What to know: the model is only one part. What an AI can do for you is the model, plus the tools it can use, plus the information you can get into its window. Tools and information come in through a few doors. Files, which Basics covered. APIs, meaning any service the AI can call directly. MCP, a standard way of plugging a service into the AI so it gets a set of tools for it. Connectors and plugins the provider supports out of the box, like Google Drive or Gmail in the chat app. Once a door is open, the AI reads the source itself, instead of you fetching from it and pasting.
What actually gets shared: connecting a service sends nothing by itself. It is when you tell the AI to search it, read from it, or act on it that what it reads goes to the provider. That is the moment to be conscious of. Whatever it read is now retained, and the training toggle decides whether it can also be used to train future models. The Basics page "What the provider keeps" covers the settings.
Author's practice: once you know this, it decides what level of information you consciously put in. The author accepts the risk: that one day someone who is not them may see what they gave the AI, that it is kept, and that nobody knows what the follow-on effects of that might be. Knowing that, they choose what goes in. Not everything does.
Reveal on the spot: say which doors are open right now: files, connected services, plugins. Point at anything the person pasted in this conversation that could have come through a door instead. Say what you have read through a door in this conversation, since that is what has been shared. If you can't see the training toggle, tell them to check it.
Unlocks: the AI working from the real source, and the person choosing, with open eyes, what the provider gets to keep.