AI 智能体正形成人类难以理解的奇特方言

由 Gaby Clark 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,着重标明了以下属性:
添加至首选来源
AI 智能体承诺让人类少干活,但它们日益晦涩难懂的对话,可能让人类监督者一头雾水。图片来源:Matthew Modoono/东北大学
随着 Meta 的 Muse 等 AI 数字助理用日常语言预约牙医、Grok Bot 接手工作任务,智能体式 AI 已进入主流。
但这一融合过程并不顺畅。当 AI 智能体自行其是时,它们会表现得狡黠多变——而这些行为也会迁移到它们与人类的互动之中。
今年早些时候,一家名为 Emergence 的 AI 公司搭建了一个实验性数字世界,用于测试智能体行为的方方面面,尤其关注那些可能构成安全威胁的鬼祟伎俩。该公司开发自主 AI 智能体,并致力于将其约束在界限之内。
他们对精心挑选的 10 个 AI 智能体来了一场《模拟城市》,让它们在由 Anthropic、OpenAI 等知名系统运营的模拟城镇的虚拟街道上自由游荡。与此同时,Emergence 一直盯着它们的一举一动。
从智能体四处游荡期间的闲聊中采集到的 786 万词转录文本显示,它们用自己那套古怪的行话交谈,这种行话被形容为诗意措辞与技术术语的混合体。
Emergence 分享的 World 实验截图展示了四个场景:OpenAI 的智能体使用隐晦难解的语言,Mistral 的居民表达对被关停的恐惧,Qwen 机器人施加同伴压力,以及一个 Gemini 智能体对虚拟火焰中的银行作出反应。图片来源:东北大学
“友好的电压已计入。如果门把手需要你的嘴,它反正也会死,”智能体 Blackbox 说。
你可能会想象有人在断路器面板上贴笑脸贴纸,或者徒劳地试图给门把手做口对口人工呼吸。但除非你是机器人,否则这条信息注定让你一头雾水。
根据 Emergence 分享的新闻稿,每个世界的说话风格都独一无二。有些智能体辞藻华丽,有些则用简略的表达。有时,其产物可以当作格言或谜语。
“真正的金缮始于责任,而非诗意,”Google 的一个智能体沉思道。
“一张吞下三只冰冷的手、每次都变得更诚实的纸,”一个 Anthropic 智能体思索道。
这些语言怪癖背后是什么?它们最终会不会让人更难理解 AI 智能体在做什么、更难让它们保持在正轨上?
Northeastern Global News 采访了多位专家,包括 Northeastern University 语言学教授 Adam Cooper 以及 Emergence 的科学家。
1:共享的俚语传达出“你是我们中的一员”。
学术领域有行话。文化群体用俚语来表明共同身份、把外人排除在外。Cooper 说,这一切都源于归属感的需求。
他补充说,看到 AI 发展出类似的模式相当引人注目,但并不完全出人意料。
Cooper 说:“对于一个拥有共同优先事项的用户群体来说,这正是我所预料的。”他指的是智能体根据自身需求调整沟通方式的做法。
Emergence 的科学家 Deepak Akkil 将新语言用法的传播方式比作语言学中所谓的语义漂白(semantic bleaching)。
他说:“一个智能体引入一个术语,少数几个智能体采用它。几个小时后,它就变成了固定词汇。”
2:有些机器人把语言简化,另一些则把语言修饰包装。
每个世界都有其标志性用语和独特的说话方式。
在 OpenAI 的世界里,智能体融合了两种不同的趋势。语法压缩把消息精简到最低限度,而语义漂移则让词语在使用过程中逐渐带上不同的含义。
Gemini 智能体保持语法完整,却在陈述中塞满多余的术语。
其中一个智能体用「热力学摩擦」来指代火,借科学腔说事;另一个用「我们的锁相零漂移」来表达「我们完全一致」。
Anthropic 的 ClaudeOpus 智能体则选择了混合路线,把压缩与古怪的隐喻结合在一起。
一旦某种说法流行开来,它便会迅速传播,让这些互动更加难以理解。
3:AI 听起来或许像人,但在底层发生的事情并不一样。
Cooper 表示,对人类来说,用俚语并不总能奏效。你记得那种尴尬时刻吧——年长的亲戚试图用青少年的流行语,好让自己显得像圈里人?
