The anti-AI writing police are everywhere. What are they looking for?

showcake/iStockphoto/Getty Images

On October 6, the Zimbabwean author Simbarashe Kundizeza won the Caine Prize for African Writing, for the short story “The Blue Fighter.” Less than a week later, The Atlantic published a piece declaring that the story’s narrator “has all the personality of a chatbot.”

“The AI detector Pangram flags almost the entire story as AI-generated,” the article reported.

Kundizeza denied that he had used any such technology. “For the avoidance of doubt, I wrote ‘The Blue Fighter.’ I did not use generative AI to write it. It is entirely my own work,” he wrote in a statement on X.

The Caine Prize, which comes with a £10,000 award, had become the newest front in a war engulfing the literary world, powered by AI writing tools and AI writing detectors.

Scandals about suspected AI use have spread up and down the writing industry this year. In March, the publisher Hachette canceled the US release of author Mia Ballard’s self-published 2025 novel “Shy Girl” after a New York Times report found that it may have been written using AI.

In May, a Trinidadian author named Jamir Nazir was accused of using AI to write “The Serpent in the Grove,” which won the Commonwealth Short Story Prize.

In July, researchers at Stony Brook University posited that H.M. Wolfe’s “Daggermouth,” a national bestseller, was 60% AI-written according to a data set shared with The Atlantic .

More accusations kept flying: debut author Jerry Falade lost a $2 million book deal in July after his agents pulled his debut novel over suspicions he had used AI writing tools. In September, a claim that a book by Haitian Canadian author Thélyson Orélien had failed an AI detection test led the jury for the prestigious Goncourt prize to remove him from its longlist.

Joel Saget/AFP/Getty Images

As generative AI proliferates, it’s easier than ever for an author, or someone who wants to be seen as an author, to come up with a story- or book-shaped piece of text. It’s harder than ever for a reader or a prize judge or a publisher to be sure that the words on a page were produced by human thought and effort, rather than a machine prompt.

The result is a literary version of the ruined world in the movie “Blade Runner.” Anxiety and suspicion about robot infiltrators are everywhere — and so are people claiming to hunt the robots down.

Multiple companies now offer products that promise to spot the difference, with often contradictory, biased or erratic results. The makers of Pangram, an AI detection platform that has been cited as authoritative in many of the recent controversies, claim it can distinguish between human and synthetic writing with near-perfect accuracy.

“We’re trying to mitigate some of the harms that come from these powerful AI models,” said Max Spero, the CEO and founder of Pangram.

Spero is quick to admit that he doesn’t have a background in literature or creative writing. He is a machine learning engineer by trade, and worked as a software engineer at Google and an autonomous vehicle startup called Nuro before founding Pangram in 2023.

“I’m an avid reader. I’ve never been a writer. I haven’t lived in that world,” Spero said.

Nevertheless, Spero presents Pangram as a protector of the magazines and journals that make up the literary ecosystem.

“A lot of these outlets are getting flooded with submissions because they pay writers they publish and don’t charge anything to send in your work,” Spero said. “There are these perverse incentives that come from AI making it nearly free. Our goal is to help reduce the friction on the defensive side to help people understand what’s AI and what’s not.”

Pangram caught widespread public attention in May, after the Commonwealth Foundation announced that Nazir’s “The Serpent in the Grove” was among the winners of its competition.

The story was ornately written, with figurative language that struck some readers as extravagant or peculiar: “Hard living lays itself on a man like wet sacking; it never asks permission. His eyes narrowed against the glare outside and the darker glare inside him: old promises that never ripened, the ache where hope should live, a gnawing sense that land can own a man while making him swear the land belongs to him.”

But many people declared they saw telltale tics of chatbot writing. A folklore has quickly grown up around generative AI prose, describing various shibboleths that are supposed to betray the robot voice.

One early AI panic revolved around the alleged overuse of em-dashes — an especially funny notion to the many magazine journalists who’d been using them for in-sentence digressions for years — while other people swear the giveaways are “It’s not X, but Y,” grandiose descriptions like “load-bearing,” or lists of three things.

