Photo via 404 Media

Inside ‘Project Lily’: The Humans Reading Your ChatGPT Chats

Read the original at 404 Media ↗
Cite

Cox, J. (2026, September 14). Inside ‘Project Lily’: The Humans Reading Your ChatGPT Chats. 404 Media. https://www.404media.co/inside-project-lily-the-humans-reading-your-chatgpt-chats/

Dated when The Aporia collected the article, usually within hours of publication. Check the original for its exact date and byline.

404 Media · collected 2026-09-14 · by Joseph Cox
Content note This article describes suicide or self-harm.

Quick Summary

404 Media reports that OpenAI has hired hundreds of contractors to read through real user interactions with ChatGPT, including sensitive personal information, in order to improve the chatbot’s responses and address issues like anthropomorphization and sycophantic behavior. With over 900 million users potentially unaware that their chats may be reviewed by actual people, this practice raises significant privacy concerns, especially as some users share intimate details with ChatGPT. Contractors do not see usernames but can still access personal data before it is removed, highlighting an overlooked aspect of how AI models like ChatGPT are refined beyond internet scraping and engineering efforts alone.
Written by AI on 2026-09-14, from this article's own text only (the story summary below draws on all the coverage).

AI analysis

Story summary

404 Media reported that OpenAI, the company behind ChatGPT, has hired hundreds of contractors to review user prompts and conversations. These reviews aim to improve ChatGPT's responses, though they may include sensitive personal information from users who treat ChatGPT as a therapist or professional assistant. With over 900 million users, most are unaware that their chats might be read by actual people. Despite efforts to remove personal data before review, privacy concerns persist due to the potential for contractors to access intimate details shared with the chatbot.

Written for “Human Review Of Chat Gpt Chats” on 2026-11-21, grounded in this article and the 1 other(s) covering the same event.

Compared with similar articles

This article reads leans left and hedges 18% of its claims. Each row says how that neighbour differs.
Semafor
Hedges 1 of 32 claims Outlet record 64/100
“The articles discuss different topics related to AI and its regulation, with Article A focusing on human review of ChatGPT interactions for improvement purposes, while Article B discusses broader views on AI companies' stances towards regulation.”
The Straits Times
Leans left Hedges 17 of 70 claims Outlet record 64/100
“Article A reports on a specific hiring practice at OpenAI involving contractors reviewing ChatGPT prompts, while Article B discusses broader societal concerns and fears about AI control and security.”
404 Media
Hedges 1 of 9 claims Outlet record 95/100
New York Post
Leans left Hedges 3 of 4 claims Outlet record 66/100
“The articles discuss different aspects and concerns related to AI development by OpenAI, but do not describe the same specific incident.”
New York Post
Leans left Hedges 4 of 43 claims Outlet record 66/100
“Article A discusses the hiring of contractors by OpenAI to review ChatGPT prompts, while Article B focuses on the broader implications and privacy concerns related to users sharing personal information with AI chatbots like ChatGPT.”

Story

Human Review Of Chat Gpt Chats
Technology · 2 article(s) covering the same event. See how they differ ↑
How this article was measured signals, leaning charts, publisher record, byline, topics and claims

Signals How these are calculated →

Claims extracted
11
sentences that state something as happening or true
Uncertain
18%
2 of 11 hedged ("reportedly", "may", "according to")
Leaning
Leans left
of the writing, not the subject · beta estimate
Publisher signals
95/100
404 Media's corrections and hedging, across everything collected from it. Not a measure of accuracy.
Outlets on this story
2
Technology
Same framing elsewhere
1
articles that frame the event the way this one does
Analyzed 2026-09-14 · source text last changed 2026-09-14 · how these are computed

How this is being covered How these are calculated →

Article leaning vs. publisher reliability
Source leaning vs. consistency

Publisher

404 Media · 223 article(s) · 0 correction(s) detected
No corrections detected for this publisher. That may mean careful reporting, or simply that nothing has been checked.

