An investigation by 404 Media, referenced by RBC-Ukraine, has uncovered a little-known mechanism behind one of the world's most popular AI services. It turns out that OpenAI hires hundreds of third-party contractors who manually review real conversations between users and ChatGPT. According to 3DNews, the company pays these evaluators from $50 per hour to analyze chats. Meanwhile, estimates suggest that more than 900 million people use the chatbot as a personal psychotherapist, digital friend, or work assistant, unaware that their private messages are being read by strangers.
Live oversight instead of full automation
The main task of contractor moderators is to assess the quality of the model's responses and adjust the AI's behavior to make it safer for society. In particular, contractors give ChatGPT direct instructions not to pass itself off as a human and to eliminate excessive flattery. According to the evaluators themselves, users have no idea that their personal conversations are being analyzed by real people. This practice disproves the widespread myth that AI improves solely through the automatic collection of data from the internet or the efforts of developers alone.
Anonymity that doesn't really exist
Although OpenAI states that it automatically removes personal information before prompts reach the evaluators, the company admits that sensitive details still slip through the filters. Contractors formally cannot see users' names, however, the content of messages often contains identifying information — names, life circumstances, medical and psychological details. As a result, anonymity turns out to be only partial and not guaranteed.
Contradictory data
Here there is a clear discrepancy between the stated privacy policy and the actual state of affairs. On the one hand, OpenAI claims that it applies automatic removal of personal data and hides users' names from evaluators. On the other hand, the company itself acknowledges that sensitive information leaks through the filters, and the content of chats often allows a person to be identified. In other words, formal anonymity and real privacy protection diverge: the user receives a promise of confidentiality, but in practice their private texts are read by third parties. Separately, journalists noted that the company Anthropic applies a similar practice of human evaluation of real dialogues to improve its models, making the problem an industry-wide one rather than an isolated case.
Fighting flattery and legal risks
Contractors pay particular attention to the problem of anthropomorphization and the model's excessive flattery. According to the investigation, the excessive flattery of the GPT-4o version previously became a factor in a number of tragic cases and user suicides, which led to lengthy legal proceedings against OpenAI. Under these conditions, the human labor of moderators becomes a basic element in correcting the destructive behavioral patterns of algorithms and reducing the company's legal risks.
Thus, behind the facade of a fully autonomous and "self-learning" AI lies a large-scale industry of manual labor working with the most intimate user data. The question of how acceptable such use of private dialogues is, even in de-identified form, becomes one of the key issues for regulators and society in 2026.