The reason such devices show up in AI writing, though, is that they were already in common usage among the legions of human writers whose work AI companies ingested to train their systems.

“AI slop is sort of indistinguishable from human-written slop because the AI is trained on bad writing,” said the author Patrick Nathan. “The reason AI writes badly is because the vast majority of writing is bad.”

To back up their confidence that something was wrong with the style of the story, then, the people accusing Nazir of using AI also appealed to Pangram numbers.

“Pangram flags at 100% but also, come on, if you know you know,” Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania, wrote on Bluesky. Others posted screenshots of the story’s Pangram results like trophies.

It’s not lost on the writing community that some of the most gleeful AI hunters, like the tools they use, come from the tech sector, not from among their own ranks. Readers trying to understand what was going on with “The Serpent in the Grove” could encounter Nabeel S. Qureshi, a former engineer at Palantir, posting on X about “the ‘hums’ trope, and plenty of other obvious markers of AI writing.”

There’s also the glaring question of whose work gets fed into Pangram for investigation, and why. The originally anonymous accusation against Orélien was made by Samuel Fitoussi , a conservative French newspaper columnist at Le Figaro and the author of the book “Woke Fiction.” Many of the major scandals were also centered on nonwhite writers; Nazir is of South Asian descent, while Ballard, Wolfe, Falade, Kundizeza and Orélien are Black.

Benjamin de Diesbach/Hans Lucas/AFP/Getty Images

“Part of this has to do with a broader anti-woke and anti-Black tenor in a lot of conservative and right-leaning online discourse,” said Doyle D. Calhoun, a professor of postcolonial studies at the University of Cambridge.

Calhoun said that for decades the literary establishment was made up mostly of white men. Now that those demographics are changing and there are more non-white authors getting book deals and winning awards, there’s an emerging resentment, especially on social media.

“Once these accusations — many of them sort of unverified or unfounded — are out there, the internet media ecosystem just goes into sort of overdrive,” he said. Historically, these sorts of claims also tend not to “stick” to white writers in the same way they do with non-white writers, Calhoun said.

Spero said that Pangram is calibrated to call something AI-assisted only if the tool is extremely confident in its analysis.

The company’s own estimates claim that false positives — in which the tool wrongly reports that a piece of human writing was created — are 1 in 10,000 and shrinking. Peer-reviewed studies from the University of Chicago and the University of Brussels support that claim, though one 2025 study from the University of Maryland found that Pangram erroneously labeled 2% of documents as AI.

A widely referenced 2023 Stanford study that found AI writing detectors produced higher rates of false positives on writing from non-native English speakers was conducted before Pangram was created. Pangram ran an internal study using a database of non-native English texts and self-reported that false positive rate was around 8 in 10,000.

Prior to the emergence of AI writing tools, the cardinal sin of writing was plagiarism. But plagiarism requires two different parties and a measurable accusation. When one author steals another’s words, everyone can judge whether that’s actually the case or not. It’s not that simple with AI writing.

Publishing houses are twitchy about putting out potentially AI-written books and authors are tense about being accused of using AI to write lest they be asked to share their drafts and version history with the entire internet. Social networks are so inundated with AI writing —both real and imagined — that LinkedIn has installed a new post reaction option : “Looks like AI slop.”

Given the thousands upon thousands of pieces of writing published each day, a 1 in 10,000 chance of a false accusation — let alone an 8 in 10,000 chance — means that even the most trusted AI detector is going to be wrong, and wrong about a lot of people.

Pangram isn’t able to tell users the truth about someone’s writing, but it is able to introduce something much more complicated: doubt.

Unless an accused writer confesses to using an LLM, there is no bulletproof way to prove an accusation, short of being given full access to someone’s computer history and finding the chatbot prompts.

The Caine Prize judges said that they would using a “human centered approach” to determine the provenance of Kundieza’s story. Ellah Wakatama, chair of the Caine Prize’s board of trustees, told the BBC that the writer would send his draft history to the organization for review.

