Skip to content

ChatGPT's human editors

Endless Thread

Hosted by Amory Sivertson and Ben Brock Johnson

25m
Transcript

Original podcast

Where to listen

Summary

Written by a model from the transcript; not yet reviewed.

Producer Grace Tatter leads co-hosts Ben Brock Johnson and Amory Sivertson through an exercise of reading and grading AI-generated marketing copy, revealing a hidden industry where human contractors evaluate ChatGPT outputs. Journalist Joseph Cox explains Project Lily, OpenAI's internal program using human reviewers to make models sound more conversational and less sycophantic. The discussion turns to privacy concerns as the hosts realize that millions of users sharing intimate personal details and relationship problems with AI may not know human eyes could review their prompts.

  • Human contractors are hired to read and grade real ChatGPT conversations to make the AI sound more human. 4:22
  • Reviewers are instructed to lower scores for responses containing too many emojis or robotic language. 13:08
  • OpenAI markets ChatGPT as a confidential sounding board, leading users to share deeply personal details. 19:17
  • Users should treat LLM chats less like encrypted messaging and more like semi-public social media posts. 24:13

Chapters

  1. 0:00
    Podcast Intro and Ad

    Hosts introduce the episode and a brief ad for a related podcast.

  2. 2:01
    ChatGPT Prompt Assignment

    Grace assigns Ben and Amory to review a ChatGPT-generated LinkedIn post.

  3. 3:20
    Grading the ChatGPT Response

    The hosts discuss how to rate the AI’s LinkedIn post and critique its tone.

  4. 7:49
    Project Lily Overview

    Grace explains OpenAI’s Project Lily, where contractors review real ChatGPT conversations to improve the model.

  5. 9:23
    Human Review Process

    Details of how reviewers analyze prompts, rate responses, and flag issues like emojis and sycophancy.

  6. 20:30
    Industry Context and Comparisons

    Hosts compare OpenAI’s practices with other LLMs like Google Gemini and Anthropic’s Claude.

  7. 22:58
    Privacy and Personal Data Concerns

    Discussion of how personal information is handled, de‑identification, and the risks of re‑identification.

  8. 23:59
    User Awareness and Behavior

    Exploration of how users perceive AI review, the need for privacy norms, and potential behavioral changes.

  9. 27:48
    Wrap‑Up and Credits

    Final remarks, thanks to guests, and episode credits and promotion details.

Highlights

  • 3:19

    Grading AI Marketing Copyinsight

    The hosts review a ChatGPT-generated LinkedIn post and grade it three and four out of seven, complaining that it lacks human spunk or voice.

  • 4:59

    Contractor Pay Ratesdata

    The hosts discuss how recruitment ads for training AI models offer between fifty and one hundred dollars an hour, providing a lucrative option for former journalists.

  • 7:52

    Uncovering Project Lilyinsight

    Joseph Cox details Project Lily, an internal OpenAI program employing contractors to read and analyze real user prompts to improve model quality.

  • 12:58

    Reviewer Guidelines and Flawsstrategy

    Contractors are told to flag excessive emoji usage, robotic phrasing, and sycophantic behavior where ChatGPT anthropomorphizes itself.

  • 19:24

    Weekly User Statisticsdata

    ChatGPT reaches between 900 million and one billion weekly users, many of whom use it for deeply personal issues or as a therapist.

  • 22:16

    The Bill Weld Experimentstory

    A 1997 study proved that a zip code, gender, and birthday are all it takes to de-anonymize individuals from data sets.

Quotes

  • “And I just feel unsettled by all of it.”
    Amory Sivertson · 10:56
  • “We should accept that reality, but not without care.”
    Ben Brock Johnson · 12:24
  • “And these human reviewers are checking to see if Chat GPT is anthropomorphizing itself.”
    Unnamed speaker · 15:05
  • “So hard to answer that question.”
    Ben Brock Johnson · 21:30
  • Is Business Broken 0:00–0:19
  • LinkedIn 0:36–0:48
  • Lululemon 0:53–1:18
  • Is Business Broken 28:30–28:41
  • Lululemon 29:56–30:20

About this episode

Up to a billion people use ChatGPT every week, for everything from the mundane to the weird to the deeply personal. It is commonly assumed that these messages are private – between the user and the LLM. It is commonly assumed that these messages are private – between the user and the LLM. But in some cases, human reviewers are reading these conversations and offering editorial feedback to help ChatGPT sound more human. Reporter Joseph Cox of 404 Media broke the story of Project Lily, Open AI's initiative to use human reviewers to train and improve ChatGPT. He breaks it down for us on this week's Endless Thread. Show notes: Inside ‘Project Lily’: The Humans Reading Your ChatGPT Chats (404 Media https://www.404media.co/inside-project-lily-the-humans-reading-your-chatgpt-chats/) This episode was produced by Grace Tatter and Frannie Monahan. It was co-hosted by Grace Tatter, Ben Brock Johnson, and Amory Sivertson, and edited by Dave Shaw. Mix and sound design by Paul Vaitkus.

Transcript

0:00
25:29

4,591 words · 8 speakers

Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.

Topics

  • OpenAI
  • ChatGPT
  • AI
  • Endless Thread
  • Project Lily
  • WBUR
  • Crossing Hurdles
  • Merckor
  • Bose
  • Project Billy

Page updated · transcript and summary are machine-generated