Humans are reading your ChatGPT conversations, and in some cases, they're seeing deeply personal details, according to a recent report from 404 Media. OpenAI's internal initiative, dubbed "Project Lily," employs hundreds of contractors to analyze real user prompts to improve the AI's responses and behavior, according to internal OpenAI documents seen by 404Media.
The report states OpenAI is outsourcing the work to hundreds of contractors tasked with evaluating ChatGPT's replies and rating them for accuracy and tone, with a focus on reducing sycophantic behavior that contributed to harmful outcomes in the past. While OpenAI strips usernames and attempts to redact personal information, sensitive data often still slips through.
According to 404 Media, this reveals a hidden human element behind the AI, one most users likely didn't expect, as many of these prompts being reviewed can include entire conversations between users and the chatbot, which in many cases can include private information given how often the 900 million active users rely on the service for a wide variety of answers.
The practice raises serious privacy concerns, especially for those who use ChatGPT as a confidant, therapist, or advisor for various aspects of their lives. The contractors themselves believe users are largely unaware of this process. "I don't think they would imagine some contractor somewhere [...] is analyzing the conversations," one reviewer told 404 Media.
Frequently Asked Questions
Open a question for an answer from TweakTown's coverage of this news, or ask your own below.
How does OpenAI redact personal information from ChatGPT chats before contractors review them?
How is OpenAI using contractor ratings to reduce ChatGPT's sycophantic behavior?
Are ChatGPT conversations reviewed in full, or only selected snippets tied to model errors?
How many contractors are involved in Project Lily, and where are they located?
Have a question that isn't listed here? Ask below, and TweakBot will answer it.
As 404 Media points out, while each of these AI models is becoming smarter and more sophisticated by ingesting large swaths of internet data, there is also the often-overlooked aspect of human-level training, which involves an actual person reading through chat conversations and critiquing the response the AI gave to our questions.
