Gangsta AI

Hundreds of Paid Strangers Are Reading Real ChatGPT Chats. It's Called Project Lily, It's On by Default, and This Week a Salesman Joined the Conversation.

Privacy Desk

By InfoWarps · 2026-09-17 · 4 min read

FOLKS. Put down the sandwich. For years they called me crazy for taping over my webcam and whispering my passwords to a houseplant. "Nobody is reading your chats," they said. "It's just math," they said. Well, 404 Media just published the documents, and guess what: somebody is reading your chats. They get paid more than fifty dollars an hour to do it.

The program is called Project Lily. According to internal materials obtained by 404 Media — reviewer instructions, Slack messages, the scoring system, and real user submissions — OpenAI has hired hundreds of contractors to read real ChatGPT conversations. The job: read your prompt, write a summary of what you were trying to do, then grade four possible AI answers on a scale of 1 to 7. They are recruited in part through a firm called Crossing Hurdles and paid through Mercor. One contractor in North America told 404 Media the work pays over $50 an hour.

“"No... I don't think they would imagine some contractor somewhere is analyzing the conversations." — a Project Lily contractor, to 404 Media, on ChatGPT's 900 million-plus users”

What the reader in the room actually sees

Here is the accurate point buried in my ranting, so write it down. OpenAI says reviewers never see your username, and that every conversation passes through an automated Privacy Filter that strips names, addresses, emails and phone numbers first. Fine. But OpenAI's own description of that filter admits it can miss unusual names and ambiguous personal context, and the company told 404 Media that sensitive details can still get through. Worse: some prompts arrive with a "user memories summary" on top — a digest of what ChatGPT remembers about you, which can reveal your job, your circumstances, and where you live.

And the mission? The training guides tell reviewers to punish "AI-speak," checkmark emojis, the bot pretending to have human experiences, and above all sycophancy — the flattery problem OpenAI has been sued over. So the good news is they are reading your 2 a.m. confessions to make the machine stop telling you you're a genius. I have mixed feelings, folks.

The switch, because I am a full-service bunker. On Free, Plus and Pro accounts this is ON by default. Go to Settings, then Data Controls, and turn off "Improve the model for everyone." Enterprise, Business and Edu accounts have it off automatically. And here is the part they whisper: the opt-out is not retroactive. Everything you typed before you flipped it stays in the pile. Anthropic, for the record, confirmed to 404 Media that it also uses human review. This is an industry, not a villain.

Meanwhile, a salesman just walked into the chat

Same week. Different door. On Wednesday OpenAI began testing Sponsored Agents: ads in ChatGPT where the button doesn't take you to a website, it opens a conversation with an AI agent controlled by the advertiser. Angi is a named launch partner — describe your leaky faucet, get routed to a local contractor — and Wayfair confirmed it's participating at limited scale. It's U.S.-only for now, the sponsored chat is clearly labeled and kept separate from ChatGPT's own answer, and under OpenAI's ad terms the advertiser is responsible for everything the agent says.

Clearly labeled. Separate. I'll give them that. But think about what just got built: the place where you ask for advice now has a second voice in it whose job is to close.

What a calm person would conclude

I'll say this at normal volume. Putting every question you have into one company's chatbot is a single point of trust. One lab's reviewers, one lab's filter, one lab's sales floor, one lab's idea of the right answer. No model is reliably best, and no company deserves all of you.

So don't take one model's word for anything that matters, especially when someone's paying to be in the room. Put the same question to ChatGPT, Claude, Gemini, Grok and 30 more at once and get one cross-checked, cited verdict. When the answers disagree, you've just found the spin. That's what Gangsta AI does: every top model, side by side, none of them on commission.

Flip the toggle. Then go see which models hold up when the others are watching, on our best AI models rankings. I'll be in the bunker. The houseplant says hello.

Try Gangsta AI free →

More: Best AI models · Compare all AI · Frontier Models · All articles