Which tool to choose
The useful tools are three, and you control them. First: the memory switch, where it exists, to decide whether the assistant saves or not. Second: the memory section, to view and delete what it has already noted down. Third, the most important, is your own judgment: deciding before writing whether a piece of information can stay there. The settings help, but the real defense is not entering what you don't want out: what you don't write can't be saved, exposed, or misunderstood.
How to do it
- Before giving the AI a piece of data, ask yourself: would it bother me if it ended up elsewhere? If so, don't give it or anonymize it.
- Replace sensitive data with placeholders: instead of the real name, "the client"; instead of the real amount, an example figure.
- Tell the assistant what it must not memorize ("don't save this data"), and check in the memory section that it doesn't remain.
- Every now and then reread the memory and delete what's too personal to stay there.
A concrete example
Stefano wanted help writing a delicate letter about a family member's health issue. Instead of putting in the name, the condition, and the real details, he described the situation in generic terms: "a relative with a serious health problem", without specifying which one or who. The AI helped him find the right tone all the same, because writing a letter takes emotions, not the medical file. He got what he needed without entrusting an external tool with information concerning another person.
When it DOESN'T work (and how to fix it)
If you realize you've already given sensitive data
Go to the memory section and delete the entry; if it was in a chat, consider deleting that conversation too. You can't go back in time, but you can reduce what stays saved from here on.
If you really need the AI to reason about real data
Anonymize as much as possible and give only the minimum the reasoning needs. Often, to get good advice, you don't need the real data but a realistic version: the AI reasons on the structure of the problem, it doesn't need the exact names and numbers.
If you handle other people's data
Here the caution doubles: information about clients, patients, family members isn't yours to share lightly. Always anonymize, and for professional contexts check which rules apply in your field before entering other people's data.
A tip from someone who really uses it
Decide on privacy before writing, not after. Once a piece of information has been entered, even if you delete it, you're in the hands of how that service handles it: the real control you have only in the instant before pressing enter. Get into the habit of pausing for half a second over delicate data and asking yourself whether it's really needed or whether an anonymous version is enough. Almost always it is. That half-second pause is worth more than any setting, because it acts in the only moment when you're the one in command: before the data goes out.
Frequently asked questions
Is what I write to the AI private?
Don't take it for granted. Depending on the service and the settings, conversations may be kept or reviewed to improve the tool. Treat what you write as potentially not private and adjust accordingly on delicate data.
Can I tell the AI not to remember something?
Yes, and where the memory exists you can also delete the entries or turn it off entirely. But the most solid defense remains not entering what's truly sensitive.
Does anonymizing ruin the quality of the answers?
Almost never. For most reasoning the AI doesn't need the real data: exact names and figures can be replaced with realistic placeholders without losing anything of the answer.
It's a private conversation between me and the AI, isn't it?
No, and it's the misunderstanding that leads to sharing things one shouldn't. A chat with an assistant isn't like a thought in your head: it's data that travels to an external service, where it can be kept, saved in memory, or handled by rules you don't decide. "Private" doesn't mean safe. The only information you're sure won't get out is the one you don't write: that's why privacy starts from the choice of what to enter, not from a setting turned on afterward.