Project Lily: Why humans may review your ChatGPT chats

Someone may be reading your ChatGPT conversations. That is the privacy question raised by Project Lily, a reported OpenAI initiative in which contract workers review real ChatGPT conversations to assess the quality and safety of the chatbot’s replies.

According to the report, the goal is to determine whether responses are clear, useful and aligned with OpenAI’s rules. But the practice raises an obvious concern: people increasingly use ChatGPT for personal, professional and sensitive questions.

Why human reviewers are looking at chats

Project Lily is said to involve human reviewers assessing ChatGPT conversations and judging how well the AI performs. Human feedback is common across the AI industry, helping companies identify whether models are useful, handle harmful requests appropriately and follow safety standards.

The main problem is what those conversations can contain. Users may type health worries, workplace documents, family issues, financial details, legal questions, code, business ideas or personal identifiers into ChatGPT. Even when review is intended to improve the service, users may not expect a person to see parts of their conversations.

OpenAI’s privacy filter has a role

OpenAI uses privacy and safety systems to reduce exposure of sensitive information. The NewsBytes report says the company has a “Privacy Filter” designed to hide sensitive details in chats before they are reviewed, although it may not always work perfectly.

That limitation has added to concerns around Project Lily. OpenAI’s own help pages state that conversations may be reviewed to improve systems and ensure compliance with policies and safety requirements.

The company also says content from individual services such as ChatGPT may be used to train models unless users opt out. Human review, therefore, is not entirely absent from OpenAI’s policies, but many users may not read those settings closely.

How to limit the use of your chats

Users who do not want their conversations used to improve OpenAI’s models can disable “Improve the model for everyone” in ChatGPT. The option is available under Data Controls for signed-in users. OpenAI says future conversations will not be used for training after it is disabled.

Temporary chats offer another option for sensitive questions. OpenAI says these chats are not used to train models and do not contribute to ChatGPT memory. Still, users should avoid sharing highly sensitive information unless it is necessary.

The bottom line

Project Lily highlights a basic reality of AI assistants: they are online services that process information to answer questions, improve performance and enforce rules.

For users in India and elsewhere, the practical approach is to check privacy settings, avoid pasting confidential documents and think carefully before sharing sensitive information. As AI becomes part of everyday work and study, companies will need to make human review practices clearer and easier for users to understand and control.