Why are we giving AI access to everything?

I remember the articles about AI running out of training data. In 2024, researchers warned that the supply of public human-written text could limit how far companies could keep scaling their models. That made me wonder where the next valuable data would come from.

Now I see people connecting AI to their email, documents, private notes, and work systems. Some give agents broad access to their computers. I can’t prove this is a way to collect training data, but it makes me uncomfortable.

In my corner of the internet, people used to talk about encrypted messengers and protecting their data. Some of the technical people I now see celebrating these tools seem happy to give them access to almost everything, as long as it saves time.

I understand the appeal. Having something handle boring work sounds useful. But I’m surprised by how little concern I see about what we’re handing over.

Training policies vary. Some services use conversations to improve models, depending on the product and settings. Turning training off can still leave data stored on a company’s servers. Google’s Gemini documentation, for example, says temporary chats aren’t used for training but are retained for 72 hours.

I can read those policies. I can’t personally verify everything happening behind them, and I’m not comfortable connecting my private life to a service just because the company says it’s safe.

There are security risks too. An agent reading your email can encounter malicious instructions that try to make it share private information. Anthropic describes this risk in its own research.

I’d consider connecting a local model to my private apps if I could verify that the whole setup kept my data on my computer and restrict what it could do. That’s the level of control I’d want.

Maybe my online bubble gives me the wrong impression. Still, before connecting another account, I’d want to know exactly what the tool can access and what happens to the data it reads.

If I’m missing something, send me a message. I’d like to see the evidence.