Who wrote this

Joseph Cox
31 article(s) here · 1 carrying a prediction
It says AT&T analysts will transform the AT&T data “into a suitable format” while “ensuring accurate technical interpretation to accomplish HSI’s investigative mission.”
Nazarenko would sometimes sit for more than 12 hours at his computer, obsessively tinkering away, with a single, high-quality fake identity document taking up to a week of work.
2026-11-05 · assertive framing · The Confessions of an AI Fraud Kingpin
You might remember in 2024 we revealed the existence of OnlyFake, an underground site that let anyone generate realistic photos of fake IDs using ‘neural networks’.
2026-11-05 · assertive framing · An AI Fraud Kingpin’s Confessions
On Friday, Andy Yen, the CEO of Proton, cancelled a series of nearly company-wide ‘re-interviews’ that would have forced employees to explain how they are using AI, with the clear intention that they may be fired if not using it enough in the eyes of Proton leadership, according to six current Proton employees and audio of an around two hour all-hands meeting obtained by 404 Media.
2026-10-30 · assertive framing · Proton’s CEO Cancels AI ‘Re-Interviews’
After the break, Sam tells us why former Yahoo CEO Marrisa Mayer can’t keep pretending ‘gender doesn’t matter.’
2026-10-29 · assertive framing · Podcast: Flock CEO Says the Backlash Isn’t Real
In the last few months, Yen set up an email address for employees to contact if they come across anything blocking adoption of AI at Proton, so leadership can “ruthlessly smash everything that stands in the way of AI velocity;” Proton installed software on workers’ computers that track how much AI code the company is generating; and Yen said this is the last year that Proton will hire engineers at all, and instead only hire people to build things with AI.
A letter written by Telegram, and providing a specific user’s phone number, user ID, and IP address, reads: Dear Sir or Madam, For this particular case, the following information was available: User #7149269705 : +201155664529 [Mahmoud Almohtadi] IP 184.186.72.239, 8 Oct 2025 2:08:12 We hope this information will be helpful in your investigation. Kind regards, The Telegram Team The letter says Telegram provided the data “voluntarily” and that by providing it, Telegram is “not conceding that we are subject to U.S. jurisdiction for this or any other request or any purpose.”
This is pretty normal in the advertising world, but I seriously doubt that when most of ChatGPT’s many, many users see an ad, they realize that might be because someone, somewhere, uploaded their phone number into OpenAI’s systems.
2026-10-15 · assertive framing · ChatGPT Ads Target Your Phone Number, Email
Multiple Democrat, Republican, and Independent lawmakers have introduced new pieces of legislation that, if passed, would dramatically curb federal agencies’ access to Flock cameras, with the lawmakers referencing reporting by 404 Media in their explanations for why the legislation is needed.
Although any potential damage will be less if ShinyHunters doesn’t publish the data publicly, the theft happening at all still presents much of those same national security risks, with the data providing granular insight into how the FBI operates. “Since the very beginning we had made our decision that we would never publish this data.
Also by Joseph Cox
An AI Fraud Kingpin’s Confessions
2026-11-05 · 404 Media
Nothing else under this byline is closely related to this article, so these are simply their most recent.
All 31 articles by Joseph Cox →

Topics

404 Media Anthropic ChatGPT OpenAI

Subjects

OpenAI ORG · 4× 404 Media ORG · 3× Anthropic ORG · 1× ChatGPT ORG · 1×

Narrative

The news presents a major privacy risk for ChatGPT’s users, with people often using ChatGPT as a therapist, professional assistant, or digital friend, and providing it with all sorts of intimate details about their lives.
framing: assertive · carried by 1 article(s) · first seen 2026-09-14
The prompts these people review can include whole conversations between users and the chatbot, conversations that most of ChatGPT’s more than 900 million users probably don’t realize may be read by actual people.
2026-09-14 · 404 Media
Inside ‘Project Lily’: The Humans Reading Your ChatGPT Chats · assertive framing

Claims (11 extracted, 2 hedged)

OpenAI is hiring hundreds of contractors who read a massive stream of real users’ ChatGPT prompts, with the prompts sometimes including sensitive personal information, 404 Media has learned. asserted
Media → hire → information
The prompts these people review can include whole conversations between users and the chatbot, conversations that most of ChatGPT’s more than 900 million users probably don’t realize may be read by actual people. uncertain
most → review → people
The goal of these prompt review teams is to improve the responses ChatGPT gives to its users, with the contractors rating and critiquing the chatbot’s generated replies. asserted
contractors → improve → replies
Internal documents seen by 404 Media show contractors training ChatGPT to not anthropomorphize itself, and to be less sycophantic, a key problem for OpenAI whose over-sycophantic 4o model led in part to multiple peoples’ suicides, according to various lawsuits. uncertain
model → see → lawsuits
The news presents a major privacy risk for ChatGPT’s users, with people often using ChatGPT as a therapist, professional assistant, or digital friend, and providing it with all sorts of intimate details about their lives. asserted
people → present → lives
The contractors don’t see ChatGPT usernames, and OpenAI says it tries to remove personal information before prompts reach the reviewers, but the company acknowledged sensitive details can still get through. asserted
details → see → reviewers
The news also dispels the misconception that these models are improving only because of OpenAI’s mass scraping of the internet, the talent of its well-paid engineering and AI teams, or the power of its newer models. asserted
models → dispel → models
An important and overlooked part are the outside contractors paid to read and review ChatGPT responses to real prompts over and over again. asserted
part → overlook → prompts
Anthropic confirmed to 404 Media it is also using human review to improve its models. asserted
it → confirm → models
“No,” someone who works with the prompts said when asked if they think ChatGPT users know that humans are reading their chats. asserted
humans → work → chats
“I don’t think they would imagine some contractor somewhere [...] is analyzing the conversations.” asserted
contractor → think → conversations
Give feedback · no sign-in needed

No sign-in needed for now. Your note is sent without a name, so it cannot be answered.

History Support