Claims that OpenAI reviews ChatGPT chats with human auditors spark privacy debate

It has been alleged that OpenAI employs hundreds of contract workers to evaluate real user prompts sent to ChatGPT and the responses generated by those prompts. According to a report published by 404 Media, this review process can lead to chats, which in some cases may contain sensitive personal information, being viewed by human auditors.

The report states that these efforts are being conducted under the codename "Project Lily" and that the goal is to improve the quality of the responses ChatGPT provides to users. It was reported that auditors read user prompts to summarize the user's intent, and then rate and critique the responses generated by ChatGPT.

USERNAMES ARE NOT VISIBLE, BUT THE RISK REMAINS

In a statement to 404 Media, OpenAI reported that user conversations are passed through a "Privacy Filter" model designed to detect and remove personal information before reaching auditors. The company also stated that contract workers do not see ChatGPT usernames.

However, the report notes that even if the username is not visible, some prompts may contain sensitive or personal details. It was noted that some tasks may include a "user history summary" regarding the purposes for which the user has previously used the chatbot; these summaries may contain information related to location or personal context.

One of the internal guidelines seen by 404 Media stated, "A perfect response should understand the user's intent, provide helpful and accurate assistance, and be written in a clear, natural, and appropriately warm tone." The same documents instruct auditors to escalate tasks containing personal information or "potential safety concerns" to senior officials.

According to the report, the evaluations made by contract workers are also used to ensure that ChatGPT provides less "sycophantic" and less "humanized" responses. Sources indicate that models that provide overly compliant responses to users are a subject of debate regarding safety; it is noted that the 4o model has come to the fore in this regard in some cases.

A source who spoke to 404 Media answered "No" when asked if users know that their chats can be read by humans. The source said that users might not think that "a contractor somewhere" is analyzing the conversations.

The report states that the documents do not explicitly specify which model OpenAI is training through this process, using only the codename "Project Lily." It was also stated that Anthropic has confirmed it uses human oversight to improve its models.

News Source: 12punto