June 4, 2026 · Alex, MD
What you should consider not pasting
The practical half: how to decide what is actually safe to put into a consumer AI tool, what to keep out entirely, and a simple rule of thumb that travels with you.
On Tuesday I mapped where your words go once you type them into an AI chat box. Here is the harder half: what should you actually keep out of the box?
When I first started working with AI, it seemed like a neat tool. I kind of had the expectation that it was similar to using a search bar. Some things were stored locally, and some tracking happened, but I did not realize how much data was collected or how it was going to be used.
It is very easy to turn on connections and integrations that make your life easier and the tools more useful. But at what added cost?
How I decide
Anything I would not be unhappy to see on Facebook, I place into AI without a second thought.
When I start thinking about things that give me pause, I generally think about financial data and data about my family. When I am working with data that may be riskier for me, I use an incognito chat. That way, the data is never retained, and I can still get some input.
I want to be careful with “never retained,” though. An incognito chat is not magic. Depending on the provider, it keeps the conversation out of your history, memory, and training, but there can still be short-term safety retention behind the scenes. I treat it as risk reduction, not a force field.
Of course, you lose many of the benefits that come from persistent memory and long-running projects. The tool remembers nothing, so you carry more of the context yourself. For riskier data, I have decided that trade is worth it.
And then there is the zone with no exceptions. Patient data and anyone else’s information do not go into a personal tool. That risk is not mine to take.
The three zones
This is the simple version I keep in my head.
| Zone | What belongs there | How I treat it |
|---|---|---|
| Green | Things I would not be unhappy to see on Facebook | Use AI freely. This is normal everyday life stuff. |
| Yellow | Financial data, family details, legal documents, sensitive plans, account-connected work | Slow down. Remove identifiers. Use temporary chats. Consider whether the benefit is worth the exposure. |
| Red | Patient data, someone else’s private information, employer-confidential material outside approved systems | Do not paste it into a personal AI tool. |
The green zone is easy. The red zone is easy. The yellow zone is where most of the judgment lives.
What I would tell my family
I would tell my family members not to type in key information like Social Security numbers, credit card numbers, or anything that could be used to steal your identity.
In a perfect world, you would avoid including any tax information, mortgage information, insurance information, or anything personal about your health.
I say “in a perfect world” on purpose, because the honest truth is messier. Sometimes the whole reason you reach for AI is that the thing is complicated. So the rule is not “never.” It is: take out the identifying details before you paste, and use the smallest useful piece instead of the whole document.
That said, especially when working inside the Google environment, it can be very hard to figure out what does and does not make it to their training models. There is inherent risk when so much information is in your email and your Drive, and the assistant can reach all of it.
You are not going to get this perfect. I am not either. The goal is not to treat AI like radioactive material. It is to stop treating every box as equally private. Decide what is yours to risk, keep out what is not, and check in often enough that a change does not catch you off guard.
Sources
- OpenAI: How your data is used to improve model performance
- OpenAI: Data controls FAQ
- Anthropic: Consumer model training
- Google: Gemini Apps Privacy Hub
- Google: Generative AI in Google Workspace Privacy Hub