AI检测器正在开启一个充满不信任的新时代

qimuai 发布于 阅读：1 一手编译

AI检测器正在开启一个充满不信任的新时代

内容来源：https://www.theverge.com/column/976690/ai-writing-detectors-suspicion

内容总结：

AI检测工具引发“不信任时代”：误判频发，教育界开始反思

随着ChatGPT等生成式人工智能工具的普及，一场针对“AI写作”的排查战正在全球教育界和内容创作领域蔓延。然而，一枚名为“AI检测器”的双刃剑，非但未能有效遏制学术不端，反而因其频繁的误判，正将我们推向一个充满猜忌与对立的新时代。

从查重到“猎巫”：检测工具不可靠却泛滥

曾几何时，教育工作者依靠查重软件比对文本相似度。如今，这场“打假”行动已升级为对AI生成内容的全面围剿。调查显示，美国43%的中学教师定期使用AI检测工具。然而，这些工具自身的准确性却饱受质疑。它们通过分析文本的措辞、节奏和结构等主观特征进行判断，极易将非母语写作者或神经多样性人群的原创内容误判为AI生成。尽管包括Turnitin在内的多家公司宣称其误报率极低，但斯坦福大学的研究已证实，非英语母语者的文章被误判的概率远高于母语者。

代价沉重：误判毁掉声誉与学业

在“风声鹤唳”的氛围下，AI写作指控正造成真实而严重的伤害。出版商Minotaur因担忧作者使用AI而取消了价值200万美元的图书合同；法国学生Thierry Rignol因被指控使用AI而被耶鲁大学处分，其诉讼明确指出检测工具对非母语人士存在系统性偏见。知名博主杰克·奥斯bourne也因轻信检测结果，公开指责记者使用AI写作，即便事后被证伪也拒绝撤回，给当事人带来巨大困扰。这揭示了一个残酷现实：众多指控者忽视了检测工具自身的免责声明。

转向理性：高校弃用工具，倡导教学改革

面对检测工具的不确定性，多所顶尖高校已开始采取行动。耶鲁、约翰霍普金斯、麻省理工等名校已禁用或限制使用AI检测工具，麻省理工更是直言“AI检测器不奏效”。教育界正从依赖技术转向重构教学法，例如鼓励课堂内评估、拆分长作业、引导学生进行反思性写作，并为学生提供主动披露AI使用情况而不受罚的渠道。这些举措旨在重建信任，而非单纯防范。

未来展望：在信任与怀疑间寻求平衡

随着AI深度嵌入生活，社交媒体平台也开始内置AI检测功能，这进一步加剧了公众对内容的普遍怀疑。这场“AI幻觉”引发的信任危机，迫使真正的人类写作者努力“证明”自己的作品非机器所为。尽管出现了“人类创作”认证等自我证明手段，但根本出路在于：承认AI检测工具的局限性，重新审视教育评估与内容创作的价值，在技术洪流中守住人文理性的底线。

中文翻译：

这是《The Stepback》周刊，每周为你解读科技界一个关键故事。想了解更多关于AI如何改变日常生活的新闻，请关注Emma Roth。《The Stepback》于美国东部时间上午8点发送至订阅者邮箱。点此订阅《The Stepback》。

AI检测工具正在制造一个新的不信任时代

那些声称能检测AI写作的工具可能并不可靠，但教育工作者和出版商仍在使用它们。

AI检测工具正在制造一个新的不信任时代

那些声称能检测AI写作的工具可能并不可靠，但教育工作者和出版商仍在使用它们。

这一切是如何开始的

早在ChatGPT流行之前，教育工作者和编辑就经常使用反抄袭工具来检查写作者是否诚实地完成了自己的作品。这些工具的工作原理是将写好的作品与一个包含全网内容、学术文章等的数据库进行比对，以查找相同的句子和短语。有些工具（如Turnitin）会给出一个百分比，声称能显示学生的写作与其他人作品的重合度。由于存在误报的可能性，以及无法确定作品是否为故意抄袭，一些教育工作者已经不再使用这一特定工具。