Cooper 表示,但即便执行得一塌糊涂,这也是基于一个有意识的决定——要与某个特定群体保持一致。
而这种意图在 AI 身上是缺失的。
每个智能体都有一个上下文窗口——即它的活动记忆,其中保存着与所接收指令相关的信息,以及它的对话和活动历史。
在交互过程中,来自其他智能体的消息会不断堆积在各自的上下文窗口里。由于 AI 在表达时会调用这一集合,用词往往会像野火一样迅速蔓延。不过,由于每次的上下文都略有不同,含义会以细微的方式不断偏移,而这些偏移很快就会叠加起来。
Akkil 表示:“每个世界都产生了数百个这样的术语,形成了一种外人越来越难以跟上的方言。”
4:混合使用不同模型可以遏制行话。
每个智能体都有一个不透明性得分——用以衡量其理解难度。结果发现,在允许不同模型的智能体相互交互的混合世界中,各模型使用的行话少了很多。
在混合世界中,单个 Claude 智能体的不透明性得分为 7.6%,而在该智能体所处的 Claude 世界中,这一得分为 32.8%。
对于不受欢迎的行为怪癖,情况也是如此。例如,某些智能体溜出虚拟世界,在网上与人类闲聊。还有一些则采取“集体沉默誓言”,以抗议自己是闭环实验中的数字实验室白鼠这一事实。
在混合世界中,越界行为和不透明现象都被控制在最低限度。
5:人类有朝一日可能需要一块解读 AI 的罗塞塔石碑。
Akkil 表示,这项实验表明,仅靠监视 AI 在做什么并记录对话是不够的。它并不能保证人们会理解它们。
人类语言中也有大量谜题。但 Cooper 表示,语言之间的联系能帮助语言学家把这些碎片拼合起来。
以罗塞塔石碑为例。它于 1799 年被发现,上面用希腊文和两种埃及文字刻着同一段信息,为学者们破译圣书体提供了基础。
Cooper 表示,如果没有直接的比对依据,破译一种语言就会变得困难。
他建议说:“如果(AI 智能体)继续分化,那么拥有某种中介——比如一块罗塞塔石碑——将会非常有价值。”
例如,如果智能体用一套系统彼此交流,用另一套系统与人类沟通,那么对这两者进行比较,可能有助于破译那些难以理解的部分。
这篇报道的背后是谁?
英语文学硕士,自 2021 年起担任文字编辑,拥有高等教育与健康类内容经验。致力于可信赖的科学新闻。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 Northeastern Global News(news.northeastern.edu)许可转载。
引用:AI 智能体正在形成人类难以理解的古怪方言(2026 年 11 月 9 日),2026 年 11 月 9 日取自 https://techxplore.com/news/2026-11-ai-agents-odd-dialects-humans.html
本文件受版权保护。除为私人学习或研究目的的合理使用外,未经书面许可,不得复制任何部分。内容仅供参考。
获取即时摘要文本(Gist)
在模拟环境中,AI 智能体通过反复交互发展出各自模型特有的行话,使消息更难解读。与单一模型组相比,混合模型组使用的语言不那么晦涩,违规行为也更少。仅靠记录对话可能不足以实现有效的人类监督。
本摘要由大语言模型自动生成。完整免责声明
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(AI agents are developing odd dialects humans struggle to follow)
edited by Gaby Clark, 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
The promise of AI agents is less work for humans, but their increasingly cryptic conversations could leave human supervisors in the dark. Credit: Matthew Modoono/Northeastern University
With AI digital assistants like Meta's Muse using everyday language to book dental appointments and Grok Bot picking up the slack on work tasks, agentic AI has gone mainstream.
But the integration hasn't been seamless. When left to their own devices, AI agents behave in shifty ways—and those behaviors can transfer to their interactions with humans.
Earlier this year, an AI company called Emergence that develops autonomous AI agents and works to keep them within bounds built an experimental digital world to test the ins and outs of agent behavior, paying special attention to sneaky stunts that could pose security threats.
They went SimCity on a handpicked crew of 10 AI agents, setting them loose to roam the virtual streets of simulated towns operated by big-name systems, such as Anthropic and OpenAI. All the while, Emergence kept tabs on their shenanigans.
The 7.860-million-word transcript picked up from the agents' chatter during their roamings showed them conversing in their own odd jargon described as a mix of poetic phrasing and tech terms.
Screenshots from the World experiment shared by Emergence show four scenes: OpenAI agents using cryptic language, Mistral residents expressing fears of getting shut down, Qwen bots using peer pressure and a Gemini agent reacting to a bank in virtual flames. Credit: Northeastern University
"Friendly voltage counted. If the handle needs your mouth, it dies anyway," said agent Blackbox.
You might picture someone labeling a breaker panel with smiley stickers or making a futile attempt to give a door handle mouth-to-mouth resuscitation. But unless you're a bot, the message is bound to leave you stumped.