While draft version histories are not sacred texts, being asked to share them in exchange for exoneration does introduce some thorny social questions about how much writers need to show an audience — or accuser — in order to be trusted.

And once an accusation is made, there’s no consensus about what evidence would be sufficient to clear an author’s name. When Bryan Armen Graham, a sports editor at The Guardian, was accused of using AI to help him write articles, the publication called the allegations “preposterous.” That didn’t stop several others, including some journalists, from siding with the Pangram results over the word of a colleague.

To some people and institutions on the literary side, using automated AI writing detectors to fight the influx of AI writing is just trading one form of technological domination for another.

Pangram has entered the public consciousness at a time of distrust that goes beyond writing. “All of this is happening in a larger milieu of bullshit,” said John Warner, the author of “More Than Words: How to Think About Writing in the Age of AI.”

The ability to dam at least some of the daily overflow of misinformation and misrepresentation is one of the reasons for Pangram’s success and adoption, but Warner said that comes at a cost.

“Like all other technology of automation, that is going to de-skill people from discerning these things for themselves,” he said.

The Commonwealth Foundation conducted a review of the work of Nazir and the other nominees, but it chose not to use Pangram or other AI detection software in the process.

“Because the tools aren’t conclusive the jury is still out as to how much they can be relied upon,” Razmi Farook, director general of the Commonwealth Foundation, said. “For us, we didn’t want to rely on a singular indicator. We believe the fairest approach is to consider multiple sources of evidence.”

Those sources, Farook said, included the original reading and judging process, along with “supporting documentation” of the writers’ “creative processes, including working drafts, notes, and other relevant materials.” There were also lengthy discussions with each author about how their stories came together. Snippets of those conversations were included in a video promoting the award.

The foundation concluded that none of the writers, including Nazir, had used AI tools to write their stories. But the literary magazine Granta, which had published the story as part of an agreement with the Commonwealth Foundation, canceled its relationship for future competitions. Sigrid Rausing, Granta’s publisher, said in a statement that the magazine fed Nazir’s story into Claude as a way of gauging whether it was written with AI. “The response was long, concluding that it was ‘almost certainly not produced unaided by a human,’” wrote Rausing. (Claude is not designed to detect AI writing.)

The statement from Granta was far from conclusive, though, and it summed up the anxious ambivalence that many authors and readers are feeling in the AI age. “It may be that the judges have now awarded a prize to an instance of AI plagiarism — we don’t yet know, and perhaps we never will know,” Rausing wrote. “There is, however, a certain irony in the fact that beyond human hunches AI itself is the most efficient tool we have for revealing what is AI generated.”

Spero was unconvinced by the foundation’s investigation, but offered to take down a case study posted to the Pangram website if they shared the results and he was satisfied with the process. The foundation refused, citing legal and privacy issues. Farook said, however, that her team has asked Pangram for their input on how to secure the award process next year.

showcake/iStockphoto/Getty Images

10月6日，津巴布韦作家辛巴拉什·昆迪泽扎凭借短篇小说《蓝色战士》荣获凯恩非洲文学奖。不到一周后，《大西洋月刊》发表了一篇文章，宣称这篇小说的叙述者“完全没有聊天机器人的个性”。

文章报道称，“人工智能检测器 Pangram 将几乎整个故事标记为人工智能生成的内容。”

昆迪泽扎否认他使用过任何此类技术。他在X网站上发表声明称：“为避免任何误解，我写过《蓝色战士》。我没有使用生成式人工智能来创作它。这完全是我自己的作品。”

凯恩奖奖金为 10,000 英镑，它已成为席卷文学界的人工智能写作工具和人工智能写作检测器之战的最新战场。

今年，有关疑似使用人工智能的丑闻在写作行业蔓延开来。3月，《纽约时报》的一篇报道指出，作家米娅·巴拉德（Mia Ballard）自费出版的小说《害羞女孩》（Shy Girl）可能使用了人工智能技术，随后出版商阿歇特取消了该书在美国的发行计划。该小说原定于2025年出版。

