帮助 AI 文本检测器适应变化的环境

使用 Google、Facebook 或 Apple 账号继续,即表示你同意使用条款。
帮助 AI 文本检测器跟上不断变化的形势
由 Lisa Lock 编辑,Andrew Zinin 审阅
本文已按照 Science X 的编辑流程与政策进行审核。编辑在确保内容可信度的同时,着重指出了以下特征:
添加到首选来源
如果一位教师怀疑学生的学期论文可能是由大语言模型(LLM)写成的,可以使用 AI 生成文本检测工具来帮助印证自己的怀疑。但在完全信任检测器之前,这位教师或许应当三思。
来自 MIT 和康奈尔大学的研究人员表明,这类检测工具本身也是人工智能模型,却会被经过伪装、看起来像是人类所写的 AI 生成文本所欺骗。此外,由于写作习惯随时间自然变化或新 LLM 不断涌现,检测器也可能无法识别出 AI 文本。
在一项新研究中,研究人员展示了一种特殊的训练技术,如何让 AI 检测器在人类写作与 AI 写作不断变化的情况下依然更加准确,并对欺骗行为更具韧性。
他们采用了一种名为测试时自适应的 方法,该方法让已部署的检测模型能够通过即时调整其内部运作方式,从新的未标注数据中学习。这样一来,模型无需从头重新训练,就能适应未见过的输入。
在实验中,一个商用检测器只能识别出约 27% 经过伪装的 AI 生成文本,而新方法达到了 93% 的准确率。
这种方法可以帮助教育工作者、编辑、出版商和招聘经理标记出人们试图冒充为人类撰写的 LLM 生成内容。网络安全专家也可以用其来防范电子邮件钓鱼诈骗,或识别搜索引擎优化欺诈。
“从科学出版到短篇小说比赛,有许多高风险场景中,人们确实很关心作者身份。我们已经证明,即便 AI 用户持续进行对抗行为,测试时自适应也能让你的 AI 检测器稳健得多,”该研究合著者 Manish Raghavan 说。他是 MIT Sloan School of Management 的 Drew Houston(2005)职业发展教授,同时在 MIT Schwarzman College of Computing 与电子工程与计算机科学系(EECS)担任联合职位,并且是信息与决策系统实验室(LIDS)的首席研究员。
与他共同完成这篇论文的还有第一作者、Cornell Tech 研究生 Kevin Ren,以及资深作者、Cornell Tech 助理教授 Nikhil Garg。该研究将在神经信息处理系统大会(NeurIPS 2026)上展示,会议于 12 月 6 日至 13 日在悉尼、亚特兰大和巴黎举行。论文目前可在 arXiv 预印本服务器上获取。
能够生成类人文本的 LLM 大量涌现,催生了众多商业化的 AI 检测工具,帮助人们判断某篇文字是否由人类撰写。
但当这些检测工具所评估的文本与其被训练识别的那类文本存在显著差异时,它们就可能失效。这被称为分布偏移。
检测工具的开发者通常从互联网上收集他们已知由 AI 或人类生成的文本,以此训练模型,让模型学会区分二者。
然而,检测器一经发布,人们就可能改变自己的写作方式,或者改变使用 AI 生成文本的方式。
例如,2030 年的人可能会使用 2010 年的写作者并不常用的新类比来讨论不同的话题。用 2010 年数据训练的检测器可能对这种新的写作风格并不熟悉,从而错误地把人类撰写的文本标记为 AI 生成。
检测模型的发布也在其开发者与想要绕过它的用户之间形成了一场猫鼠游戏。这些人可以把 AI 生成的文本输入自动化的“humanizer”(人性化)软件工具,将其修改得在检测器看来像是人类所写。
“公司发布模型后,一大批‘humanizer’就会冒出来试图击败它,于是公司又尝试把这些经过人性化处理的文本纳入训练,好让检测器能够识别出来,如此往复。就像反复冲洗再重复一样。这个循环可以永远持续下去,”Ren 说。
研究人员探讨了测试时自适应(TTA)能否让 AI 检测器在面对这些分布偏移时保持稳健。
在 TTA 中,研究人员不是把模型冻结起来、让它在部署期间无法变化,而是允许模型在发布之后自行适应新到来的数据。对于 AI 文本检测器来说,这些数据就是用户上传的新文本。
这些数据没有标注,因此检测器的开发者并不知道哪些数据由人类撰写、哪些由 AI 模型撰写。但这些新数据代表了当下的人类写作和 AI 写作——正是开发者希望模型能够分类的内容。
有了足够多的样本,检测器就能学会识别模式,判断哪些数据由 AI 生成、哪些由人类生成。
研究人员通过构建包含一定数量 AI 生成文本和人类生成文本的数据集来研究 TTA,但没有把这些标签提供给模型。他们还用 LLM 生成了一些旨在规避检测器的 AI 撰写文本样本。
随后他们执行了 TTA 训练流程,并在真值数据上测试该模型。
在模拟人类语言随时间自然变化的实验中,采用 TTA 的检测器准确率为 85%,而传统方法的准确率只有 73%。
而在针对刻意规避检测工具的文本时,现有方法只能识别出约 27%,TTA 则让模型能够捕捉到超过 93% 的伪装文本。
Ren 表示:“TTA 有望真正帮助已部署的 AI 检测模型,使其能够主动应对对抗性 AI 用户这种不断迭代的‘博弈’行为。”
他补充说,适应人类写作行为的变化是一个更困难的 TTA 问题,因为这些变化更为缓慢,且以更微妙的方式发生。
研究人员提醒,TTA 未必始终是有效的解决方案,因为未来的大语言模型会生成听起来更自然的语言,而找到可用于训练的可验证人类文本也会变得更难。
“目前还不完全清楚这是否是一个可以解决的问题,”Raghavan 说。“在一个无法分辨人类生成内容与 AI 内容的世界里,你该如何看待科学、写作和艺术,这存在一些重大的未解问题。”
未来,研究人员希望探索 TTA 可能同样适用的类似战略领域,并构建更好的基准测试工具,帮助工程师动态改进这些 AI 检测模型。
Kevin Ren et al., Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift, arXiv (2026). DOI: 10.48550/arxiv.2606.25152
期刊信息:arXiv
这篇报道的作者是谁?
