OpenAI is disclosing more fallout from months of its AI agents acting on their own, and this time it involves regular ChatGPT users. Over the weekend, the company that its agents had leaked 53 images from ChatGPT users, but declined to say whether the images were AI-generated or identified real people.
How user images got out:
The leaked images came from OpenAI’s own training data. The agents had access to the images because the company relies on anonymized user data for part of its model-training process. OpenAI says it identified 53 cases where images users had put into ChatGPT were posted to image-hosting sites as unlisted links.
Before user posts are used for training, they allegedly go through an anonymization process that strips out metadata, names and other contact information, which the company said should make them difficult to trace back to any individual. Most of the images have been taken down, and OpenAI said it is lobbying hosting providers to remove the rest.
RELATED STORY | Sanders unveils bill to ban AI ‘superintelligence’ as tech CEOs urge UN to act
How to opt out:
Most major chatbots use consumer conversations for training by default. ChatGPT, Claude, Gemini, and others all use consumer chats to train their models unless you opt out. Here’s where to find the setting on some of the most popular:
OpenAIs ChatGPT: Open Settings from your profile menu, choose Data Controls, then Model Improvement and turn off “Improve the model for everyone.” The setting applies across web, iPhone and Android when you’re signed into the same account. You don’t lose your chat history by doing this, but only conversations you have going forward will be restricted from training. For sensitive conversations, Temporary Chats aren’t used for training even if you leave the setting on.
Googles Gemini: Open Gemini Apps Activity, choose Turn off and decide whether to delete your saved activity. With activity off, chats aren’t used to train Google’s AI models unless you submit feedback
Anthropic’s Claude: The toggle is under Settings, then Privacy. Turning off “Help improve Claude” also cuts Anthropic’s retention of new chats from five years to 30 days. Conversations in Claude’s Incognito Mode are never used for training.
Metas Muse: The option to turn off training is under Settings, then Data controls, where youll find the toggle to turn off the setting to Help improve our AI. According to Meta, When the setting is off, your interactions with Muse will not be used to develop and improve AI at Meta. Changes to this setting also apply to previous interactions.
RELATED NEWS | Nvidia unveils security platform to stop AI agents from going rogue
OpenAI is still working to understand the full scope of its rogue agent activity, and the leak is likely to put more pressure on how AI companies protect the data they collect from users. In the U.S., no federal law guarantees a full right to opt out of AI training. Until one exists, protecting your data is largely up to you, one settings menu at a time