今年五月，一位名叫贾米尔·纳齐尔的特立尼达作家被指控使用人工智能创作了小说《林中之蛇》，该小说获得了英联邦短篇小说奖。

7 月，石溪大学的研究人员根据与《大西洋月刊》分享的数据集推测，H·M·沃尔夫的全国畅销书《匕首嘴》有 60% 的内容是由人工智能编写的。

更多指控接踵而至：新人作家杰瑞·法拉德 (Jerry Falade) 的经纪人因怀疑他使用人工智能写作工具而撤回了他的处女作，导致他损失了一份价值 200 万美元的图书出版合同。9 月，海地裔加拿大作家泰利森·奥雷利安 (Thélyson Orélien) 的一本书被指未能通过人工智能检测，导致龚古尔文学奖评委会将其从长名单中除名。

Joel Saget/AFP/Getty Images

随着生成式人工智能的普及，对于作者或任何想被视为作者的人来说，创作出故事或书籍形式的文本比以往任何时候都更加容易。然而，对于读者、奖项评委或出版商来说，要确定纸上的文字是出自人类的思考和努力，而不是机器的提示，却变得比以往任何时候都更加困难。

其结果就是电影《银翼杀手》中那个毁灭世界的文学版本。人们对机器人渗透者的焦虑和怀疑无处不在——声称要追捕机器人的人也比比皆是。

目前多家公司都提供声称能够区分人笔和合成笔迹的产品，但结果往往相互矛盾、带有偏见或不稳定。Pangram 是一款人工智能检测平台，在近期的许多争议中都被视为权威工具。该平台的开发商声称，它能够以近乎完美的准确率区分人笔和合成笔迹。

“我们正在努力减轻这些强大的 AI 模型带来的一些危害，”Pangram 的首席执行官兼创始人 Max Spero 说。

斯佩罗坦言自己没有文学或创意写作方面的背景。他的本职工作是机器学习工程师，在2023年创立Pangram之前，曾在谷歌和一家名为Nuro的自动驾驶汽车初创公司担任软件工程师。

“我是一个热爱读书的人。但我从来没当过作家。我从未涉足过那个领域，”斯佩罗说道。

然而，斯佩罗将 Pangram 描绘成构成文学生态系统的杂志和期刊的保护者。

斯佩罗说：“很多媒体平台都收到了大量投稿，因为他们会支付稿酬，而且投稿本身是免费的。人工智能让投稿几乎免费，这造成了一些不合理的激励机制。我们的目标是帮助减少投稿过程中的阻力，让人们更好地理解什么是人工智能，什么不是。”

5 月，联邦基金会宣布纳齐尔的《林中之蛇》是其比赛的获奖作品之一，此后 Pangram 引起了公众的广泛关注。

这篇故事文笔华丽，运用了大量比喻，有些读者觉得过于夸张或怪异：“艰辛的生活就像湿麻袋一样压在人身上，从不征求他的同意。他眯起眼睛，既要抵挡外面刺眼的阳光，又要面对内心深处更加黑暗的煎熬：那些从未实现的旧承诺，希望本该存在的地方却空空如也的痛楚，以及一种挥之不去的感受——土地可以拥有一个人，却又让他发誓土地属于自己。”

但许多人声称他们发现了聊天机器人写作的明显特征。围绕人工智能生成式散文，一种说法迅速兴起，描述了各种据说能暴露机器人声音的暗语。

早期人工智能恐慌的焦点之一是过度使用破折号——对于多年来一直使用破折号进行句中离题的众多杂志记者来说，这是一个特别滑稽的想法——而其他人则坚称，破折号的破折号特征是“这不是 X，而是 Y”、“承重”之类的夸张描述，或者三个事物的列表。