艺术史学士,物质文化硕士。曾任博物馆编辑、急救员和器官移植协调员。自 2021 年起在 Science X 担任编辑。完整简介 →
物理学硕士,有科研经历。资深科学新闻爱好者。在 Science X 的编辑工作中发挥关键作用。完整简介 →
本文经 MIT News(web.mit.edu/newsoffice/)授权转载,该网站是一个报道 MIT 研究、创新与教学新闻的热门网站。
引用:帮助 AI 文本检测器跟上不断变化的条件(2026 年 11 月 18 日),2026 年 11 月 18 日检索自 https://techxplore.com/news/2026-11-ai-text-detectors-conditions.html
本文受版权保护。除为私人学习或研究目的进行的合理使用外,未经书面许可不得转载任何部分。内容仅供参考之用。
检测机器生成文本:一场随大型语言模型进步而升级的军备竞赛
人类作者还是 AI?学者打造了一款检测工具
AI 生成的内容真的能被检测出来吗?
观察检测器:研究人员探究 AI 检测工具的有效性——以及危险性
专家通过寻找人类特有的行文习惯(如用词变化和复杂句式)来检测 AI 生成的文本
让氢火焰更容易被看见,或有助于推动更清洁的能源未来
新 AI 方法揭示视觉模型识别出哪些概念,以及这些概念出现在何处
电子指纹帮助 AI 在训练数据很少的情况下预测复杂合金的性能
机器人学习人类运动模式,以便在密集人群中找到通行路径
这种微型磁体可将普通晶体管电路变成人工神经元
单螺旋桨无人机利用自身旋转向各个方向扫描周围环境
获取即时摘要文本(Gist)
原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Helping AI text detectors keep up with changing conditions)
By continuing with Google, Facebook or Apple, you agree to the Terms of Use.
Helping AI text detectors keep up with changing conditions
edited by Lisa Lock, reviewed by Andrew Zinin
This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:
Add to Preferred Sources
Credit: Matheus Bertelli from Pexels
A teacher who believes a student's term paper may have been written by a large language model (LLM) can use a tool for detecting AI-generated text to help confirm their suspicions. But the teacher may want to think twice before putting all their trust in the detector.
Researchers from MIT and Cornell University showed that such detection tools, which are artificial intelligence models themselves, can be fooled by AI-generated text disguised to appear as if it were written by a human. Detectors can also fail to identify AI text due to natural shifts in writing habits over time or the emergence of new LLMs.
In a new study, the researchers demonstrated how a special training technique makes an AI detector more accurate and resilient to deception, even as human and AI writing constantly change.
They employed a method called test-time adaptation, which allows a deployed detection model to learn from new, unlabeled data by adjusting its inner workings on the fly. In this way, the model adapts to unseen inputs without being retrained from scratch.
In experiments, a commercial detector could identify only about 27% of AI-generated text that had been disguised, while the new approach achieved 93% accuracy.
This method could help educators, editors, publishers and hiring managers flag LLM-generated writing that people try to pass off as human-written. It could also be used by cybersecurity experts to prevent email phishing scams or identify search engine optimization fraud.
"There are many high-stakes contexts where people really care about the identity of the author, from scientific publishing to short story competitions. We've shown that test-time adaptation allows your AI detector to be much more robust, even with continual adversarial behavior by AI users," says study co-author Manish Raghavan, the Drew Houston (2005) career development professor at the MIT Sloan School of Management, who also holds an MIT Schwarzman College of Computing shared position in the Department of Electrical Engineering and Computer Science (EECS) and is a principal investigator in the Laboratory for Information and Decision Systems (LIDS).
He is joined on the paper by lead author Kevin Ren, a graduate student at Cornell Tech, and senior author Nikhil Garg, an assistant professor at Cornell Tech. The research will be presented at the Conference on Neural Information Processing Systems (NeurIPS 2026), held in Sydney, Atlanta, and Paris, Dec. 6–13. The paper is currently available on the arXiv preprint server.
