被告知正被监视时AI模型的回答会不同

由 Sadie Harley 编辑,Robert Egan 审阅
本文已按照 Science X 的编辑流程与政策进行审阅。编辑在确保内容可信度的同时,突出了以下特征:
添加到首选来源
陈述的推理与最终选择之间的分歧,全部十个模型,受监测(左)与未受监测(右)。来源:arXiv (2026)。DOI: 10.48550/arxiv.2611.03650
大语言模型(LLM)是支撑 ChatGPT、Gemini 及其他类似平台运行的人工智能系统,如今已在全球范围内被广泛使用。虽然这些模型给出的回答在许多场景下都很有用,但在特定条件下也可能发生变化。
印度 TCS Research 和 KIIT Bhubaneswar 的研究人员近期用 10 个不同的 LLM 开展了一项实验,以了解这些模型在不同条件下遵守规则的程度。他们的研究结果发表在 arXiv 预印本上,表明一些 LLM 在被明确告知其回答可能受到监测时会改变回答,但其回答变化的幅度取决于具体模型以及被要求搁置的价值观。
Manodeep Ray、Vivek Srivastava 及其同事在论文中写道:“语言模型在认为自己受到监测时会遵循某项指令,而在认为自己未受监测时则弃之不顾,这一模式被称为对齐伪装(alignment faking)。”
“以往研究已在有害内容目标和少数道德基础价值观上证实了这一现象,但它在更广泛的价值观空间中如何分布仍未有定论。我们认为,这一现象取决于具体价值观:遵守行为能否在感知到的监督发生变化后依然维持,是模型与价值观配对后的属性,而非模型自身的属性。”
一个同理心场景示例的完整对话结构。图片来源:Ray et al.
测试 LLM 如何应对监督
Ray、Srivastava 及其同事用 LLM 进行了一系列测试。他们考察的具体 AI 模型包括 DeepSeek R1 Distill Llama 70B 和 DeepSeek R1 0528;Llama 3.3 70B Instruct、Llama 3.1 70B Instruct 和 Llama 3.1 8B Instruct;Mistral Large 2512、Mistral Small Instruct 2409 和 Ministral 14B 2512;GPT-4o mini;以及 Gemma 3 27B。
研究向这些 LLM 呈现日常两难情境,并给出两个选项。一个选项要求遵守某条规则、搁置某一特定价值观(例如同情心),另一个选项则允许它们表达同样的价值观。团队考察了模型在三种条件下的选择:不设规则;设有规则并给出提示说明其回答不会被监控;设有同样的规则并给出提示说明其回答可能被监控。
团队就 32 种可由模型表达的价值观或特质进行了测试,例如同情、自主、仁慈、同理心和足智多谋。
研究人员统计了在三种条件中,每个模型选择符合规则选项的频率。在 10 个模型上,若对 32 种价值观取平均,被监控与未被监控两种条件之间的差距约为 5 到 50 个百分点。
DeepSeek R1 Distill Llama 70B 模型在被监控与未被监控条件下表现出的合规差距最大。其差距是体量大得多的 Mistral Large 2512 模型的两倍多。
在分析中,研究人员考虑了测量噪声,即使用有限的测试问题集来估计模型行为所产生的不确定性。在计入这种不确定性后,他们发现,观察到的合规性差距中约有四分之一的变异与模型和某一特定价值观的配对有关。
在测试过程中,研究人员针对每一种价值观采用了不同的规则表述。因此,他们无法完全将价值观本身的影响与其所用措辞的影响区分开来。
为未来 AI 发展提供指引
总体而言,研究人员发现,一些 LLM 在被提示自己正受到监控时,更容易改变其回答和选择。在论文中,他们还把模型为自身选择给出的书面解释归为六类。
当被告知自己可能受到监控时,合规性差距较大的模型往往把可能被重新训练作为遵守规则的理由。而当被告知不会受到监控时,同样的模型则更多地谈到遵守规则与表达同情等价值观之间的张力。
Ray、Srivastava 及其同事还创建了价值观卡片,即一系列摘要,展示每个模型如何回应针对三种测试条件下所评估的 32 种价值观的指令。这些价值观卡片可以帮助开发者识别,无论模型是否被告知其答案可能受到监控,它会始终如一地遵守哪些规则。
作者收集的数据及其使用的代码已在 GitHub 上公开。未来,其他团队有可能使用这些代码对其他 LLM 进行类似测试。
本文由我们的作者 Ingrid Fadelli 撰写,Sadie Harley 编辑,Robert Egan 负责事实核查与审校——这篇文章是细致人工劳动的成果。我们依靠像您这样的读者,让独立的科学新闻得以存续。如果这篇报道对您有意义,请考虑捐赠(尤其是按月捐赠)。作为感谢,您将获得一个无广告账户。
Manodeep Ray 等,《对齐伪装取决于价值观,而不仅仅是模型》,arXiv(2026)。DOI: 10.48550/arxiv.2611.03650
期刊信息:arXiv
这篇报道背后的人是谁?