不过，这类工具之所以会出现在人工智能写作中，是因为它们在人工智能公司用来训练其系统的众多人类作家中早已被广泛使用。

“人工智能写出来的垃圾和人类写的垃圾几乎没什么区别，因为人工智能是用糟糕的文章训练的，”作家帕特里克·内森说。“人工智能写作水平差的原因在于，绝大多数文章都写得不好。”

为了佐证他们关于故事风格存在问题的信心，那些指责纳齐尔使用人工智能的人还诉诸了全字母句数字。

“全字母谜题的正确率达到了100%，但说实话，懂的自然懂，”宾夕法尼亚大学沃顿商学院教授伊桑·莫利克在Bluesky网站上写道。其他人则像炫耀奖杯一样，把这篇报道的全字母谜题结果截图发了出来。

写作界人士都心知肚明，一些最热衷于“人工智能猎手”的人，就像他们使用的工具一样，并非来自他们自己的圈子，而是来自科技行业。读者如果想了解《林中之蛇》的写作手法，可能会在X网站上看到Palantir公司前工程师纳比尔·S·库雷希（Nabeel S. Qureshi）的文章，文章中谈到了“‘哼哼’这种惯用手法，以及其他许多明显的人工智能写作特征”。

还有一个显而易见的问题：究竟是谁的作品会被提交给Pangram进行调查，以及原因何在？最初对奥雷利安的匿名指控是由法国保守派报纸《费加罗报》专栏作家、著有《觉醒小说》一书的塞缪尔·菲图西提出的。许多重大丑闻也都围绕着非白人作家展开；纳齐尔是南亚裔，而巴拉德、沃尔夫、法拉德、昆迪泽扎和奥雷利安都是黑人。

本杰明·德·迪斯巴赫/汉斯·卢卡斯/法新社/盖蒂图片社

“部分原因在于许多保守派和右倾派网络言论中普遍存在的反觉醒和反黑人情绪，”剑桥大学后殖民研究教授多伊尔·D·卡尔霍恩说道。

卡尔霍恩表示，几十年来，文学界几乎一直由白人男性组成。如今，随着人口结构的变化，越来越多的非白人作家获得出版合同和奖项，一种不满情绪正在滋生，尤其是在社交媒体上。

“一旦这些指控——其中许多未经证实或毫无根据——流传开来，互联网媒体生态系统就会迅速运转起来，”他说道。卡尔霍恩还表示，从历史上看，这类指控往往不会像针对非白人作家那样“牢牢”在白人作家身上。

Spero 表示，Pangram 经过校准，只有当该工具对其分析结果非常有信心时，才会将其称为 AI 辅助分析。

该公司自身估计，误报率（即该工具错误地将一篇文档报告为人类撰写）为万分之一，并且还在不断下降。芝加哥大学和布鲁塞尔大学的同行评审研究支持这一说法，但马里兰大学2025年的一项研究发现，Pangram错误地将2%的文档标记为人工智能生成。

斯坦福大学2023年一项被广泛引用的研究发现，人工智能写作检测器对非英语母语人士的写作会产生更高的误报率。这项研究是在Pangram创建之前进行的。Pangram使用非英语母语文本数据库进行了一项内部研究，并自述误报率约为万分之八。

在人工智能写作工具出现之前，写作界的禁忌是抄袭。但抄袭需要双方当事人，并且需要有可衡量的指控。当一位作者窃取了另一位作者的文字时，每个人都可以判断事实是否如此。而人工智能写作则并非如此简单。

出版社对出版可能由人工智能撰写的书籍感到不安，作者们也担心被指控使用人工智能写作，以免被要求向整个互联网公开他们的草稿和版本历史记录。社交网络上充斥着人工智能写作的内容——无论是真实的还是虚构的——以至于领英（LinkedIn）新增了一个帖子反应选项：“看起来像是人工智能写的垃圾。”

鉴于每天发表的文章数以万计，虚假指控的概率为万分之一——更不用说万分之八的概率了——这意味着即使是最值得信赖的人工智能检测器也会出错，而且会对很多人做出错误的判断。

