按年被提到
20252026 · 14 次2026
AI 安全、伦理与社会影响›大模型与人类行为·arXiv:2503.23674 · cs.CL

Large Language Models Pass the Turing Test

大语言模型通过了图灵测试
Cameron R. Jones, Benjamin K. Bergen · 2025

三方图灵测试中 GPT-4.5 有 73% 的时间被判为真人

开始阅读 ↗PDFalphaXiv 讨论
被提及 · 16 人首次 2025-04-02 · 最近 2026-10-03
16不同的人
12活跃月份
21讨论串
33被引(OpenAlex)

作者在两个独立人群上做了随机对照、预先注册的三方图灵测试,评估 ELIZA、GPT-4o、LLaMa-3.1-405B 和 GPT-4.5:参与者同时与一名真人和其中一个系统各聊 5 分钟,再判断谁是人类。被提示扮演类人角色时,GPT-4.5 有 73% 的时间被判为人类,显著高于真人被选中的比例;LLaMa-3.1 为 56%,与真人无显著差异;基线 ELIZA 和 GPT-4o 只有 23% 和 21%,显著低于随机水平。作者称这是首个人工系统通过标准三方图灵测试的实证。

HN 为什么提它

评论者拿它说明 LLM 已经通过标准的三方图灵测试,常引用 GPT-4.5 的 73% 并以 ELIZA 作基线对照,也有人批评该测试其实只能证明很少。

实证研究难度 2 / 5约 30 分钟有时效性
读之前最好知道
1图灵测试(Turing test)
2随机对照实验设计
3LLM 人设提示(persona prompt)
4统计显著性

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
批评Kotlopou · 2026-10

“I think it only proves very little, because a typical conversation they studied looks like this:”

批评这项图灵测试证明的东西很少,并以典型对话为例
Vote on which of Hacker News' challenges for AI have been met原评论 ↗
引用akoboldfrying · 2025-06

“ChatGPT 4.5 was judged to be the human 73% of the time in this RCT study, where human interrogators had 5-minute conversations with a human and an LLM: Large Language Models Pass the Turing Test”

引用 RCT 结果:5 分钟对话中 GPT-4.5 有 73% 的时间被判为人类
The Illusion of “The Illusion of Thinking”原评论 ↗
引用ufmace · 2025-06

“I only skimmed it, but I don't see anything clearly wrong about it.”

只略读过,没看出明显错误,论文称 GPT-4.5 加人设提示能通过图灵测试
What happens when people don't understand how AI works原评论 ↗
引用sigmoid10 · 2025-06

“but some of the most recent models have actually passed this test [1].”

拿它说明一些最新模型实际上已经通过图灵测试
The Consciousness Gradient: When Machines Begin to Wonder原评论 ↗
引用tim333 · 2025-10

“I think it illustrates both the limits of the test and that LLMs are getting quite good.”

认为它既体现了该测试的局限,也说明 LLM 已相当出色,75% 的评判者选 GPT-4o 为人类
John Searle has died原评论 ↗
引用ben_w · 2025-12

“The latter: Large Language Models Pass the Turing Test”

贴出论文,说明三方图灵测试中 LLM 被判为人类
Ask HN: Do you believe any LLM's pass the Turing test? How?原评论 ↗
看全部 20 条 →
引用bonplan23 · 2025-12

“With regards to strong AI, there was a widely accepted definition - i.e. passing the Turing Test - and we are way past that point already: ( see Large Language Models Pass the Turing Test )”

指出强 AI 公认的定义是通过图灵测试,而 LLM 早已越过,附上这篇论文
Show HN: Use Claude Code to Query 600 GB Indexes over Hacker News, ArXiv, etc.原评论 ↗
引用canjobear · 2026-01

“At this point LLMs usually beat humans at the Turing Test! People are more likely to pick the LLM as the human, rather than the human. Large Language Models Pass the Turing Test”

指出 LLM 通常已能在图灵测试中胜过人类,人们更常把 LLM 当作人
LLM Problems Observed in Humans原评论 ↗
引用coldtea · 2026-02

“That researchers haven't done Turing-test studies (taking the setup from turing and even called them that) is patently false. Including openly testing LLMs:”

反驳说研究者没做过图灵测试是明显错误,包括公开测试 LLM 的这项研究
Frontier AI agents violate ethical constraints 30–50% of time, pressured by KPIs原评论 ↗
引用shakna · 2026-02

“And still had people think it was a real person, over 27% of the time.”

指出在对抗性测试中 ELIZA 也让人把它当真人超过 27% 的时间
I don't know how you get here from "predict the next word."原评论 ↗
引用wat10000 · 2026-03

“such a test was run last year and GPT-4.5 and LLaMa-3.1-405B both passed.”

指出去年就做过这类测试,GPT-4.5 和 LLaMa-3.1-405B 都通过了
Gemini Said They Could Only Be Together If He Killed Himself. Soon, He Was Dead原评论 ↗
引用ufmace · 2026-03

“I had already found this study, published almost a year ago, saying that they do:”

说自己早已找到近一年前发表的这项研究,结论是 LLM 能通过
Based on its own charter, OpenAI should surrender the race原评论 ↗
引用tim333 · 2026-03

“In the test mentioned in nearby comments ELIZA only got 27% suggesting the test wasn't that easy to fool.”

指出该测试中 ELIZA 只得 27%,说明测试并不容易被骗过
Marriage over, €100k down; AI users whose lives were wrecked by delusion原评论 ↗
引用jmalicki · 2026-04

“73% judged GPT 4o to be the human.”

引用论文数据:73% 的人判定 GPT 4o 是人类
We found an undocumented bug in the Apollo 11 guidance computer code原评论 ↗
引用antonvs · 2026-05

“By most normal standards, they can definitely pass the Turing test.”

认为按大多数通常标准,LLM 当然能通过图灵测试
When Dawkins met Claude – Could this AI be conscious?原评论 ↗
引用akoboldfrying · 2026-05

“When prompted to adopt a humanlike persona, GPT-4.5 was judged to be the human 73% of the time”

引用摘要:用人设提示时 GPT-4.5 有 73% 的时间被判为人类
GitLab Announces Workforce Reduction and End of Their CREDIT Values原评论 ↗
引用akoboldfrying · 2026-08

“From the abstract: "When prompted to adopt a humanlike persona, GPT-4.5 was judged to be the human 73% of the time”

引用摘要:用人设提示时 GPT-4.5 有 73% 的时间被判为人类
Qwen 3.8 27B is excellent, but it defaults to overthinking things原评论 ↗
引用akoboldfrying · 2026-08

“From the abstract: "When prompted to adopt a humanlike persona, GPT-4.5 was judged to be the human 73% of the time”

引用摘要:用人设提示时 GPT-4.5 有 73% 的时间被判为人类
I'm Becoming AI-Blind原评论 ↗
引用drusepth · 2026-09

“LLMs started meaningfully passing the Turing test a year or two ago, around GPT-4.5.”

指出 LLM 大约在 GPT-4.5 前后已能真正通过图灵测试,加人设时 73% 被判为人类
GPT-6 Astra原评论 ↗
引用Topfi · 2026-10

“has been beaten by all manner of models.”

指出这一测试已被各类模型击败,包括 ELIZA 之类的旧系统
An AI agent emailed researchers for help. It told us why原评论 ↗