This practice highlights a significant privacy gap for the platform’s 900 million users, many of whom may share sensitive information under the assumption of digital privacy.
The data training setting that enables this human review is active by default for all free, Plus, and Pro accounts, requiring users to manually opt-out to prevent their chats from being reviewed.
This reliance on human labor underscores that AI development depends not just on massive data scraping and engineering, but on constant manual intervention by contractors—some earning over $50 an hour—to refine the software’s social behavior and helpfulness.
Reviewers follow a three-stage process: reading a prompt, summarizing the user’s intent, and rating four potential AI responses on a scale of one to seven.
They are instructed to penalize "AI-speak," such as the excessive use of emojis, and to ensure the chatbot remains professional without implying it has emotions.
While Enterprise and Education accounts have these features disabled by default, individual users are affected by the company's data usage policies, which allow de-identified content to be used for model improvement even after a conversation is deleted.