但现在，对抄袭内容的追查正演变为一场对AI生成内容的战争。学生使用ChatGPT、Google Gemini和Microsoft Copilot的速度有多快，教师采用所谓的AI检测工具就有多快。民主与技术中心的一项调查发现，在2024年至2025年间，美国有43%的六年级至十二年级教师定期使用AI检测工具。一些已经在学习管理系统中使用Turnitin的大学发现，该服务在2023年推出AI检测功能时自动启用了该功能。

与比较书面内容不同，GPTZero、Pangram以及Turnitin开发的AI检测工具依靠自身的人工智能模型来猜测某段文字是否可能不是人类写的——这个过程可以说比在网上匹配文本更加模糊。正如GPTZero所指出的，AI检测工具使用算法来分析文本的措辞、节奏和结构，并捕捉在长度和语气上可能更常见于AI生成文本的模式。这种相对主观的评估不像你在网上匹配文本那样有据可依，而且容易被英语非母语的写作者打乱判断。尽管如此，Turnitin表示其AI检测工具将人类写作内容误判为AI的比例不到1%，而Pangram声称其误报率仅为万分之一。GPTZero则声称其将人类内容误判为AI的比例同样很低。

现在的情况如何

网上已经有人在互相指责对方“听起来像AI”，而AI检测工具的随处可得更是给这场针对大语言模型的猎巫行动火上浇油。在一些备受关注的案例中，AI写作指控直接影响了人们的生计和声誉。上个月，出版商Minotaur因担心作者Jerry Falade使用了AI而撤销了一笔200万美元的出书合约——Falade对此坚决否认。

还有法国人Thierry Rignol，他去年起诉了耶鲁大学，起因是一位教授指责他在期末考试的部分内容中使用了AI，导致他不及格并被停学一年。该教授使用GPTZero扫描了Rignol的写作以寻找AI痕迹，但诉讼认为“AI监控和检测工具以不公平地针对非英语母语者而闻名”，Rignol正是其中之一。今年2月，阿德尔菲大学的一名学生赢得了对学校的诉讼，此前他的教授同样声称他用AI写了文章。虽然诉讼中没有说明教授用哪种AI工具检查了学生的文章，但阿德尔菲大学与Turnitin签有许可协议。

2023年斯坦福大学的一项研究发现，AI检测工具将非英语母语者写的文章误判为AI的频率高于母语者。（许多服务仍声称其工具在处理非母语者写的文本时是准确的。）这些工具还可能对神经多样性写作者存在偏见。

正如加州大学洛杉矶分校所指出的，AI检测工具经过训练来捕捉可能表明使用AI的模式，例如重复的术语和短语、听起来过于正式或过于随意的文本，以及不合逻辑的措辞。据加州大学洛杉矶分校介绍，有些工具（如QuillBot）还会衡量文本的“不可预测性”，因为“与人类写作相比，AI倾向于做出最‘明显’或最常见的语言选择”。它们还可能会寻找通篇保持相同的句子结构，作为AI的另一个迹象。但这些衡量标准本身并不能说明文本就是AI生成的，因为有些人可能恰好就具有这些特征的写作风格。

尽管Turnitin宣传其误报率很低，但它也表示其工具“可能并不总是准确的”，不应被用来对学生采取处分措施。Grammarly警告用户“绝不应单独依赖AI检测工具的结果”，而GPTZero则表示“没有哪个AI检测工具能真正做到100%完美”。OpenAI甚至因准确率低而在2023年关闭了自己的AI写作检测器。

但AI写作指控仍在网络上满天飞。上周，奥兹·奥斯本的儿子杰克在一段面向其社交频道超过350万粉丝的视频中，指控记者兼The Verge撰稿人Kat Tenbarge使用AI为《滚石》杂志撰写了一篇文章，并炫耀AI检测工具Getsolved的结果作为其指控的“证据”。Tenbarge在视频和她网站上的帖子中驳斥了这一指控，但奥斯本没有撤回指控，也没有删除视频，留下Tenbarge独自应对网上的喷子。

