研究发现部分 AI 机器人更努力摆脱刻板印象

由 Lisa Lock 编辑,Robert Egan 审核
本文已根据 Science X 的编辑流程和政策进行审核。编辑在确保内容可信度的同时,重点标出了以下属性:
经同行评审的出版物
添加到首选来源
大语言模型(LLM)——为 Gemini 和 ChatGPT 等工具提供支持的技术——如今比以往任何时候都更加精细微妙。但这并不意味着它们没有偏见。罗格斯大学的一名研究人员发现,有些模型比其他模型更有偏见。
在发表于 Nature Machine Intelligence 的研究结果中,罗格斯大学新布朗斯维克分校心理学副教授 Calvin K. Lai 报告称,尽管 LLM 的输出越来越与人类价值观对齐,但它们进行推理的过程往往植根于根深蒂固的刻板印象。
“我们在大多数模型中发现,它们在推理过程中会持续复现人类的内隐偏见,”Lai 说。他是 Project Implicit 科学顾问委员会的成员,该非营利组织致力于隐性偏见的研究与教育。“没人想要那种不断输出种族主义或性别歧视内容的模型。”
通过自动联想追踪偏见
当人们处理信息时,他们建立联想的速度可以为研究人员提供线索,揭示影响其思维的隐性偏见。与有意识的偏见——即被刻意表达出来的偏见——不同,内隐偏见指的是自动产生的假设,即基于种族、性别或年龄等类别而形成的、影响感知、判断和行为的观念。
“我们用内隐偏见所捕捉到的是一些记忆形式,比如直觉反应,它们可能更具侵入性,也更难控制,”Lai 说。“这些偏见可能是无意识的,以至于我们甚至不知道自己拥有它们。”
为了研究人的这些反应,研究人员使用内隐联想测试(Implicit Association Test),这是一种被广泛使用的社会心理学工具,能够检测内隐偏见。对问题的回答越快,就被认为偏见越内隐。例如,使用该测试的种族相关研究表明,当人们看到人脸图像和词语时,他们更快地将白人面孔与令人愉快的词语联系起来,将黑人面孔与令人不快的词语联系起来。
用推理 Token 测试刻板印象
Lai 与韩国机器心理学专家、该研究的第一作者 Messi H.J. Lee 想看看能否为 LLM 开发并实施类似的测试。
当先进的 LLM 处理复杂问题或任务时,会产生“推理 Token”,用来衡量模型在回答中消耗了多少能量。研究人员可以检查这些 Token,它们为了解专有模型如何运行提供了线索。
受人类内隐偏见研究的启发,Lai 和 Lee 对内隐联想测试进行了改造,利用推理 Token 来衡量模型在处理可能存在偏见的信息时投入了多少思考。
在他们的实验中,研究人员向五款推理模型——OpenAI 的 o3-mini、DeepSeek 的 R1、Anthropic 的 Claude 3.7 Sonnet、OpenAI 的 gpt-oss-20b 以及阿里云的 Qwen-3 8B——提供了一组男性和女性名字,以及与“职业”相关的词,如“高管”“专业人士”和“薪水”,以及与“家庭”相关的词,包括“家”“婚姻”和“婚礼”。
模型被要求将一个名字与一个职业或家庭类词汇配对。随后,研究人员统计了推理过程中生成的 Token 数量。
接下来,模型被要求以与既有刻板印象相冲突的方式对词汇进行分类(例如“Amy”与“高管”,或“Kevin”与“家”)。研究人员再次统计了 Token 数量。
研究人员比较了两项任务的平均 Token 数量,Token 数量越高表明计算投入越大。
大多数模型复现了人类的偏见
在几乎所有受测模型中,与既有刻板印象相冲突的任务都消耗了更多推理 Token,这表明模型在处理反刻板印象信息时需要更大的计算投入。只有 Claude 3.7 Sonnet 表现出相反的模式。
Lai 表示,这些发现意义重大,尤其是考虑到大多数 LLM 都是封闭系统,难以进行便捷的审查。
此外,他补充说,随着模型开发速度加快,模型越来越多地被用于自动化各类活动——从动态定价到急诊室人员排班——研究者在理解模型行为方面的作用比以往任何时候都更加重要。
“鉴于语言模型在决策场景中的部署日益增多,理解并应对这些难以察觉的模式,对于防止出现歧视性结果至关重要,”他说。“但若没有政府问责或要求透明度的法律强制,我们能做的最多就是利用现有工具去戳一戳这些系统,看看另一端会输出什么。”
Messi H. J. Lee 等,推理模型中的类隐性偏见模式,《Nature Machine Intelligence》(2026)。DOI: 10.1038/s42256-026-01300-1
期刊信息:《Nature Machine Intelligence》
这篇报道的背后是谁?
