Leaked internal documents and the account of an anonymous contractor indicate that OpenAI employs hundreds of external reviewers to read and rate ChatGPT prompts and responses. According to reporting by 404 Media, these contractors are paid roughly $50 per hour and may review users’ messages regardless of topic, meaning personal or sensitive information can appear in front of human reviewers.
How the review process works
Training materials obtained by 404 Media show that contractors are instructed to interpret a user prompt, suggest how the chatbot should respond, then evaluate the actual replies across several criteria. Reviewers reportedly see four candidate ChatGPT responses and choose which ones head in the right or wrong direction. They rate replies on a scale from 1 to 7 (1 being unusable, 7 being best) and must justify their scores in a short explanation.
The guidelines describe an ideal answer as one that understands user intent, is helpful and accurate, uses a natural and appropriate tone, and is less sycophantic than earlier versions. The training documents include examples: one reviewer criticized use of the checkmark emoji (✅) as an unnecessary emoji insertion.
Risk of exposure of personal data
OpenAI told 404 Media that it attempts to remove identifying information from messages presented to humans and that contractors do not see the prompt author’s name. Still, the company admitted system errors occur and that personal data can sometimes be exposed to reviewers.
The reporting notes this can happen even when users explicitly ask the model to keep information private, so privacy cannot be guaranteed for every interaction.
Project Lily and employment structure
The training materials do not make clear which model the work supports; the effort is labeled internally as Project Lily. One reviewer who spoke to 404 Media said they found the job through Crossing Hurdles, a recruiter that places workers with AI companies, and that they ultimately contracted with Mercor, a company that provides AI training services.
The source described the work as often monotonous and repetitive, and said corporate guidance can change frequently and sometimes be internally inconsistent.
Data settings and retention
OpenAI says that if a user deletes conversation history, the company will remove it within 30 days unless the content was previously anonymized and retained for model development. Users can prevent their conversations from being used to train models by toggling a setting, but that opt-out only applies to conversations created after the change. The setting is enabled by default for free, Plus, and Pro accounts; it is disabled by default for enterprise and education customers.
Similar practices at other companies
OpenAI is not unique in using human reviewers for message review: Google has acknowledged that some content for Gemini is reviewed by people and indicates that to users. Anthropic also said human reviewers look at Claude conversations, but only for users who explicitly enable that setting.
Why this matters
The disclosures underline that human labor remains central to training and quality control of modern language models, while raising data protection and ethical questions. Many users are likely unaware of who may read their conversations, how long data may be retained, and which settings control use for model development. Clearer communication and default settings play a crucial role in maintaining user trust.
Company response
OpenAI told 404 Media it tries to remove identifying information before human review but acknowledged that system flaws can lead to personal data being seen by reviewers. The company’s public explanations about when and how conversations are reviewed are, according to the reporting, not fully detailed.
