Hundreds of contract workers review real ChatGPT conversations and score responses from one to seven, according to an investigation by 404 Media reported by The Decoder. The work includes efforts to reduce excessive flattery and overly human-like behavior in model output.
OpenAI anonymizes prompts and applies a privacy filter, but acknowledges that filtering can make mistakes. Reviewed conversations can therefore still contain sensitive personal details, including exchanges in which users explicitly asked the chatbot to keep information private.
Users can prevent new conversations from being used for model improvement—and potentially reaching human reviewers—by turning off “Improve the model for everyone,” which is enabled by default. The choice does not apply retroactively. OpenAI also offers temporary chats that it says are not used for training.
An OpenAI FAQ says authorized personnel and service providers may view user data to improve performance and advises people not to enter sensitive information. The criticism is that this disclosure and the setting’s label do not make human review sufficiently obvious. Human feedback is common across major AI services, but users should treat consumer chatbots as systems where authorized people may inspect eligible conversations, not as confidential vaults.