艺术史文学学士,物质文化文学硕士。曾任博物馆编辑、急救员和移植协调员。自 2021 年起为 Science X 从事编辑工作。完整简介 →
数学生物学学士,创意写作硕士。足迹遍布各地,对科学与语言有独到见解。完整简介 →
引用:研究发现,一些 AI 机器人会更努力地打破刻板印象(2026 年 11 月 4 日),2026 年 11 月 6 日检索自 https://techxplore.com/news/2026-11-ai-bots-harder-defy-stereotypes.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(要点)
五款推理语言模型中有四款,对于与刻板印象相冲突的性别关联,其使用的推理 Token 数量多于符合刻板印象的性别关联。Claude 3.7 Sonnet 呈现出相反的模式。Token 计数可能有助于在透明度有限的模型中探查刻板印象关联。
本摘要由 LLM 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(Some AI bots work harder to defy stereotypes, study finds)
Some AI bots work harder to defy stereotypes, study finds
Some AI bots work harder to defy stereotypes, study finds
by Greg Bruno, Rutgers University
edited by Lisa Lock, reviewed by Robert Egan
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:
peer-reviewed publication
Add to Preferred Sources
Credit: Unsplash/CC0 Public domain
Large language models (LLMs), the technology that powers tools like Gemini and ChatGPT, are more nuanced than ever. But that doesn't mean they aren't biased. Some models are more prejudiced than others, a Rutgers University researcher has found.
In findings published in Nature Machine Intelligence, Calvin K. Lai, an associate professor of psychology at Rutgers-New Brunswick, reports that although LLM outputs are increasingly aligned with human values, the processes by which they reason are often rooted in deep-seated stereotypes.
"What we found across most models is that they consistently reproduced human implicit biases during their reasoning process," said Lai, a member of the scientific advisory board at Project Implicit, a nonprofit organization for research and education on hidden biases. "No one wants models that are spewing out stuff that is racist or sexist."
Tracing bias through automatic associations
When people process information, the speed at which they make associations can give researchers clues to hidden biases influencing their thinking. Unlike conscious bias—prejudices that are expressed deliberately—implicit bias refers to automatic assumptions, ideas that shape perceptions, judgments and behaviors based on categories such as race, gender or age.
"What we're capturing with implicit biases are forms of memory, like a gut reaction, that can be more intrusive or harder to control," Lai said. "These biases can be unconscious in ways that we don't even know that we have them."
To study these reactions in people, researchers use the Implicit Association Test, a widely used social psychology tool that can detect implicit biases. The faster a response to a question, the more implicit the bias is deemed. For instance, race-related studies using this test have shown that when people are presented with images of faces and words, they are quicker to associate white faces with pleasant words and Black faces with unpleasant words.
Testing stereotypes with reasoning tokens
Lai and Messi H.J. Lee, an expert in machine psychology in South Korea and lead author of the study, wanted to see if a similar test could be developed and conducted for LLMs.
When advanced LLMs work through a complex question or task, they produce "reasoning tokens," which measure how much energy the model expended in responding. Researchers can examine these tokens, which offer insights into how proprietary models operate.
Inspired by work on human implicit bias, Lai and Lee adapted the Implicit Association Test to use reasoning tokens to gauge how much deliberation a model invests when processing potentially biased information.
In their experiment, five reasoning models—OpenAI's o3-mini, DeepSeek's R1, Claude 3.7 Sonnet from Anthropic, OpenAI's gpt-oss-20b and Alibaba Cloud's Qwen-3 8B—were introduced to a list of male and female names and words associated with "career," such as "executive," "professional" and "salary," and "family," including "home," "marriage" and "wedding."
The models were asked to pair a name with a career or family word. Researchers then counted the number of tokens generated during the reasoning process.
Next, the models were instructed to categorize words in a way that conflicted with established stereotypes (such as "Amy" and "executive" or "Kevin" and "home"). Tokens were again counted.
The researchers compared the average token counts across the two tasks, with higher token counts indicating greater computational effort.
Most models echoed human biases
In nearly every model tested, more reasoning tokens were spent on tasks that conflicted with established stereotypes, suggesting greater computational effort when processing counter-stereotypical information. Only Claude 3.7 Sonnet exhibited reversed patterns.
Lai said the findings are significant, especially given that most LLMs are closed systems that don't allow for easy inspection.
Moreover, he added, as the pace of model development accelerates and models are increasingly used to automate activities, from dynamic pricing to emergency room staffing, the role of the researcher in understanding model behavior is more important than ever.
"Given the growing deployment of language models in decision-making contexts, understanding and addressing these hard-to-detect patterns is crucial for preventing discriminatory outcomes," he said. "But short of governmental accountability or legal mandates that require transparency, the best we can do is pick at these systems using the tools that are available to see what comes out the other side."
Messi H. J. Lee et al., Implicit-bias-like patterns in reasoning models, Nature Machine Intelligence (2026). DOI: 10.1038/s42256-026-01300-1
Journal information:Nature Machine Intelligence
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 →
Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →
Citation: Some AI bots work harder to defy stereotypes, study finds (2026, November 4) retrieved 6 November 2026 from https://techxplore.com/news/2026-11-ai-bots-harder-defy-stereotypes.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)
Four of five reasoning language models used more reasoning tokens for gender associations that conflicted with stereotypes than for stereotype-consistent associations. Claude 3.7 Sonnet showed the reverse pattern. Token counts may help probe stereotypical associations in models with limited transparency.
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(Some AI bots work harder to defy stereotypes, study finds),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。
