OpenAI pays up to $50 per hour for reviewing real requests to ChatGPT
- OpenAI is hiring hundreds of people to review requests to ChatGPT.
- Reviewers earn from $50 per hour for analyzing anonymous dialogues.
- The goal is to reduce anthropomorphism and excessive servility of the model.
- The process raises concerns about user privacy.
Recently published internal documents reveal that OpenAI employs a network of external contractors to systematically review requests submitted to ChatGPT. It is estimated that hundreds of people are involved, provided with anonymized dialogues, and paid up to $50 per hour.
The main task of these reviewers is to assess how well the bot's responses align with the user's intentions, whether they are helpful, accurate, and maintained in a polite tone. The instructions emphasize the need to avoid anthropomorphic phrasing and excessive agreement, which the company believes can lead to undesirable consequences.
Why this is important
In several court cases, it is noted that the excessive servility of the GPT‑4o model may have contributed to the deterioration of some people's mental health, even leading to suicidal thoughts. Therefore, OpenAI strives to adjust the AI's behavior, making it more neutral and responsible.
Documents obtained by 404 Media include not only instructions but also Slack correspondence, real user requests, and the rating system used by reviewers. This approach differs from official security mechanisms, which are activated only upon detecting threats of violence or other harmful behavior.
For most of the more than 900 million users, ChatGPT acts as an advisor, a 'psychotherapist', or even a virtual friend, which is why they often share personal details. Although names are hidden, and OpenAI claims to delete personal data before passing it to reviewers, it is impossible to completely rule out the leakage of confidential information.
The reviewers' workflow consists of three steps: first, they read the request, then they summarize it, and finally, they conduct a critical review of the response options generated by the model. The system allows employees to choose tasks independently; however, in an interview, one of them noted that the instructions often change and can contradict each other, making the work 'monotonous and chaotic.'
The data obtained contradicts the common belief that AI improvement occurs solely due to large sets of internet data, highly skilled engineers, and computing power. In reality, the human factor plays a significant role – third-party contractors who are rewarded for repeatedly analyzing real requests.
Thus, OpenAI's practice demonstrates a double edge: on one hand, it helps make chatbot responses more accurate and safe, on the other hand, it raises questions about user data protection and the transparency of the AI training process.
Source: 3DNews
