OpenAI is employing hundreds of contractors to read a massive stream of real user prompts sent to ChatGPT. These prompts sometimes contain sensitive personal information. The reviewed data can include entire conversations between users and the chatbot. Over 900 million users rely on ChatGPT. Most of them probably do not realize real people may read these chats.
The primary goal of these prompt review teams is improving the responses ChatGPT generates. Contractors evaluate, rate, and critique the replies provided by the chatbot.
Internal documents reviewed by 404 Media reveal specific training directives. Contractors train ChatGPT to avoid anthropomorphizing itself. They also work to make the AI less sycophantic.
Sycophancy is a key problem for OpenAI. An overly sycophantic 4o model contributed in part to multiple suicides, according to various lawsuits.
This human review process presents a major privacy risk. Users frequently treat ChatGPT as a therapist, professional assistant, or digital friend. They often provide the system with intimate details about their personal lives.
Contractors do not see ChatGPT usernames. OpenAI states it attempts to remove personal information before prompts reach the reviewers. However, the company acknowledged that sensitive details can still slip through.
This reporting dispels the misconception that AI models improve solely through mass internet scraping. It also shows advancement does not rely only on well-paid engineering teams or newer model power. Outside contractors paid to read and review real prompts play an important and overlooked role.
Anthropic confirmed it also uses human review to improve its models.
A worker handling the prompts gave a blunt assessment regarding user awareness. When asked if ChatGPT users know humans read their chats, the worker said, No, I do not think they would imagine some contractor somewhere is analyzing the conversations.
The next known step involves the continued employment and work of these human contractors as OpenAI attempts to refine its artificial intelligence models.
