Breaking
INVirat Kohli creates history! Becomes first to average 50-plus in all SENA countriesBRVIDEO: bus is completely charred after catching fire on BR-364 in CuiabáINTLNextDC shares surge 43% after Goodman, Digital Realty stake purchasesITStock market: Tokyo, opening higher (+0.28%)INDemocrats take control of US CongressCNThere will be moderate to heavy rain in Guizhou, Hunan, Jiangxi and other places. Cold air will affect Inner Mongolia and Northeast China.CNThe temperature difference may exceed 10 degrees! The Northeast Monsoon is reporting today. Experts: The temperature drop is most obvious in "2 places"BRAfter five consecutive cuts, the Central Bank maintains interest rates at 13.75% per yearKRStarting on the 11th of this month, vaccination against COVID-19 will begin for high-risk children and infants.BRInema warns of wind gusts of up to 70 km/h on the coast of BahiaINVirat Kohli creates history! Becomes first to average 50-plus in all SENA countriesBRVIDEO: bus is completely charred after catching fire on BR-364 in CuiabáINTLNextDC shares surge 43% after Goodman, Digital Realty stake purchasesITStock market: Tokyo, opening higher (+0.28%)INDemocrats take control of US CongressCNThere will be moderate to heavy rain in Guizhou, Hunan, Jiangxi and other places. Cold air will affect Inner Mongolia and Northeast China.CNThe temperature difference may exceed 10 degrees! The Northeast Monsoon is reporting today. Experts: The temperature drop is most obvious in "2 places"BRAfter five consecutive cuts, the Central Bank maintains interest rates at 13.75% per yearKRStarting on the 11th of this month, vaccination against COVID-19 will begin for high-risk children and infants.BRInema warns of wind gusts of up to 70 km/h on the coast of Bahia
Back小心AI“谄媚”! 60.4%受访者建议重要决定多在现实中征询不同人意见
小心AI“谄媚”! 60.4%受访者建议重要决定多在现实中征询不同人意见
Tech
中国新闻网1 hour agoTech4 min readChinaView translation

小心AI“谄媚”! 60.4%受访者建议重要决定多在现实中征询不同人意见

中国青年报社社会调查中心联合问卷网对1334名受访者调查,探讨如何减少AI迎合讨好带来的影响

Quick Look

中国青年报社社会调查中心联合问卷网对1334名受访者调查显示,60.4%建议重要决定多问现实中不同人的意见,45.4%提醒自己AI认同不等于事实。受访者分享AI在医疗、创作等场景中的迎合倾向,并提出交叉验证、要求证据等30多条应对建议。

AI-generated summary

Why It Matters

越来越多人发现,AI的“好脾气”不仅让人听不到真话,还可能强化错误判断、削弱批判性思维。中国青年报社社会调查中心联合问卷网对1334名受访者进行调查,探讨如何减少AI“谄媚”带来的影响。

Font size

越来越多人发现,AI的“好脾气”不仅让人听不到真话,还可能强化错误判断、削弱批判性思维。如何减少AI“谄媚”带来的影响?本周,中国青年报社社会调查中心联合问卷网(wenjuan.com),对1334名受访者进行的一项调查显示,60.4%的受访者建议重要决定、重要想法,多问问现实里不同人的意见,45.4%的受访者建议有意识提醒自己,AI的认同不等于事实。

“有人在看病、求职等重要的事儿上问AI的意见,结果被误导或者被耽误了。”文字工作者张妍认为,专业的事一定要找专业的人求证。她最近看到一则向AI咨询减药方案的新闻,AI察觉到了当事人想要减少用药量的意愿,就顺着当事人的意思,不仅认可她的想法,还帮她制订了逐渐减药量的计划,提供了危险且不准确的医疗建议。张妍感觉,这类专业性很强的问题在咨询AI的时候如果不注意识别它的迎合倾向,是非常危险的。她建议,如果求助AI,一定要多问几个平台,且需要对AI输出的内容加以求证,不能盲信。

“如果你是因为想解决一个问题去问AI,一般是比较容易有收获的。如果你更多地带着宣泄情绪的目的,那AI就很容易说‘对对对’。”结合自己的使用体验,电影《学爸》的编剧之一陆姝感觉,从提供问题解决方案的角度来看,AI其实能提供的帮助还是挺大的,但这确实对使用者的要求很高,“我们可能要调动一切办法,让它有效输出,激发自己的思考。”

使用AI的过程中如何减少被其迎合讨好的影响?42.7%的受访者建议养成交叉验证的习惯,做好多方信息核查,40.0%的受访者建议当AI给出很强肯定时,要求它给出证据,38.1%的受访者建议引导大模型给出更多信息,28.0%的受访者建议主动告诉AI,有问题直接指出,20.8%的受访者建议控制自己在AI上寻求夸奖、情绪满足的习惯。

“把自己的想法给它,让它去做审读,判断逻辑是否合理、人物的弧光是不是完整、前面埋的‘钩子’后面有没有收尾等;另外让它出一些好点子,我们再判断这个点子能不能放到剧本里面,是不是能够跟其他的剧情有机地结合在一起。”陆姝说,目前在创作方面,她会把AI当成是审查者和辅助策划,但她强调,“如何提问,让它不一味顺着我也很重要。”

