Do Son 
According to a comprehensive report published by 404 Media, a secretive initiative dubbed Project Lily is actively operating within OpenAI. To relentlessly refine the quality of its responses, OpenAI employs hundreds of outsourced prompt reviewers. These individuals are explicitly tasked with manually reading authentic conversations taking place between real users and ChatGPT.
Privacy Filters Remain Imperfect
Under normal operational circumstances, these prompt reviewers cannot view actual usernames. The dialogue initially passes through a designated privacy filter to systematically eradicate personally identifiable information. However, OpenAI candidly admits that this automated filtering mechanism might occasionally overlook sensitive personal details. Consequently, short conversations are particularly susceptible to such alarming privacy lapses.
Furthermore, the review interface typically includes a persistent user memory summary. This summary efficiently encapsulates the user’s historical inquiries, established interests, and situational context. Occasionally, reviewers can even deduce the user’s geographical location entirely from these memory fragments. Thus, the reviewers frequently observe the complete conversation complete with profound contextual depth.
Reviewers Grade Style, Not Facts
Participants frequently describe this labor as intensely mechanical and repetitive, despite a reported hourly compensation exceeding fifty dollars. The primary objective of the reviewer revolves around meticulously grading the stylistic quality of the response, rather than rigorously verifying its factual accuracy. For instance, reviewers must determine whether the model’s response remains topically relevant, avoids an overt “AI tone,” abstains from pedagogical lecturing, eschews the abuse of emoticons, avoids obsequious flattery, and refrains from portraying the model as an entity possessing genuine life experiences.
Crucially, determining whether the model’s response harbors factual inaccuracies falls entirely outside the reviewer’s purview. Ultimately, reviewers are not encyclopedic entities capable of comprehending the absolute truth of every conceivable event. Therefore, it remains entirely unsurprising when users encounter model responses riddled with significant factual errors.
Default Data Sharing Drives the Initiative
By default, the free, Plus, and Pro tiers of ChatGPT all activate the feature permitting the use of chat data to improve product functionality. Once activated, this crucial setting explicitly authorizes OpenAI to harness the user’s conversational content for continuous product enhancement. The examination of dialogue by outsourced reviewers within Project Lily merely represents one specific application of this vast data reservoir. Naturally, the company utilizes this invaluable data for myriad other purposes, including the ongoing, relentless training of its models.
Countless users treat ChatGPT as an ephemeral confidant, openly discussing familial issues, intimate health concerns, and profoundly private matters. Some reviewers have starkly observed that many users likely remain blissfully unaware that their deeply personal conversations might be scrutinized by actual human beings. Contemplating this reality invariably evokes a profound sense of horror.
Support Our Threat Intelligence
Find our threat intelligence and malware analysis helpful? Support our work today and unlock a 100% ad-free reading experience!
Buy Me a Coffee