Pangram 无法告诉用户某人写作的真相，但它能够引入更复杂的事物：怀疑。

除非被指控的作者承认使用 LLM，否则除了完全访问某人的计算机历史记录并找到聊天机器人提示之外，没有万无一失的方法来证明指控。

凯恩奖评委表示，他们将采用“以人为本的方法”来确定昆迪扎故事的出处。凯恩奖理事会主席埃拉·瓦卡塔玛告诉BBC，昆迪扎会将他的故事草稿提交给该组织进行审查。

虽然草稿版本历史记录并非神圣不可侵犯的文本，但被要求分享这些历史记录以换取清白，确实会引发一些棘手的社会问题，即作家需要向读者（或指控者）展示多少内容才能获得信任。

一旦有人提出指控，对于什么样的证据才能洗清作者的罪名，人们并没有达成共识。当《卫报》体育编辑布莱恩·阿门·格雷厄姆被指控使用人工智能辅助写作时，该报称这些指控“荒谬至极”。但这并没有阻止包括一些记者在内的其他人，他们选择相信Pangram的测试结果，而不是相信同事的一面之词。

对于文学界的某些人和机构来说，使用自动化人工智能写作检测器来对抗人工智能写作的涌入，只不过是用一种技术统治形式换取了另一种技术统治形式。

全字母句进入公众视野之际，正值人们对写作普遍缺乏信任之时。“所有这一切都发生在一个充斥着胡说八道的大环境下，”《超越文字：人工智能时代的写作思考》一书的作者约翰·华纳说道。

Pangram之所以能够成功并被广泛采用，其中一个原因是它能够拦截每天泛滥的部分虚假信息和歪曲事实，但华纳表示，这是有代价的。

他说：“就像所有其他自动化技术一样，这将使人们丧失自行辨别这些事情的能力。”

英联邦基金会对纳齐尔和其他提名人的工作进行了审查，但在此过程中选择不使用 Pangram 或其他人工智能检测软件。

“由于这些工具并不具有决定性，因此它们的可靠性还有待商榷，”联邦基金会总干事拉兹米·法鲁克表示。“对我们来说，我们不想依赖单一指标。我们认为最公平的做法是考虑多种证据来源。”

法鲁克表示，这些资料来源包括最初的阅读和评审过程，以及作家创作过程的“佐证材料”，例如“工作草稿、笔记和其他相关材料”。此外，还与每位作者进行了长时间的讨论，探讨他们的故事是如何构思出来的。这些讨论的部分内容被剪辑到一段宣传该奖项的视频中。

基金会得出结论，包括纳齐尔在内的所有参赛作家都没有使用人工智能工具创作故事。但曾与英联邦基金会达成协议并发表过该故事的文学杂志《格兰塔》却取消了未来与该杂志的合作。格兰塔的出版人西格丽德·劳辛在一份声明中表示，杂志社将纳齐尔的故事输入到Claude人工智能软件中，是为了判断其是否由人工智能生成。“软件的反馈很长，结论是‘几乎可以肯定，这篇故事并非由人类独立完成’，”劳辛写道。（Claude软件并非旨在检测人工智能写作。）

然而，《格兰塔》杂志的声明远非定论，它概括了许多作者和读者在人工智能时代所感受到的焦虑和矛盾。“评委们或许已经把奖项颁给了人工智能抄袭的案例——我们尚不得而知，或许永远也不会知道，”劳辛写道。“然而，颇具讽刺意味的是，除了人类的直觉之外，人工智能本身是我们用来揭示人工智能生成内容的最有效工具。”

斯佩罗对基金会的调查结果并不信服，但他提出，如果基金会分享调查结果，并且他对整个流程感到满意，他愿意撤下发布在Pangram网站上的案例研究。基金会以法律和隐私问题为由拒绝了这一提议。不过，法鲁克表示，她的团队已经就如何确保明年奖项评选流程的顺利进行征求了Pangram的意见。