic writing proved particularly difficult: depending on the detector, 24% to 29% of AI-generated texts went undetected. Ironically, academic writing – one of the areas where AI detection is likely to be used most frequently – was the hardest to identify.

The findings point to a fundamental problem: detectors do not identify authorship; they identify statistical patterns. When typical AI characteristics are weakened through style instructions, editing or paraphrasing, detection rates fall. With every new generation of models and every new method of rewriting text, the linguistic signals considered characteristic of AI-generated content may also change.

Earlier research has reached similar conclusions:

• A 2025 study showed, for example, that paraphrasing can drastically reduce the detection performance of several tools. The researchers examined 19 so-called humaniser and paraphrasing tools and found that many established detectors were considerably less effective at identifying rewritten AI-generated text.
• A further study from the previous year helps explain why. According to the researchers, paraphrasing shifts a text semantically away from its source material, while some linguistic generation patterns still remain. The result is a kind of grey area that detectors find difficult to classify.

What does this mean for the use of AI in professional research?

A detector score alone should never be treated as proof of either human or machine authorship. For professional research, provenance records, verifiable sources, version histories and transparent documentation of AI use provide more robust evidence. Detectors can, at most, serve as a warning signal that prompts further checks. What matters is not whether a text “sounds like AI”, but whether its claims are supported, its sources are traceable and its conclusions can be independently verified.

By Maria Kleiner