The work is described in instruction guides, Slack messages and reviewer materials under an internal initiative known as Project Lily. Contractors are shown user prompts and, in some cases, complete exchanges with ChatGPT, then asked to summarise the user’s intent, assess several candidate answers and score their quality.
The leaked material indicates that reviewers are recruited through outside staffing and work platforms, including Crossing Hurdles and Mercor. At least one contractor said the work paid more than $50 an hour. The documents do not establish that every reviewer receives the same rate, and OpenAI has not publicly detailed the compensation structure.
OpenAI says it removes account names and applies an automated privacy filter before material reaches human reviewers. The company has acknowledged, however, that sensitive information can still pass through those safeguards. Some review tasks may also include contextual information drawn from a user’s previous interactions, increasing concern about how much personal detail can be visible even when direct identifiers are stripped.
The purpose of the review programme is model improvement rather than customer support. Contractors are instructed to judge whether answers are accurate, appropriately sourced and aligned with behavioural goals, including avoiding excessive flattery, unnecessary anthropomorphism and formulaic language. Their evaluations can then be used as training or feedback data for later model versions.
OpenAI’s published policies state that content from consumer services such as ChatGPT may be used to improve and train its models unless users opt out. The company allows signed-in users to disable “Improve the model for everyone” under Data Controls. Once switched off, new conversations remain in chat history but are not used for model training.
That distinction is important because opting out does not necessarily erase data that may already have been collected or processed for model improvement. OpenAI’s privacy materials advise users not to share information they would not want used or reviewed, and its consumer privacy guidance says conversations can help improve model quality only when users permit that use under available controls.
Temporary Chat offers a separate privacy setting. OpenAI says those conversations do not appear in chat history, do not create memories and are not used to train its models. Temporary chats are retained for 30 days for safety purposes before deletion, subject to legal or security requirements.
Human access is not limited to training workflows. OpenAI’s documentation for ChatGPT agent says a limited number of authorised personnel and trusted service providers may access user content for abuse investigations, security incidents, support, legal matters or model improvement when the user has not opted out. The company says such access is restricted on a need-to-know basis and is logged.
Business, Enterprise and Edu customers receive different treatment. OpenAI says data from those offerings is not used to train its models by default. Its privacy portal also distinguishes consumer services from business products, where contractual terms govern data handling.
The new disclosures focus attention on a longstanding feature of artificial-intelligence development: human feedback remains a significant part of improving model behaviour. Contractors across the sector commonly label data, compare responses and identify errors that automated systems may miss.
Anthropic has also confirmed that it uses human reviewers to improve its models, though individual companies differ in how they select data, anonymise it and allow users to opt out. The broader practice is increasingly sensitive as chatbots are used for health questions, legal issues, workplace communications and personal matters.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