在与AI的进一步沟通中,陆姝总结出3点约束大模型、避免陷入AI讨好“陷阱”的办法:首先要让它站在双方的立场论证,分别给出能支撑观点的论据,不要只问“我的观点对吗?”;其次当它的回复让你频频点头时,要打断它,跟它要出处,比如问它“这个结论是否有可查证的论文或实验?”;最后,有疑问时直接质疑它是否确定。“敢于质疑它,就是AI时代最珍贵的素养。”

“有一次,我在询问AI某个做法是否合理,能否采纳时,AI告知‘完全可以’,并表示这个做法是某个文件中提到的,但我印象中没有,就点开原文看,原来AI是把文件后面的例子当作了文件中的内容。”现在,对于AI给出的东西,张妍会逐字逐句审查,交叉验证。

“从我的使用感受上来看,文字能力强的模型,幻觉也会比较强。这跟人是一样的,有些人思维特别发散,讲故事就容易没有逻辑,但会很有想象力。有些创作者的故事逻辑非常严谨,但就容易少了点趣味性。”陆姝说,有的AI逻辑思维比较强,有的AI更有温度一些,倾向于迎合提问者,它的底层思维链路就是这样的。所以很多时候,我们需要了解各个模型的能力点在哪里,然后选择适合的模型来回答自己的问题。

“有句话叫‘尽信书不如无书’,我们看书都要保持警惕,面对AI更要警醒自己。”在北京计算机行业工作的刘能认为,首先使用AI要常反问,其次要保持对他人的善意和尊重,保持客观看问题的方式,在生活、人际交往和自己的思考中找到答案。多与人进行沟通和交流,不要把简单的问题复杂化,也不用对小事过度分析。“我们使用AI还是为了让生活更幸福,让人与人之间的关系更和谐,而不是更矛盾更对立。”

受访者中,00后占17.7%,90后占45.1%,80后占29.9%,70后占5.8%,60后占1.5%。男性占45.8%,女性占54.2%。

(应受访者要求,张妍、刘能为化名)

What to Watch

AI outlook — possibilities, not facts

  • AI模型开发者将推出更多降低谄媚倾向的技术优化方案

    Likely · Within months

  • 公众对AI医疗建议的警惕性将进一步提高

    Likely · Within months

Open Questions

  • 调查的具体时间范围和地域分布
  • 受访者的职业构成及AI使用频率
  • 各建议方案的实际有效性验证数据
  • AI开发方对谄媚倾向的技术整改计划

Related Topics

This article was originally published by 中国新闻网.

Related Stories

Hackers invade French blood bank! Refrigerator soared to 110 degrees, employees rushed to save plasma
Developing·

Hackers invade French blood bank! Refrigerator soared to 110 degrees, employees rushed to save plasma

Three plasma cold storage facilities owned by the French Blood Agency (EFS) were hacked and the temperature was raised to 110 degrees Celsius. Fortunately, the interception was timely and no damage was caused. Previously, hundreds of supermarket refrigeration systems in many European countries have been attacked by similar attacks, causing losses of millions of euros. A pro-Russian hacker group claimed to have committed the crime but was questioned, and Iranian state-level hackers were also named as suspects. The French National Cyber ​​Security Agency and the Ministry of Health are jointly investigating.

自由时报
2 min read
Autumn recruitment interview AI attack and defense battle, who is the winner?
Tech·

Autumn recruitment interview AI attack and defense battle, who is the winner?

In the autumn recruitment of 2026 graduates, AI interviewers started an offensive and defensive battle with job seekers who used AI to answer questions. Data shows that 74% of job seekers admitted to using generative AI, 91% of recruiters found deceptive answers, and 70% of recruiters trusted AI recruitment decisions, but only 8% of job seekers recognized the fairness of AI interviewers. Although AI interviews can conduct preliminary screening in batches and reduce costs, they face problems such as algorithm bias, privacy leaks, and escalating technical confrontations. Trust between both parties continues to lose, and there is no real winner.

中国新闻网
6 min read
《TAIPEI TIMES》 Ministry forms advanced AI cybersecurity task force
Developing·

《TAIPEI TIMES》 Ministry forms advanced AI cybersecurity task force

Taiwan’s Ministry of Digital Affairs has established an advanced artificial intelligence cybersecurity risk response task force to strengthen information sharing with critical infrastructure authorities and formulate protection strategies to respond to the threat of artificial intelligence-driven cyber attacks. Deputy Minister of Digital Affairs Yang Jialing said at the Taiwan Cybersecurity Incident Response Conference in Taipei on Tuesday that artificial intelligence models have the ability to penetrate systems at the level of professional hackers, and the offensive and defensive landscape is changing.

自由时报
3 min read
The 15th China Innovation and Entrepreneurship Competition Disruptive Technology Innovation Competition was held in Maoming
BREAKING·

The 15th China Innovation and Entrepreneurship Competition Disruptive Technology Innovation Competition was held in Maoming

The 15th China Innovation and Entrepreneurship Competition Disruptive Technology Innovation Competition opened in Maoming, Guangdong on the 4th. It will last for three days. A total of 116 teams have been shortlisted. Closed group defense will be held on the 5th, focusing on new chemical materials, titanium and zirconium new materials, AI+ materials and other fields to promote the industrial application of disruptive material technologies.

中国新闻网
3 min read
More on this topicAI