OpenAI may be sending entire chat logs of ChatGPT conversations to contractors to help improve its models' responses. According to documents seen by 404Media, OpenAI uses contractor feedback extensively to help train its models, and logs obtained make it clear that most users aren't aware that the data is being shared with anyone.
Annotated data and A/B testing of chatbot responses have served as training systems for major AI developers for years, but it's never been entirely clear how much access they have to private chat logs. You can restrict access in settings, and OpenAI claims it puts all logs through a "privacy filter" before sharing anything with training contractors. But 404Media observed deeply personal information, even if users asked ChatGPT to keep all that information private.
Contractor training documents encourage them to escalate anything they come across "with potential safety concerns," or personal information. However, OpenAI has the "Improve the model for everyone" setting enabled by default on the Free, Plus, and Pro plans, so it's something you need to opt out of, and it isn't applied retroactively. Even when you do, older conversations can still be used for training and may well be sent to contractors for review.
OpenAI also doesn't make it clear that humans may review your conversations when it's enabled or if you re-enable it. Since being contacted by 404, however, it has updated its help page to better explain how to opt-out of model training. It also highlights that using a temporary chat will ensure that the log won't appear in your chat history, won't be used to create memories, and won't be used in model training.
Contractors that 404Media interviewed said that most of the time, the prompts they reviewed were very "rote" and kind of boring. Sometimes, however, they did contain "amusing" personal information. They also described the work as being "all over the place," with regularly changing guidelines that, in some cases, were even contradictory to one another or to previous guidelines.
They aren't the kind of underpaid data annotators; 404Media reports the contractors are paid more than $50 an hour, though some analysts suggest that may change in due time.
How to Opt Out of ChatGPT Training #
On Web (Signed-in):
- Click your profile icon
- Select Settings
- Go to Data Controls
- Turn off "Improve the model for everyone"
On Web (Signed-out):
- Click the ? icon in the bottom-right corner
- Select Settings
- Turn off "Improve the model for everyone"
On Mobile:
- Open the side-bar menu
- Tap your profile icon
- Select Data Controls
- Turn off "Improve the model for everyone"
A better way may be to just keep any sensitive chats off cloud-based AI in general. Certainly, make sure that within the individual prompts, you don't self-identity.
Disclosure: Ziff Davis, PCMag's parent company, filed a lawsuit against OpenAI in April 2025, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.