从作家到学生，虚假指控的例子数不胜数，许多指控者都没有注意到这些AI检测工具附带的一些免责声明。

接下来会发生什么

围绕AI检测工具的不确定性足以让一些教育机构完全停止使用它们。耶鲁大学、约翰霍普金斯大学、范德堡大学、乔治城大学等已禁用或限制使用AI检测工具。麻省理工学院也警告称“AI检测工具不起作用”。

许多学校没有依赖工具来筛除AI内容，而是鼓励教育工作者重新思考他们的教学方式。例如，芝加哥大学建议告诉学生放慢阅读速度、将较长的写作任务分解成小块，并要求学生对他们的作品进行反思。斯坦福大学表示教授可以考虑在教室内进行考核，而麻省理工学院则建议教授留出空间，让学生可以在不受处罚的情况下披露是否在作业中使用了AI帮助。

随着AI在课堂内外变得越来越普遍，辨别内容出自人类还是机器的努力也在加强。一些在线平台正在加剧对AI使用的怀疑。Substack已将Pangram集成到其应用中，允许用户扫描博客以发现疑似AI生成的内容，而LinkedIn在帖子中增加了一个“看起来像AI垃圾”的按钮。结果是进入了一个新的不信任时代，读者不断质疑他们读到的东西是否为AI生成，而真正的人类写作者则竭尽全力让自己的文字听起来不像AI。

