Opt Out of AI Training on Your Data
By default, ChatGPT, Claude, Gemini and Perplexity may use your inputs to train their models. Every one of them lets you turn that off, and every one does it a little differently. Here is the right guide for all four.
Why it is worth doing
Once something has flowed into the training of an AI model, it can practically no longer be retrieved in a targeted way. With Gemini, human reviewers also read a sample, and Perplexity states outright that training data already collected cannot be removed. If you enter third-party data professionally, such as customer or employee details, you remain responsible for it under data protection law. In all four cases the opt-out takes only a few minutes.
The guides
ChatGPT
OpenAI trains on your conversations by default. How to turn it off in Data Controls, including the formal opt-out.
ChatGPT guide →Claude
Since autumn 2025, Anthropic trains on your chats by default and keeps them for up to 5 years. How to switch it off.
Claude guide →Gemini
Google uses your chats for training, and human reviewers read a sample. How to turn off Keep Activity.
Gemini guide →Perplexity
Perplexity uses your search data for AI training by default, and nothing already collected can be deleted. How to switch it off.
Perplexity guide →How the four differ
The switch is called something different everywhere, and more importantly the catch is different everywhere. On ChatGPT the setting sits in Data Controls, and there is an additional formal opt-out form for your account. On Claude, Anthropic flipped the default in autumn 2025: anyone who has not objected since is training the models, and the chats are kept for up to five years. On Gemini it is not only about training but about human reviewers: once something has been in the sample it stays for up to three years, even if you delete your history. And on Perplexity the opt-out works only going forward, because training data already collected cannot be removed, by their own account.
One thing follows from all of that: the earlier you object, the less is already gone. It is also worth re-checking the setting after major product updates. All four providers have changed their defaults in the past, and a new switch is rarely set to "off".
What the opt-out does not do
Important context: the opt-out only prevents training. Your inputs are still transmitted to the provider and processed there. This is exactly where AEGIS comes in: the browser extension detects personal data and pseudonymizes it before it even reaches the AI. Training opt-out and pseudonymization complement each other. See how it works in the interactive AEGIS demo.