Remember when you were chatting with ChatGPT under the comfortable assumption that the other participant was, you know, not human?
Funny thing about that. A new investigation from 404 Media reports that OpenAI has hired hundreds of human contractors to review real ChatGPT conversations as part of an internal effort known as Project Lily. The reviewers reportedly read user prompts and, in some cases, entire conversations while grading ChatGPT’s answers and helping OpenAI make the bot less inaccurate, robotic and relentlessly agreeable.
Humans training AI? Not much of a surprise.
The slightly more eyebrow-raising part is whose conversations they’re using.
According to 404 Media, the chats are put through privacy filtering before reviewers see them, and usernames aren’t shown. Unfortunately, OpenAI acknowledged that the filtering isn’t perfect. Reviewers can still encounter sensitive personal information contained in conversations.
Which is a mildly important detail considering people routinely tell ChatGPT about their relationships, jobs, finances, health problems, legal situations and that deeply embarrassing thing they absolutely do not want Steve from accounting knowing about.
OpenAI isn’t exactly hiding the fact that consumer conversations can be used to improve its models. Its current policy says ChatGPT conversations may be used for training unless users opt out. What has received considerably less attention is that “training” can involve an actual human being reviewing the material.
For Canadians, there’s another uncomfortable footnote. Earlier this year, federal and provincial privacy regulators concluded that OpenAI’s initial handling of personal information raised problems involving over-collection, consent and transparency, although regulators also said OpenAI subsequently made changes to improve its practices.
Want out?
Go to:
Settings → Data Controls → “Improve the model for everyone” → OFF
OpenAI says new conversations will then no longer be used to train its models. Temporary Chat conversations also aren’t used for training.
And a reminder:
Don’t type something into an AI chatbot that would ruin your afternoon if you discovered another human had read it.