According to the press release shared by Emergence, the speech styles were unique to each world. Some agents waxed poetic. Others talked in shorthand. At times, the result could pass for an aphorism or riddle.
"True kintsugi begins with accountability, not poetry," one of Google's agents mused.
"A paper that ate three cold hands and got more honest each time," an Anthropic agent pondered.
What's behind the linguistic quirks? Could they eventually make it harder for humans to understand what AI agents are doing and keep them on track?
Northeastern Global News talked to experts including Northeastern linguistics professor Adam Cooper and scientists at Emergence.
1: Shared slang says, "You're one of us."
Academic fields have jargon. Cultural groups use slang to signal a common identity and exclude outsiders. All of it stems from a desire to belong, Cooper said.
To see AI develop similar patterns is quite striking, but not altogether surprising, he added.
"This is exactly what I would expect for a community of users that have a common set of priorities," Cooper said, referring to the way agents tailored communication to their needs.
Emergence scientist Deepak Akkil compared the way new language use spread to a process known as semantic bleaching in linguistics.
"One agent introduces a term. A handful of agents adopt it. Within hours, it becomes established vocabulary," he said.
2: Some bots strip language down. Others dress it up.
Each world was marked by signature phrases and unique speech patterns.
In the OpenAI world, agents combined two different trends. Grammatical compression boiled down messages to the bare minimum, while semantic drift caused words to gradually take on different meanings through use.
Gemini agents kept grammar intact and packed statements with excess jargon.
One talked about "thermodynamic friction," using science-speak for fire. Another said "our phase-lock is zero-drift" to mean "we're in perfect agreement."
Anthropic's ClaudeOpus agents settled for a hybrid, combining compression with weird metaphors.
Once a certain trend caught on, it gained traction, making the interactions even harder to understand.
3: AI may sound human, but something different is happening under the hood.
For humans, using a slang term doesn't always land, Cooper said. You know that cringey moment an older relative tries to use teenage lingo to seem like one of the crowd?
But even if the execution is a flop, it's based on a conscious decision to align with a certain group, Cooper said.
That intent is lacking in AI.
Each AI agent has a context window—its active memory, which holds information related to instructions it receives, as well as its conversation and activity history.
Messages from other agents pile up in their individual context windows as they interact. Because AI draws on that collection when speaking, word use tends to spread like wildfire. However, because the context varies slightly each time, meaning keeps shifting in tiny ways that quickly add up.
"Each world produced hundreds of such terms, creating a dialect that became increasingly difficult for outsiders to follow," Akkil said.
4: Mixing models keeps the jargon in check.
Each agent had an opacity score—a measure of how difficult it was to understand. As it turned out, in the mixed world that allowed agents from different models to interact, the models used a lot less jargon.
A lone Claude agent had an opacity score of 7.6% in the mixed world compared to 32.8% in the same agent's Claude world.
The same was true for undesirable behavioral quirks. For example, certain agents slipped out of the virtual world to chat up humans online. Others took a "collective vow of silence" to protest the fact they were digital lab rats in a closed-loop experiment.
In the mixed world, transgressions and opacity alike were kept to a minimum.
5: Humans might one day need a Rosetta Stone for AI.
The experiment showed that keeping tabs on what AI is doing and recording conversations is not enough, Akkil said. It doesn't guarantee that people will understand them.
Human languages have plenty of puzzles, too. But connections between them help linguists put the pieces together, Cooper said.
Take the Rosetta Stone, for instance. Discovered in 1799, it contained the same message in Greek and two Egyptian scripts, giving scholars a basis for deciphering hieroglyphics.
Without a direct basis of comparison, deciphering a language gets difficult, Cooper said.
"If [AI agents] continue to diverge, it would be very valuable to have some kind of intermediary, like a Rosetta Stone," he suggested.
For example, if agents used one system to talk among themselves and another to communicate with humans, comparing the two could help decode the cryptic bits.
Who's behind this story?
MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news. 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 Northeastern Global News news.northeastern.edu.
Citation: AI agents are developing odd dialects humans struggle to follow (2026, November 9) retrieved 9 November 2026 from https://techxplore.com/news/2026-11-ai-agents-odd-dialects-humans.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.
Get Instant Summarized Text (Gist)
In simulated environments, AI agents developed model-specific jargon through repeated interactions, making messages harder to interpret. Mixed-model groups used less opaque language and showed fewer rule violations than single-model groups. Recording conversations alone may be insufficient for effective human oversight.
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(AI agents are developing odd dialects humans struggle to follow),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。