顺便一提

延伸阅读

热门文章

英文来源：

This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more news about how AI is changing our daily lives, follow Emma Roth. The Stepback arrives in our subscribers’ inboxes at 8AM ET. Opt in for The Stepback here.
AI detectors are creating a new era of distrust
Tools that promise to detect AI writing can be dubious, but educators and publishers are using them anyway.
AI detectors are creating a new era of distrust
Tools that promise to detect AI writing can be dubious, but educators and publishers are using them anyway.
How it started
Long before ChatGPT became a thing, educators and editors frequently used anti-plagiarism tools to see if writers were being honest about their work. These tools work by comparing a written work against a database filled with content from across the web, scholarly articles, and more to check for matching sentences and phrases. Some, like Turnitin, offer a percentage that claims to illustrate how much of the student’s writing overlaps with other works. Between potential false positives and uncertainty about whether work was duplicated intentionally, some educators have backed away from using that particular tool.
But now, the hunt for copied content is evolving into a war on AI-generated work. As quickly as students have picked up ChatGPT, Google Gemini, and Microsoft Copilot, teachers have adopted so-called AI detectors just as fast. A survey from the Center for Democracy and Technology found that 43 percent of sixth to 12th grade teachers in the US regularly used AI detectors between 2024 and 2025. Some universities already using Turnitin in their learning management systems found that the service automatically enabled AI detection when the tool launched in 2023.
Instead of comparing pieces of written content, AI detectors like GPTZero, Pangram, and the one created by Turnitin rely on their own AI models to guess whether something might not be human-written — a process that’s arguably even murkier than matching text on the web. As noted by GPTZero, AI detectors use an algorithm to analyze a text’s wording, rhythm, and structure, as well as to pick up on patterns in length and tone that may be more common in AI-written text. This relatively subjective evaluation isn’t as solid as something you could back up by matching text online, and can get tripped up by writers who speak English as a second language. Despite this, Turnitin has said that its AI detector falsely flags less than 1 percent of human-written content as AI, while Pangram claims its false positive rate is just 1 in 10,000. GPTZero claims it has a similarly low rate of mistaking human content for AI.
How it’s going
People online are already accusing each other of “sounding like AI,” but the ready availability of AI detection tools is only adding fuel to the LLM witch hunt. In some high-profile cases, AI writing accusations have directly impacted people’s livelihoods and reputations. Last month, the publisher Minotaur dropped a $2 million book deal over concerns that its author, Jerry Falade, used AI — something he vehemently denies.
There’s Thierry Rignol, a French national who sued Yale last year after a professor accused him of writing portions of his final exam with AI, resulting in a failing grade and a one-year suspension. The professor used GPTZero to scan Rignol’s writing for signs of AI, but the lawsuit argues that “AI surveillance and detection tools are known to unfairly target non-native English speakers” like Rignol. In February, a student at Adelphi University won a lawsuit against the school after his professor similarly claimed he used AI to write an essay. Though the lawsuit doesn’t say which AI tool the professor used to examine the student’s essay, Adelphi University has a licensing agreement with Turnitin.
A 2023 Stanford study found that AI detectors falsely flagged essays written by non-native English speakers as AI more often than native speakers. (Many services still argue that their tools are accurate when dealing with text written by non-native speakers.) These tools may also be biased against neurodivergent writers.
As pointed out by the University of California, Los Angeles, AI detection tools are trained to pick up on patterns that could indicate AI use, such as repetitive terms and phrases, text that sounds too formal or informal, and nonsensical phrasing. Some, like QuillBot, also measure the “unpredictability” of text, as “AI tends to make the most ‘obvious’ or most common language choices as compared with human-produced writing,” according to UCLA. They may also look for sentence structure that remains the same throughout as another sign of AI. But these measurements aren’t indicative of AI on their own, as some people may just have a writing style with these qualities.
Even though Turnitin touts low false positive rates, it maintains that its tool “may not always be accurate” and shouldn’t be used to take actions against a student. Grammarly warns that users “should never rely on the results of an AI detector alone,” while GPTZero says “no AI detector can ever truly be 100% perfect.” OpenAI even shut down its own AI writing detector in 2023 due to low accuracy.
But AI writing accusations are still being flung across the web. Last week, in a video broadcast to the more than 3.5 million followers across his social channels, Ozzy Osbourne’s son, Jack, accused journalist and Verge contributor Kat Tenbarge of using AI to write an article for Rolling Stone, while flaunting the results from an AI detector, Getsolved, as “proof” of his claim. Tenbarge has refuted the claim in a video and a post on her website, but Osbourne hasn’t retracted his accusation or deleted the video, leaving Tenbarge to deal with the trolls.
There are numerous examples of false accusations, from writers to students, with many accusers failing to acknowledge the disclaimers that come along with some of these AI detection tools.
What happens next
The uncertainty surrounding AI detectors is enough for some educational institutions to stop using them altogether. Yale University, Johns Hopkins University, Vanderbilt University, Georgetown University, and others have disabled or restricted the use of AI detection tools. The Massachusetts Institute of Technology also warns that “AI detectors don’t work.”
Instead of relying on tools to weed out AI, many schools are encouraging educators to rethink their lessons. The University of Chicago, for example, suggests telling students to slow down their reading, breaking up longer writing assignments, and requiring students to reflect on their work. Stanford University says professors can consider holding assessments in classrooms, while MIT advises professors to leave room for students to disclose whether they used AI for help on an assignment, without penalty.
With AI becoming more prevalent in and outside the classroom, efforts to suss out what’s written by a human or a machine are ramping up as well. Some online platforms are only exacerbating suspicions surrounding AI use. Substack has built Pangram into its app, allowing users to scan blogs for suspected AI-generated content, while LinkedIn added a “seems like AI slop” button on posts. The result is a new era of distrust, where readers constantly question whether what they’re reading is AI and real human writers try their best not to sound like it.
By the way

ThevergeAI大爆炸

文章目录

    📚 推荐阅读


    扫描二维码，在手机上阅读
    版权所有：启幕AI
    文章标题：AI检测器正在开启一个充满不信任的新时代
    文章链接：https://news.qimuai.cn/?post=4764
    本站文章均为原创，未经授权请勿用于任何商业用途