大模型
www.aibase.com ·
你的ChatGPT对话可能被真人看过：OpenAI人工审核机制曝光
据404Media调查，OpenAI雇佣数百名合同工审核真实ChatGPT用户对话，用于改进模型回复质量。审核人员按1至7分对回复评分，重点减少过度谄媚等问题。合同工由Crossing Hurdles招募、AI训练公司Mercor支付薪酬，有北美评论者称时薪超过50美元。OpenAI的'Improve the model for everyone'设置默认开启，用户可关闭以阻止新对话记录用于模型训练及相关人工审核，临时聊天模式不使用输入数据训练。OpenAI FAQ自2023年起已说明授权人员和服务提供商可能查看用户数据，并建议不要输入敏感信息；404Media认为该披露较为隐晦。Anthropic和Google也有类似人工审核安排。
AI推荐理由 基于404Media调查，披露OpenAI雇佣数百名合同工审核真实ChatGPT对话、按1-7分评分以减少过度谄媚等问题，并给出合同工招募方（Crossing Hurdles、Mercor）和时薪超过50美元等具体细节，同时说明默认开启的'Improve the model for everyone'设置及临时聊天模式的数据使用差异。信息密度较高，涉及用户隐私与AI训练数据使用的实际权衡，对开发者和用户有参考价值。局限在于内容为转述调查报道，未提供原始文件，且OpenAI人工审核机制在FAQ中自2023年起已有披露，并非全新事实。
72
正文
Your ChatGPT Conversations May Have Been Read by Real People: OpenAI's Human Review System Exposed
AIbase基地
Published inAI News · 4 min read · Sep 15, 2026
8
According to an investigation by 404Media, OpenAI employs hundreds of contract workers to review real ChatGPT user conversations to improve the quality of the model's responses. Leaked internal documents and sources indicate that reviewers rate ChatGPT's responses on a scale from 1 to 7, focusing on helping reduce issues such as excessive flattery and over-animality.
Some commentators believe users are unaware that their chat records may be read by real people; even anonymized conversations may still contain sensitive personal information. Although OpenAI has deployed privacy filters, it has also admitted that they may sometimes make mistakes.
According to reports, these contract workers were recruited by Crossing Hurdles and paid by the AI training company Mercor. A North American commentator said their hourly wage exceeded $50. OpenAI currently has the "Improve the model for everyone" setting enabled by default. Users can turn off this option to prevent new chat records from being used for model training and related human reviews; the temporary chat mode does not use input data for model training.
It is worth noting that OpenAI has not recently disclosed its human review mechanism. Its FAQ has at least since 2023 stated that authorized personnel and service providers may view user data to improve model performance and advised users not to enter sensitive information. 404Media believes this disclosure is somewhat obscure, and users may find it difficult to intuitively understand the actual scope of human reviews.
Human feedback has become an important part of optimizing large models. Anthropic has also confirmed that it will conduct human reviews of Claude conversations with relevant privacy settings enabled and anonymize them before reviewing; Google also allows human reviewers to check some Gemini chat records. This incident once again highlights the balance between AI model training efficiency and user data privacy.
ChatGPTOpenAIAI Training CompanyPrivacy Filter
This article is from AIbase Daily
Scan to view
Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications.
—— Created by the AIbase Daily Team
© Copyright AIbase Base 2024, Click to View Source -