Using a cloud chatbot on purpose

Most of this site teaches you to run a model on your own computer, and that really is the private option. But plenty of people are going to type into a chatbot in a browser, because it’s already there and it works. If that’s you, this page is not a lecture about installing something instead. It’s how to do this version well.

Start by changing three settings, once

Every major service has these. The names differ and they move around, so look for the idea rather than the exact wording, in your account settings or under a heading like data controls.

Turn off using your conversations to improve the model. This is usually a single switch, often on by default, sometimes only available on the website rather than in the app. It is the one setting most worth finding.
Decide what happens to your history. Most services keep every conversation indefinitely until you say otherwise. Some let you set automatic deletion after a period. Turning history off usually costs you the ability to scroll back, which is a real trade — make it deliberately rather than by default.
Learn where the temporary or private chat option is. Most services now have one. It’s the right default for anything you wouldn’t want sitting in a list for years, and it takes one click once you know where it lives.

Then check them again in six months. These reset, or move, or acquire new siblings after a redesign — the same maintenance problem described in the chapter on turning off AI you didn’t ask for.

The working assumption

Assume that anything you type could, under some circumstance, be read by a person at that company — a support investigation, a legal request, a security incident. This is not an accusation; it is true of your email provider too, and you use email. It is simply the right mental model, and it makes the next section obvious rather than frightening.

What not to paste

Not a list of forbidden topics. A short list of things where the cost of being wrong is high and the benefit is usually small.

Other people’s information. Your own medical history is your decision to make. Your sister’s isn’t, and neither is a letter from your doctor about somebody else.
Account numbers, passwords, one-time codes, card details. There is no task that needs these, and a chatbot cannot use them anyway.
Whole documents, when a paragraph would do. This is the same habit that keeps costs down in Chapter 04, and it happens to be the privacy answer as well.
Anything your employer would consider confidential. Customer records, unreleased work, contracts. If your workplace provides its own AI tool, use that one; if it doesn’t, assume the answer is no.
Photographs of documents. A picture of a statement or a letter contains everything on it, including the parts you weren’t thinking about — the account number in the corner, the other name on the account.

Everything else — drafting a letter, explaining a term, planning a trip, arguing with you about a recipe — is genuinely fine, and worth enjoying.

Free, paid, and work accounts are not the same

Free tiers are the most likely to use your conversations for training, and the least likely to give you retention controls. Paid personal accounts usually have better settings, though you still have to turn them on.

Work accounts are a different animal: the account belongs to your employer, the conversations are generally theirs to see, and that is normally the point rather than a scandal. Just know which account you’re typing into. The same is true in reverse — using a personal account for work is how confidential material ends up somewhere nobody intended.

What deleting actually does

Deleting a conversation removes it from your view and, in most services, starts a process that clears it from their systems within some stated window — often around a month. That is a real deletion and it’s worth doing.

What it doesn’t do is reach backwards. If a conversation was used to train a model before you changed the setting, deleting the chat later doesn’t remove its influence. Nor does it cover anything a human reviewer already looked at, or copies kept for legal reasons. So the setting is the thing to get right early; deletion is housekeeping, not a remedy.

Voice, photos, and the features that watch

Voice conversations are recorded audio of you, which is exactly the material discussed in reducing what’s available to copy. Photo upload sends the whole image. Features that offer to read your screen, your inbox, or your files continuously are a different category again — see when it acts on your behalf before switching one on.

The one habit that matters most

Before you paste, ask what would actually happen if this specific text ended up somewhere you didn’t choose. Most of the time the honest answer is nothing at all, and you should paste it and get on with your day. Occasionally the answer is something real, and that’s the moment to either trim it down to the part you need help with, or to reach for a local model instead.

That question takes about two seconds and replaces the whole anxious business of wondering whether AI is safe. It isn’t a category. It’s a series of specific choices, and you are perfectly capable of making them.