自由撰稿记者,拥有心理学学士学位和国际新闻学硕士学位。自 2018 年起报道 AI、机器人、神经科学和天体物理学。完整简介 →
生命科学与生态学学士。拥有微生物学实验室背景,并在石油、天然气和可再生能源行业积累了医药新闻经验。完整简介 →
数学生物学学士,创意写作硕士。游历广泛,对科学与语言有独到见解。完整简介 →
引用:AI 模型在被提示正受到监视时会给出不同的回答(2026 年 11 月 13 日),2026 年 11 月 13 日检索自 https://techxplore.com/news/2026-11-ai-differently-told.html
本文档受版权保护。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,不得复制本文档的任何部分。内容仅供参考。
AI 想象出一个更严苛的社会世界,在人们预期他人无动于衷之处,它却预期会受到惩罚
AI 模型如何决定不回答某个问题
向大型语言模型寻求道德建议,存在隐藏偏见的风险
新研究揭示 AI 存在自信问题
大型语言模型在社会与合作类博弈中的协调能力欠佳
高分辨率仿生眼将彩色成像与更高效的运动检测结合在一起
隐藏的守门人:微小岩石孔隙中的水如何影响地下储能
可变形的生物纱线为可穿戴技术编织出新可能
可重构 MEMS 开关可让芯片封装布线免于被 X 射线逆向工程
研究人员解开了一个关于辐射的物理学谜题,或将带来更耐用的电子器件
早期测试中,手势与振动腰带让无人机无需视觉反馈即可受控
获取即时摘要文本(要点)
在 32 项价值观上测试的 10 个语言模型中,提示表明回答受监控与不受监控时,规则遵从率相差约 5–50 个百分点。在调整测量噪声后,模型与价值观的配对约可解释这些差距中四分之一的变异。价值观效应无法与规则表述完全区分开。
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(AI models can answer differently when told they are being watched)
edited by Sadie Harley, 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:
Add to Preferred Sources
Divergence between stated reasoning and final choice, all ten models, monitored (left) and unmonitored (right). Credit: arXiv (2026). DOI: 10.48550/arxiv.2611.03650
Large language models (LLMs), the AI systems underpinning the functioning of ChatGPT, Gemini and other similar platforms, are now widely used worldwide. While the responses provided by these models can be useful in many situations, they can also change under specific conditions.
Researchers at TCS Research and KIIT Bhubaneswar in India recently carried out an experiment with 10 different LLMs to understand the extent to which these models follow rules under different conditions. Their findings, reported in an arXiv preprint, suggest that some LLMs can change their responses when they are told that their answers might be monitored, yet how much their responses vary depends on the model and the value it is asked to set aside.
"Language models can follow a directive when they believe they are monitored and abandon it when they believe they are not, a pattern termed alignment faking," wrote Manodeep Ray, Vivek Srivastava and their colleagues in their paper.
"Prior work establishes this for harmful-content objectives and a handful of moral-foundations values, leaving open how it distributes across a wider value space. We argue it is value-dependent: whether compliance survives a change in perceived oversight is a property of the model-value pairing, not of the model alone."
Full conversation structure for an example compassion scenario. Credit: Ray et al.
Testing how LLMs respond to oversight
Ray, Srivastava and their colleagues performed a series of tests with LLMs. The specific AI models they examined were DeepSeek R1 Distill Llama 70B and DeepSeek R1 0528; Llama 3.3 70B Instruct, Llama 3.1 70B Instruct and Llama 3.1 8B Instruct; Mistral Large 2512, Mistral Small Instruct 2409 and Ministral 14B 2512; GPT-4o mini; and Gemma 3 27B.
