ChatGPT, Perplexity and Claude all ship with settings that let users keep their conversations out of future model training, and Malwarebytes Labs has now published the exact paths to reach them. The guide arrives after 404 Media reported that OpenAI is recruiting hundreds of contractors to review ChatGPT prompts, a program reportedly called Project Lily in which those contractors score the chatbot's answers.
For ChatGPT, the switch is called Improve the model for everyone, and it is on by default for Free, Plus and Pro users. OpenAI says turning it off keeps chats out of model improvement, though it adds that authorized staff and vetted service providers may still reach content for abuse and security investigations, customer support, troubleshooting and legal matters. On desktop the route runs through the profile icon, then Settings, then Data Controls, where the toggle sits; on mobile the same steps start from the sidebar and the profile icon.
Perplexity's control lives under Settings, then Preferences, then AI data retention. Perplexity says switching it off only blocks training use of data gathered from that point forward, and that anything collected earlier stays in place. Perplexity also notes that data may still be processed to run and improve the service and to meet legal obligations.
Claude users reach their setting by opening their name, then Settings, then Privacy, then Help Improve our AI Models. Anthropic says new chats and coding sessions are excluded from future training once that is off, and that conversations already stored are excluded from later training runs as well. Data that has already gone into a training run cannot be pulled back out. Anthropic adds a caveat for conversations its systems flag: "If our safety classifiers flag your conversations, they may still be used to improve our internal trust and safety models, detect harmful content, enforce our policies, or advance our safety research."
OpenAI's own account of who sees chats is narrower than the contractor report suggests. The company says only limited authorized personnel and trusted providers can access conversations, and that a Privacy Filter model strips personal details before reviewers ever see a prompt. Reviewers do not see usernames, according to OpenAI, though they may still encounter personal details along with a summary of a user's memories.
The guide does not address enterprise or API accounts, and it closes by advising readers to check the privacy policies and settings of any other AI tools they use.
