按年被提到
20252026 · 12 次2026
大模型与语言模型›对话式大模型·arXiv:2501.15654 · cs.CL

People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text

经常用 ChatGPT 写作的人是准确且稳健的 AI 生成文本检测者
Jenna Russell, Marzena Karpinska, Mohit Iyyer · 2025

常用 LLM 写作的标注者检测 AI 文本,胜过多数检测器

开始阅读 ↗PDFalphaXiv 讨论
被提及 · 16 人首次 2025-01-29 · 最近 2026-08-21
16不同的人
10活跃月份
18讨论串
31被引(OpenAlex)

作者雇用标注者阅读 300 篇英文非虚构文章,判断其为人写还是 AI 生成(GPT-4o、Claude、o1),并写出段落式理由。结果显示经常用 LLM 写作的标注者无需专门训练或反馈就很擅长检测,五位这类“专家”的多数票只误判了 300 篇中的 1 篇,明显优于多数商用和开源检测器,即使面对改写和“人性化”等规避手段也如此。定性分析显示他们既依赖具体的词汇线索(“AI 词汇”),也能察觉正式程度、原创性和清晰度等更复杂的现象。数据集与代码已发布。

HN 为什么提它

评论者在争论人类和检测工具(如 Pangram)能否可靠识别 AI 文本时,把它当作独立研究证据:常用 LLM 写作的人准确率很高、误判极少。

实证研究难度 2 / 5约 30 分钟有时效性
读之前最好知道
1LLM 文本生成基础
2AI 生成文本检测器的评测方式
3标注一致性与多数投票

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
引用Jweb_Guru · 2025-11

“Human experts can reliably detect some kinds of long-form, AI-generated text using exactly the same sorts of cues”

拿它说明专家能凭同样的线索可靠识别长篇 AI 生成文本
When your hash becomes a string: Hunting Ruby's million-to-one memory bug原评论 ↗
引用samdoesnothing · 2025-12

“However there is evidence that writers who have experience using LLMs are highly accurate at detecting AI generated text.”

拿它说明有 LLM 使用经验的写作者检测 AI 文本非常准确
“You should never build a CMS”原评论 ↗
引用spindump8930 · 2026-04

“There are recent papers from academics independently bench-marking and studying detectors e.g. People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text”

举例说明有学者独立评测和研究检测器
Ask HN: How do systems (or people) detect when a text is written by an LLM原评论 ↗
引用kaonwarb · 2026-05

“Not perfect, but I think as good evidence as any:”

拿它作证据,说明常用 ChatGPT 写作的人能较好识别 AI 文本,虽不完美
Fisker went bankrupt and owners built an open source car company from the ashes原评论 ↗
引用Groxx · 2026-05

“shows 98% for example.”

指出研究显示检测准确率约 98%
Americans Are Smashing Flock Cameras原评论 ↗
引用embedding-shape · 2026-05

“Said Pangram report literally citing the single evidence of em-dashes...”

批评 Pangram 报告只拿破折号当唯一证据,并提到这项重度用户检测研究
I'm Tired of Talking to AI原评论 ↗
看全部 12 条 →
引用Chu4eeno · 2026-06

“It's also confirmed in third-party studies like this People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text”

拿它作第三方研究,印证常用 ChatGPT 的人能准确稳健地检测 AI 文本
"No, I swear I wrote this."原评论 ↗
引用Chu4eeno · 2026-06

“It has extremely low false positive rate in various tests (e. g. People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text),”

引用测试说明检测误报率极低
Looking Ahead to Postgres 19原评论 ↗
引用Chu4eeno · 2026-07

“On the humans (vs. pangram) ability to reliably spot LLM generated content, there was an independent study: People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text”

引用独立研究,说明人类能否可靠识别 AI 生成内容
Ask HN: Add flag for AI-generated articles原评论 ↗
引用chis · 2026-08

“This has been pretty well studied. Pangram is about as good at an expert human with a 98% detection rate and <2% false positive, for flagging AI generated text.”

指出这点已有充分研究,Pangram 与专家相当,检出率 98%、误报低于 2%
Where .env Went Wrong原评论 ↗
引用Seattle3503 · 2026-08

“There are independent evaluations [1, 2] showing it works,”

引用独立评测,说明这类检测确实有效
Branchless Rust: Making a Filter 4x Faster by Removing an If原评论 ↗
引用neom · 2026-08

“People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text - People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text”

直接贴出论文,说明常用 ChatGPT 写作的人能准确稳健地检测 AI 文本
I'm Becoming AI-Blind原评论 ↗
被提交为 story · 6 次