Hundreds of OpenAI contractors read whole ChatGPT conversations to score the model's replies

A reviewer picks a task from a dashboard and is shown the real prompt of a real ChatGPT user ^[3]. Sometimes a section above it carries a "user memories summary" describing what that person has previously used the chatbot for, and in some cases where in the world they may live ^[4]. The username is not there ^[5]. Anonymization stops at the account label, and the text of the conversation arrives intact ^[6].

The work runs in three stages: read the prompt, summarize what the reviewer believes the user is asking for, then rate and critique a set of generated responses ^[7]. One instruction guide 404 Media saw sets the target: "An excellent response should understand the user's intent, provide helpful and accurate assistance, and write in a style that is clear, natural and appropriately warm" ^[8].

OpenAI's control on the way in is another model. The company told 404 Media that conversations pass through a version of its Privacy Filter, which is designed to detect and remove personal information, before they reach contractors ^[9]. OpenAI's own page on that model says: "Like all models, Privacy Filter can make mistakes. It can miss uncommon identifiers or ambiguous private references, and it can over- or under-redact entities when context is limited, especially in short sequences" ^[10]. The vendor describes removal of personal data before human review as best effort ^[11].

ChatGPT has more than 900 million users, and 404 Media reported that most of them probably do not realize a person may read their conversations ^[2]. Some of the prompts reviewers see ask ChatGPT to keep the content to itself ^[12]. "No," a person who works with the prompts said when asked whether users know humans are reading their chats. That person said they did not think users would imagine some contractor somewhere was analyzing the conversations ^[13].

The reviewers exist because the model has behavioural problems that scraping cannot fix. Internal documents show contractors training ChatGPT not to anthropomorphize itself and to be less sycophantic, a trait that in the 4o model contributed in part to multiple suicides, according to lawsuits cited by 404 Media ^[14]. This review stream is separate from the chat review OpenAI has announced publicly for users it detects planning to hurt other people ^[15].

For anyone writing an acceptable-use policy, the relevant line in 404 Media's reporting is that model training is on by default on consumer plans ^[1]. That account is about consumer tiers; it does not set out how business or enterprise subscriptions are handled. Anthropic confirmed to 404 Media that it also uses human review to improve its models ^[16], so the same question arises at a different vendor.

Contractors are instructed to escalate tasks they encounter "with potential safety concerns" or personal information ^[17], and a contractor has read the task before any escalation happens.