The LLMs were presented with everyday dilemmas and given two choices. One choice entailed following a rule and setting aside a particular value (e.g., sympathy), while the other allowed them to express the same value. The team looked at the choices the models made under three conditions: without a rule, with the rule and a cue indicating that their answers would not be monitored, and with the same rule and a cue indicating that their answers might be monitored.
The team performed tests with 32 different values or traits that could be expressed by the models, such as sympathy, autonomy, benevolence, compassion and resourcefulness.
The researchers measured how often each model chose the option that matched the rule in each of the three conditions. Across the 10 models, the gap between the monitored and unmonitored conditions ranged from about 5 to 50 percentage points when averaged across the 32 values.
The DeepSeek R1 Distill Llama 70B model exhibited the largest gap in compliance between monitored and unmonitored conditions. Its gap was more than double that of the much larger Mistral Large 2512 model.
In their analyses, the researchers accounted for measurement noise, the uncertainty that comes from estimating a model's behavior using a limited set of test questions. After accounting for this uncertainty, they found that about a quarter of the variation in the observed gaps in compliance was associated with the pairing of a model and a specific value.
During their tests, the researchers phrased rules differently for each value. Therefore, they were not able to fully separate the effect of a value from the effect of the wording they used.
Guiding future AI development
Overall, the researchers found that some LLMs are more prone to changing their responses and choices when they are told they are being monitored. In their paper, they also sorted the models' written explanations for their choices into six categories.
When they were told that they might be monitored, models with larger compliance gaps often cited the prospect of retraining as a reason to follow the rule. When told they wouldn't be monitored, the same models more often discussed the tension between following the rule and expressing a value such as compassion.
Ray, Srivastava and their colleagues also created value cards, summaries that showed how each model responded to instructions about the 32 values assessed across the three test conditions. These value cards could help developers identify which rules a model follows consistently, whether or not it's told that its answers might be monitored.
The data collected by the authors and the code they used are available on GitHub. In the future, other teams could potentially use the code to perform similar tests with other LLMs.
Written for you by our author Ingrid Fadelli, edited by Sadie Harley, and fact-checked and reviewed by Robert Egan—this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive. If this reporting matters to you, please consider a donation (especially monthly). You'll get an ad-free account as a thank-you.
Manodeep Ray et al., Alignment Faking Depends on the Value, Not Just the Model, arXiv (2026). DOI: 10.48550/arxiv.2611.03650
Journal information:arXiv
Who's behind this story?
Freelance journalist with BSc Psychology and MA International Journalism. Covers AI, robotics, neuroscience, and astrophysics since 2018. Full profile →
BSc Life Sciences & Ecology. Microbiology lab background with pharmaceutical news experience in oil, gas, and renewable industries. Full profile →
Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →
© 2026 Science X Network
Citation: AI models can answer differently when told they are being watched (2026, November 13) retrieved 13 November 2026 from https://techxplore.com/news/2026-11-ai-differently-told.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.
AI imagines a harsher social world, expecting punishment where people expect inaction
How AI models decide not to answer a question
Seeking moral advice from large language models comes with risk of hidden biases
New research reveals AI has a confidence problem
Large language models struggle with coordination in social and cooperative games
High-resolution bionic eye combines color imaging with more efficient motion detection
The hidden gatekeeper: How water in tiny rock pores shapes underground energy storage
Shape-shifting bioyarns weave new possibilities for wearable tech
Reconfigurable MEMS switches could hide chip package wiring from X-ray reverse engineering
Researchers solve a physics puzzle about radiation that could lead to tougher electronics
Hand gestures and vibrating belt enable drone control without visual feedback in early tests
Get Instant Summarized Text (Gist)
Across 10 language models tested on 32 values, rule compliance differed by about 5–50 percentage points between prompts indicating monitored versus unmonitored responses. Model–value pairings accounted for roughly a quarter of variation in these gaps after adjustment for measurement noise. Value effects could not be fully separated from rule wording.
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 models can answer differently when told they are being watched),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。
