AI文本检测器在模仿作者风格时漏检率飙升

流行的 AI 文本检测器几乎可以以接近完美的准确率检测普通的 AI 生成文本。但当语言模型有意模仿特定作者的写作风格时，多达五分之一的 AI 文本会逃脱检测。科学写作是检测器最容易失败的领域。

Epoch AI 的一个研究团队测试了三种最广泛使用的 AI 文本检测器：Pangram（版本 3.3.2）、GPTZero（模型 2026-05-11-base）和 Originality.ai（Turbo 3.0.2）。测试涵盖三个类别：真实的人类写作、从简单提示生成的 AI 文本，以及故意模仿特定作者风格的 AI 文本。

该团队构建了一个包含 99 位作者的 495 段人类文本的语料库，平均分布在博客、小说和科学写作中。所有文本均在 ChatGPT 于 2022 年 11 月发布之前写成，这有效地排除了被语言模型污染的可能性。

在处理普通 AI 生成文本时，所有三种检测器的表现几乎无懈可击：https://epoch.ai/data-insights/ai-text-detection-challenges，假阴性率最高仅为 0.7%。大部分人类文本也能被正确分类。Pangram 和 GPTZero 没有产生任何误报。然而，Originality.ai 对 495 段人类文本中有 19 段误判为 AI 生成，假阳性率高达 3.8%，令人担忧。

当语言模型获得作者的写作样本作为参考资料时，这个结果就会发生变化。对于本次测试，三款前沿模型（Claude Opus 4.8、GPT-5.5 和 Gemini 3.1 Pro）各自收到一位作者的五段真实文本，然后被要求以相同风格撰写新文本。

根据 Epoch AI 的数据，在通过这种方式生成的 297 段文本中，平均有 38 段逃脱了检测，相当于约 13% 的假阴性率。Pangram 漏_detect 了 10% 的风格模仿文本，GPTZero 漏_detect 了 11%，Originality.ai 漏_detect 了 18%。

对于小说而言，三种检测器的假阴性率仅为 1% 至 5%。科学写作则表现出完全不同的情况。Pangram 未能识别出 25% 的风格模仿学术 AI 文本，GPTZero 漏_detect 了 24%，Originality.ai 漏_detect 了 29%。

在科学写作中，最糟糕的个体表现出现在特定的模型-文类组合中。根据已发布的数据，Pangram未能检测到48%的Gemini生成的学术段落：https://github.com/jaeholee-brown/ai-text-detectors。在Originality.ai上，39%的GPT-5.5学术文本未被检测到。Ad DEC_D_Incontent-2

Pangram使用一个在人工和机器生成文本上训练的神经网络，尽管其创始人曾称该系统为黑箱：https://the-decoder.com/pangram-ceo-says-language-models-give-themselves-away-by-making-the-same-arguments/，因为其判决无法追踪。GPTZero测量词语选择的可预测性及其在文本中的变化程度，基于语言模型比人类写得更统一的理念。Originality.ai：http://originality.ai/ 搜索其训练过程中从人工和AI生成文本中学到的统计模式。Ad

尽管存在这些差异，三种检测器表现出相同的模式。它们几乎每次都能捕捉到简单提示生成的文本，但更容易漏掉模仿文本。科学写作这一类型可能是AI检测在实际应用中使用最频繁的：https://the-decoder.com/grades-dropped-from-96-to-48-percent-when-a-brown-professor-made-students-take-the-exam-without-ai/，仍然是最难正确标记的。

早期的作者协会测试：https://the-decoder.com/authors-guild-test-finds-some-ai-detectors-perfectly-identify-human-writing-while-others-fail-on-every-single-text/ 发现，Pangram和Originality.ai能够可靠地将人类文本归类为人类写作。Epoch AI的研究补充了图景的另一半：对人类写作的假阳性率低，对AI文本实际漏检的数量几乎没有说明意义。

保持对AI的关注。清晰、有用、不空洞。

关注The Decoder获取AI新闻、背景故事和专家分析。

The Decoder：https://the-decoder.com/

Popular AI text detectors catch plain AI-generated text with near-perfect accuracy. But when language models deliberately copy a specific author's writing style, up to one in five AI texts slips through undetected. Scientific writing is where the detectors fail the hardest.

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: genuine human writing, AI text generated from simple prompts, and AI text that deliberately mimicked a specific author's style.

The team built a corpus of 495 human passages from 99 authors, evenly split across blogging, fiction, and scientific writing. All texts were written before ChatGPT's release in November 2022, which effectively rules out contamination by language models. Ad

When dealing with plain AI-generated text, all three detectors performed almost flawlessly：https://epoch.ai/data-insights/ai-text-detection-challenges, with the false-negative rate topping out at 0.7 percent. Human texts were also classified correctly for the most part. Pangram and GPTZero didn't produce a single false alarm. Originality.ai, however, flagged 19 out of 495 human passages as AI-generated, a troublingly high false-positive rate of 3.8 percent. Ad DEC_D_Incontent-1

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

Of the 297 passages generated this way, an average of 38 went undetected, according to Epoch AI, which works out to a false-negative rate of about 13 percent. Pangram missed 10 percent of style-imitated texts, GPTZero missed 11 percent, and Originality.ai missed 18 percent. Ad

For fiction, the false-negative rate across all detectors sat at just 1 to 5 percent. scientific writing told a very different story. Pangram failed to catch 25 percent of style-imitated academic AI texts, GPTZero missed 24 percent, and Originality.ai missed 29 percent.

The worst individual results showed up in specific model-genre combinations within scientific writing. Pangram missed 48 percent of Gemini-generated academic passages, according to the published data：https://github.com/jaeholee-brown/ai-text-detectors. At Originality.ai, 39 percent of academic GPT-5.5 texts went undetected. Ad DEC_D_Incontent-2

Pangram uses a neural network trained on human and machine-generated text, though its founder has called the system a black box：https://the-decoder.com/pangram-ceo-says-language-models-give-themselves-away-by-making-the-same-arguments/ since its verdicts can't be traced. GPTZero measures how predictable word choices are and how much that varies within a text, based on the idea that language models write more uniformly than humans. Originality.ai：http://originality.ai/ searches for statistical patterns it learned during training on human and AI-generated text. Ad

Despite these differences, all three detectors show the same pattern. They catch text from simple prompts almost every time but miss imitations far more often. Scientific writing, the genre where AI detection probably sees the most real-world use：https://the-decoder.com/grades-dropped-from-96-to-48-percent-when-a-brown-professor-made-students-take-the-exam-without-ai/, remains the hardest to flag correctly.

An earlier Authors Guild test：https://the-decoder.com/authors-guild-test-finds-some-ai-detectors-perfectly-identify-human-writing-while-others-fail-on-every-single-text/ found that Pangram and Originality.ai reliably classified human texts as human. The Epoch AI study fills in the other half of that picture: a low false-alarm rate on human writing says little about how many AI texts actually slip through.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/