OpenAI Had Contractors Rate ChatGPT Conversations

OpenAI has been having hundreds of contractors read and rate actual ChatGPT conversations. According to reporting by media outlet 404 Media, the conversations are scored on a 1-7 scale. One purpose of this evaluation is to mitigate the tendency for AI to be excessively flattering or behave more human-like than necessary.

Improving the quality of AI models requires a process in which humans read actual outputs and judge their quality. This technique is called "Reinforcement Learning from Human Feedback" (RLHF) and is widely adopted in the development of large language models like ChatGPT. In other words, this report can be seen as part of a development process common to the entire AI industry, not a practice unique to OpenAI.

From a privacy perspective, however, there are points that warrant caution. While the conversation data used for evaluation is anonymized, it is impossible to rule out the possibility that personal information and sensitive content may remain despite the anonymization process. This is because the prompts users send on a daily basis often contain information about health, work, and relationships that they do not expect others to see.

What further complicates the issue is the default setting. In ChatGPT, the "Improve model for everyone" setting is turned on by default, and when this setting is enabled, conversation content may become subject to evaluation. In other words, users who do not want humans to read their conversations must manually disable this setting themselves. The mechanism that makes conversations subject to evaluation by default unless the user takes action can be likened to a state of consent without the user's awareness.

This structure is an example of the "opt-out" approach seen widely in digital services. A design that requires users to actively choose rather than treating data usage as consented in the default state always involves the question of where to place the balance between convenience and privacy protection. Among regulators and researchers, discussions questioning the appropriateness of such default settings continue.

This report serves as an opportunity for users who routinely use generative AI services to reexamine how their data is being handled. Currently, the responsibility falls on users themselves to check the settings menu and make choices aligned with their intentions. How AI companies will increase transparency regarding the scope of data usage in the future will likely become an important issue affecting the credibility of the entire industry.