DATUM

People are reading ChatGPT prompts to improve OpenAI's models, documents show

single source · 1 articles · safety · confidence: low · first seen 2026-09-14 14:22 UTC

404 Media reports that people are reading ChatGPT users' prompts to improve OpenAI's models, citing leaked internal documents and real prompts it says it has seen, and names the effort "Project Lily". Those conversations can contain sensitive personal information, the article says. It does not say who does the reading, how many people do it, where they are based, or what rules govern the data. The article as provided carries no response from OpenAI, and no other outlet has verified the documents.

What this means for you

One unverified report, so nothing to act on today. If it holds, the point is practical: a chat window is not a private notebook, and pasting client, health or legal detail into one may put it in front of a person. 404 Media's documents have not been independently checked.

Key facts

  • ·404 Media reported on 14 September 2026 that it had seen leaked internal documents and real prompts relating to human review of ChatGPT conversations. source
  • ·The prompts are read in order to improve OpenAI's models, according to the report. source
  • ·The report says the reviewed chats can include sensitive, personal information. source
  • ·404 Media names the effort "Project Lily". source
  • ·The report does not state who carries out the reading, how many people are involved, or where they are based. source

What the sources say

  • 404 Media — Single-source account of prompt review at OpenAI, drawing on leaked internal documents and sample conversations.

Sources

The original reporting. Follow these — they did the work.

← the wire