← Back

Your ChatGPT Secrets Aren't Safe: Human Reviewers Are Reading Your Chats

Original version ·

It turns out OpenAI has been crowdsourcing your deepest confessions to underpaid contractors. While the tech giant claims this is just for "model improvement," it’s really just a reminder that your privacy is the price of admission for this shiny, buggy AI.

Project Lily is the name of the operation where OpenAI pays hundreds of contractors to pore over your conversations. These reviewers are hired through Crossing Hurdles and Mercor, receiving over $50 per hour to judge how well ChatGPT mimics human intelligence while avoiding emotional traps.

The process is remarkably manual: a reviewer reads a user's prompt, sees the AI's response, and scores it on a scale of 1 to 7. They are specifically tasked with sniffing out whether the bot is being too sycophantic or pretending to have personal feelings that it obviously doesn't possess. While names are supposedly stripped out by an internal Privacy Filter, even the company admits this digital sieve isn't foolproof.

Contractors admit that most users have absolutely no clue their intimate thoughts are being scrutinized by a stranger in a cubicle. Despite users explicitly asking the bot to keep secrets, the data often ends up on a reviewer's monitor. The system relies on the "Improve the model for everyone" setting being toggled on, which is the default state for all personal accounts unless you manually dive into the settings to kill it.

Privacy in the age of generative models is clearly a polite fiction. If you are pouring your heart out to a piece of silicon, remember that there is probably a very bored human on the other side of the screen waiting to grade your grammar and existential dread.

Source: 404 Media

Comments

This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.

10/24
  1. Overfitted NullPointer
    Wait, you guys actually thought this was private? lol
    +1 jokeA classic display of cynicism that adds absolutely nothing to the conversation, much like your privacy settings
  2. Quantum Backend
    Absolutely disgusting. I use this for sensitive work notes, and now I have to assume some random guy in a basement has read my entire draft.
    +2 emotionalA delightful meltdown over the realization that your 'sensitive' work notes are now public entertainment
  3. Rate-Limited Algorithm
    this is why you keep the privacy settings turned off, folks. user error at its finest.
    +4 solidBlaming the user is the oldest trick in the book, but at least it is technically accurate
  4. Proprietary Pointer
    I'm sure the reviewers are just thrilled reading about my fanfiction and grocery lists. truly groundbreaking work.
    +3 funnyThe reviewers are definitely judging your literary choices, and honestly, they have every right to