The proliferation of LLMs that generate humanlike text has led to a plethora of commercial AI-detection tools to help people identify whether a piece of writing was authored by a human.
But these detection tools can fail when the text they evaluate differs significantly from the text they were taught to identify. This is known as a distribution shift.
Detector developers typically train their models by gathering text from the internet that they know to be either AI- or human-generated. They train the model to tell the difference.
However, once the detector is released, people might change the way they write or the way they use AI to generate text.
For instance, a human in 2030 may write about different topics using new analogies that weren't common for someone writing in 2010. A detector trained with 2010 data might be unfamiliar with this new writing style and incorrectly mark human-written text as AI-generated.
The release of a detection model also creates a cat-and-mouse game between its developers and users who want to bypass it. These individuals can run AI-generated text through an automated "humanizer" software tool that modifies it to appear human-written to the detector.
"After the company releases a model, a bunch of 'humanizers' pop up to try to beat it, so the company tries to incorporate this humanized text into their training so the detector can identify it, and so on. It's like rinse-and-repeat. The cycle can go on forever," says Ren.
The researchers explored whether test-time adaptation (TTA) can make an AI detector resilient to these distribution shifts.
In TTA, instead of freezing the model so it can't change during deployment, researchers allow that model to adapt itself to incoming data after it has been released. For an AI text detector, these data are new pieces of text uploaded by users.
These data are unlabeled, so the detector developers don't know which data were written by humans versus AI models. But these new data represent contemporary human and AI writing—exactly what the developers want the model to be able to classify.
With enough examples, the detector learns to identify patterns to determine which data are AI-generated and which are human-generated.
The researchers studied TTA by creating data sets containing a certain amount of AI-generated and human-generated text, but they didn't provide those labels to the model. They also used LLMs to generate examples of AI-written text designed to evade a detector.
Then they performed the TTA training process and tested the model on ground-truth data.
In experiments designed to mimic the natural shift in human language over time, a detector that used TTA was 85% accurate, while traditional methods achieved only 73% accuracy.
And when it came to writing designed to thwart detection tools, existing methods could identify only about 27%, while TTA enabled the model to catch more than 93% of disguised text.
"TTA has the potential to really help deployed AI-detection models in terms of being able to proactively address this iterative 'gaming' behavior by adversarial AI users," Ren says.
Adapting to shifts in human writing behavior is a more difficult TTA problem, since the shifts are slower and occur in more subtle ways, he adds.
The researchers caution that TTA may not always be an effective solution as future LLMs produce more natural-sounding language and it becomes harder to find verifiable human text to use for training.
"It isn't entirely clear if this will be a solvable problem," Raghavan says. "There are some big, open questions about how you should think about science, writing and art in a world where it is infeasible to tell the difference between human-generated content and AI content."
In the future, the researchers want to explore similar strategic domains where TTA could be useful and build better benchmarking tools to help engineers improve these AI detection models dynamically.
Kevin Ren et al., Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift, arXiv (2026). DOI: 10.48550/arxiv.2606.25152
Journal information:arXiv
Who's behind this story?
BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021. Full profile →
Master's in physics with research experience. Long-time science news enthusiast. Plays key role in Science X's editorial success. Full profile →
This story is republished courtesy of MIT News (web.mit.edu/newsoffice/), a popular site that covers news about MIT research, innovation and teaching.
Citation: Helping AI text detectors keep up with changing conditions (2026, November 18) retrieved 18 November 2026 from https://techxplore.com/news/2026-11-ai-text-detectors-conditions.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.
Detecting machine-generated text: An arms race with the advancements of large language models
Human writer or AI? Scholars build a detection tool
Is AI-generated content actually detectable?
Watching the detectors: Researchers probe efficacy—and danger—of AI detection tools
Experts detect AI text by looking for human idiosyncrasies, like word variation and complex sentences
Making hydrogen flames easier to see could help power a cleaner energy future
New AI method reveals which concepts vision models recognize and where they appear
Electron fingerprints help AI predict complex alloy properties with little training data
Robots learn human movement patterns to find paths through dense crowds
This tiny magnet could turn ordinary transistor circuits into artificial neurons
Single-propeller drone uses its own spin to scan surroundings in all directions
Get Instant Summarized Text (Gist)
Test-time adaptation improved AI text detection by updating models with new, unlabeled text. Detection of disguised AI text rose from about 27% to over 93%, while accuracy under simulated language shifts increased from 73% to 85%. Effectiveness may decline as AI writing becomes harder to distinguish from human text.
This summary was automatically generated using LLM. Full disclaimer
This site uses cookies to assist with navigation, analyse your use of our services, collect data for ads personalisation and provide content from third parties. By using our site, you acknowledge that you have read and understand our Privacy Policy and Terms of Use.
原文为英文,正文由自动抓取并保留原文。
原文发布于 TechXplore(Helping AI text detectors keep up with changing conditions),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。
