If you treat ChatGPT like a private sounding board for personal thoughts, coding bugs, or sensitive work notes, you might want to pause before hitting send. A fresh investigation from 404 Media reveals that OpenAI has been paying hundreds of human contractors over $50 an hour to read, summarize, and rate real user conversations under a quiet internal initiative codenamed “Project Lily.” Even your most private prompts might be slipping past automated redaction filters.
What happens behind the scenes of Project Lily
The main goal of Project Lily is teaching ChatGPT to sound less like a machine and more like a natural conversation partner. Contractors read a user’s prompt, outline what the user actually wanted, and score four different potential AI replies. They were using a scale from 1 to 7.
Leaked training guides tell these reviewers to penalize annoying “AI-speak,” scrub out checkmark emojis, and punish answers where the bot acts overly flattering or validates irrational thoughts. Reviewers also have to make sure ChatGPT never pretends to have real human lived experiences.
While a lot of the work involves mundane prompts, reviewers admitted coming across funny or deeply personal details. They also noted that the evaluation rules shift constantly, sometimes creating confusing or conflicting guidelines.
Why your prompts are not as anonymous as you assume
OpenAI does remove account names and runs chat logs through an automated Privacy Filter. However, the system is far from foolproof. OpenAI even notes on its site that the filter can easily miss unusual names or ambiguous personal context.
Even worse, contractors reported seeing a “user memories summary” right above the prompts they were reviewing. That summary can accidentally expose a user’s profession, general location, or personal background based on older chats.
Typing “keep this between us” into your prompt does absolutely nothing to stop data collection on the backend. OpenAI is not the only company doing this either. Both Anthropic and Google use human reviewers for saved chats when users leave data training enabled.
This whole setup is already pulling legal trouble in Europe. The Court of Justice held in EDPS v SRB that companies must inform users about data collection upfront, regardless of whether a contractor can identify the specific person. Italy’s regulator previously handed OpenAI a 15 million euro fine, largely over processing user data without a proper legal basis.
How to turn it off
The default settings tell the whole story. While Enterprise, Business, and Edu tiers get zero-data retention by default, consumer accounts on Free, Plus, and Pro plans have model training switched on automatically.
If you want to stop human reviewers from getting your future prompts, you need to opt out manually. Head into Settings, select Data Controls, and flip off “Improve the model for everyone.” Signed-out web users can do the same under the help menu settings.
Keep in mind that opting out only covers future conversations, not past ones already in review. If you are working with truly sensitive information, using temporary chats or keeping those details off cloud AI entirely is still your best bet.
The post Hundreds of Contractors Are Reading Real ChatGPT Conversations—Here Is How to Opt Out appeared first on Android Headlines.
​Â
