What never to share, and the 30-second check

The situation

You’re about to upload a screenshot of the class group chat to get the dates out of it, and it shows twelve other parents’ names and phone numbers. Or you’re about to paste your son’s letter from the doctor, or a photo of your passport for a form. Pause for ten seconds first.

By the end of this lesson you’ll understand what happens to what you type into an AI app, you’ll have a short list of things that never go in, the right privacy settings switched on, and a 30-second habit that catches the AI’s mistakes before anyone else sees them.

Why it’s worth it

AI apps are most useful exactly when life is most personal: school letters, medical appointments, money worries, family logistics. That’s also when it’s easiest to over-share in a hurry. The good news is that you can get almost all of the help with almost none of the risk, using a couple of ten-second habits.

There’s a second reason, which is easy to forget: a lot of family admin involves other people’s information. The other parents in the group chat, the coach’s mobile number, your mother-in-law’s date of birth. They didn’t choose to share it with an AI company. Keeping their details out is simply good manners, and in many countries it’s what privacy law expects too.

The other half of this lesson is about the AI’s mistakes rather than your data. Everything it drafts goes out under your name. The confident guessing from Lesson 1.1 means a wrong date or an invented promise can slip in, and a 30-second check catches nearly all of it.

Where your chats actually go

When you send a message or upload a photo, it travels to the company that runs the app (OpenAI for ChatGPT, Anthropic for Claude, Google for Gemini) and is stored with your account. That’s how your chat history works. Three things can then happen to it, depending on the app and your settings.

It can be used to improve future AI models. “Training” means the company uses large numbers of conversations to teach the next version of the AI patterns of language. Your chat isn’t published, but it becomes part of what the model learns from. Each app lets you switch this off.

It can be reviewed by people. Companies check some conversations for safety problems and misuse, and some of that review is done by humans.

It is kept for a while, even after you delete it, under the company’s privacy policy. Deleting a chat removes it from your history, but the company may still hold a copy for a period.

None of this makes AI apps dangerous. It just means the right mental model is a very helpful stranger: friendly, useful, discreet in principle, and still not someone you’d hand your passport to.

The never-share list

These stay out of every AI app, whatever the settings, because the harm if they leak is real and the AI never needs them to help you:

• ID numbers: passport, driving licence, national ID, tax or social security numbers. These are the raw material for identity fraud.
• Passwords and codes: PINs, one-time codes, answers to security questions. They give someone your accounts.
• Money details: bank account and card numbers.
• Medical records or letters with names on them. Health information is among the most sensitive data there is.
• Your children’s full names together with their school, address or photos. Each detail alone is harmless; together they say exactly where a child will be.
• Other people’s details: names, phone numbers and faces in screenshots of group chats. Not yours to share.

Crop and swap: the ten-second habit

You can still get help with every one of those situations. The trick is that the AI almost never needs the identifying details to do the useful part:

• Crop screenshots to the part you need, and scribble over names and numbers with your phone’s photo editor before uploading. To pull dates out of a group chat, the AI needs the dates, not the senders.
• Swap names for roles: “my son (11)”, “the coach”, “Parent A”. The AI writes just as well about “the coach” as about “Dave”.
• Describe instead of uploading: “a letter from the doctor says he can do light exercise from next week” tells the AI everything it needs to plan, without the letter itself.

Privacy settings worth two minutes

Each app has a switch that stops your chats being used to improve its AI, and a private mode for one-off sensitive chats. Names vary by version; the official help pages have the current steps.

• ChatGPT: Settings → Data controls → turn off Improve the model for everyone. For a sensitive one-off, start a Temporary Chat: it isn’t used to improve the models and is deleted from OpenAI’s systems after 30 days.
• Claude: Settings → Privacy → turn off the setting that lets your chats help improve Claude. Incognito chats aren’t used for training.
• Gemini: turn off Keep Activity (in Gemini Apps Activity) so future chats aren’t used to improve Google’s services. Google still keeps them for up to 72 hours to run the service. For one-offs, use a Temporary chat.

If your app has a memory feature switched on, it may also carry details from one chat into later ones. It’s worth opening that setting once to see what it has saved, and deleting anything you’d rather it didn’t keep.

Why the AI’s drafts need a check

Now for the other direction: what comes out. An AI draft can be wrong in a handful of predictable ways, and knowing why makes them easy to spot.

Dates and days go wrong because the AI doesn’t look at a calendar. Working out that the 14th is a Thursday, not a Friday, is calculation, and the AI writes what sounds right instead of calculating. Names go wrong when it mixes up two people from your message, or “corrects” an unusual spelling. Promises creep in because the AI is trying to be helpful and polite: “I’ll bring the form in on Monday” sounds nice, even if you never offered. Money slips when it rounds, assumes or invents an amount. And tone can be off because it can’t hear how things land with the actual person reading.

None of these are rare, and none of them take long to catch, if you look for them on purpose.

Copy this

You can make the AI do the first pass of the check itself. After any draft, paste this in the same chat:

Before I send your draft, list every name, date, day, time, place, amount of money and promise in it. Mark anything you guessed or I didn't tell you. Then tell me if anything could come across as rude or cold.

Why each part is there

“List every name, date, day…” turns the draft into a short checklist. It’s much easier to spot a wrong date in a list of five facts than buried in a friendly paragraph.

“Mark anything you guessed or I didn’t tell you” asks the AI to separate what came from you from what it filled in. It’s surprisingly good at this when asked directly, and it’s exactly where the confident guessing hides.

“Could anything come across as rude or cold?” catches tone problems you might miss when you’re tired or annoyed.

What you’ll see: a short list under the draft, with warnings next to anything it made up.

What you’ll get

An example (yours will look different):

• Names: Ms Patel and Mia (both from you)
• Date: “Friday 14th”. ⚠️ You said Friday but not the date. I guessed the 14th.
• Time: 3.30pm (from you)
• Promise: “I’ll bring the form in on Monday”. ⚠️ I added this. Delete it if you didn’t mean to promise it.
• Tone: friendly overall, but “As I said before” could sound impatient.

Notice that two of the five lines are warnings: a guessed date and an invented promise. Both would have gone out under your name. Notice too that the AI is checking its own work here, which is useful but not foolproof: it can miss a slip the same way it made it. That’s why the final check is yours.

Before you send: the 30-second check

Even with the AI’s own list, do this yourself. It’s the habit that runs through the whole course, and it gets faster every time:

1. Names: the right people, spelled right.
2. Dates and days: does “Friday 14th” really fall on a Friday? Check your own calendar.
3. Times and time zones: especially for anything involving travel.
4. Money: amounts, and who’s paying.
5. Promises: did it commit you to something you never offered?
6. Tone: read it out loud. Would you say it to their face?

Make it yours

• Save the check prompt in your notes app, or as a text shortcut on your phone’s keyboard. Most keyboards let you set a short code like “chk” that expands into a longer phrase, so the check is three letters away.
• Make it automatic. Add this line to your instructions from Lesson 1.3: Under every draft, list any dates, times or facts you weren't sure about. Then you don’t even need to ask.

If it goes sideways

• You already shared something you shouldn’t have. It happens to everyone. Delete that chat from your history and carry on; deleting removes it from your history, though the company may keep a copy for a while under its privacy policy. If it was a password or card number, change the password or contact your bank, just as you would after any accidental share.
• The AI refuses because the request looks personal. Some apps are cautious with anything that looks like private data. Rephrase without the personal bits: roles instead of names, a description instead of the document.
• The check finds nothing, but something’s still off. Trust your gut. You know the people and the plans; the AI is working from whatever you gave it.