How to Set Up Claude to Protect Your Data

If you’re using Claude or ChatGPT for your children’s activity business, this is a quick, honest guide to the settings you can control right now.

None of this is a reason not to use either of them. I use AI every day in a business with almost 650 families on the books and it has given me back more hours than I can count. It’s just worth ten minutes setting it up properly before you get going, so you can stop thinking about it and get on with the useful part.

One thing before we start, so you’re not hunting for something that isn’t there. I use Claude, so the screens I’m describing are Claude’s. If you’re on ChatGPT, every principle on this page applies to you exactly the same, and every one of these settings has an equivalent somewhere in your settings. It’ll be in a different place with a different name, and I’m not going to guess where, because I don’t use it. Both are good products. Claude is the one that stuck for me.

So take the thinking, then spend ten minutes finding the settings in whichever one you actually use.

It is not legal advice, it won’t make you GDPR compliant on its own, and it isn’t a substitute for your safeguarding policy. What it does is show you what Anthropic’s settings actually do, so you’re not guessing. Everything here comes from Anthropic’s own pages and I’ve linked them at the bottom so you can check me.

The one rule that matters most

The people stay out. The problem can come in.

Never type or upload a child’s name, a parent’s contact details, a register, a medical form, or anything else that could identify a family. Describe the situation, not the person.

The five settings worth knowing

1. Turn off “Help improve our AI models”

Where: Settings, then Privacy, then scroll down to Preferences. The toggle says “Help improve our AI models” and underneath it, “Allow the use of your chats and coding sessions to train and improve Anthropic AI models.” That’s the one. Switch it off.

What it does: with it on, Anthropic says your chats can be used to train future versions of Claude, and they can hold onto them for up to five years. With it off, they say your chats aren’t used for training and your retention drops to 30 days.

Do this one first. It takes about ten seconds. Mine’s off.

2. Go and look at what Claude already remembers

Where: Settings, then Memory. The toggle says “Generate memory from chats”.

What it does: Claude builds up a set of notes about you as you chat, so you stop re-explaining yourself. It’ll pick up how you like things written, what your business is, the things you’ve told it more than once. It’s genuinely one of the best things about using it for work, and it’s the reason I don’t re-explain my brand every single morning.

The bit worth knowing: it’s switched on by default. Nobody turns it on, so nobody thinks to go and read it. If you’ve been chatting to Claude for a few months, there’s already a list in there and you’ve never seen it.

So go and have a look. Settings, then Memory, and you’ll see the entries grouped into categories. You can delete any of them one at a time, and you can just tell Claude in a chat what you’d like it to remember or forget. Each Project keeps its own separate memory too, so what it learns in one doesn’t leak into another.

What to do about it: honestly, mostly leave it on. It’s the feature that makes Claude feel like it knows your business. Just read what’s in there once, delete anything that shouldn’t be, and remember that the rule at the top applies here as much as anywhere. If a family’s details went into a chat, they can end up in a memory. Describe the situation, not the person, and there’s nothing in there to worry about.

Mine’s on, and I check it every so often.

3. Delete conversations you don’t need to keep

Where: your chat history. Delete them one at a time, or clear old ones in bulk.

What it does: the chat disappears from your history straight away, comes off Anthropic’s systems within 30 days, and won’t be used in any future training.

What it doesn’t do: it can’t reach back into training that has already happened. Deleting is good housekeeping. It is not an undo button.

While you’re in there: that same Privacy page has a “Shared chats” section with a Manage button. If you’ve ever sent someone a link to a chat, that link is still live and most people forget they did it. Have a look and clear out anything that doesn’t need to be sitting there.

4. Know which plan you’re on

Free, Pro and Max are the consumer plans. The training setting above is yours to turn off.

Team and Enterprise are the business plans. Anthropic says data on those isn’t used to train their models by default.

The consumer plans are sold to individuals rather than to businesses, so they don’t come with the sort of paperwork a business supplier agreement usually involves. Nothing to panic about. Just another reason the rule at the top exists.

5. Incognito chat, and what it’s actually for

Where: start a new chat and click the ghost icon in the top right.

What it does: it isn’t saved to your chat history, it doesn’t go into Claude’s memory either, it’s never used to train Claude even if the setting above is switched on, and it comes off Anthropic’s systems in 30 days. You can’t reopen it once it’s closed, and it doesn’t work inside Projects.

What it’s for: your own confidential business things. A conversation with anyone you’d rather wasn’t sitting in your history.

What it’s not for: families. There is a private mode. It is still not where a child or anyone’s name goes. There is no mode where a child’s name goes.

What these settings don’t do

They control what Anthropic’s model training sees. That’s it.

They don’t turn Claude into a secure record-keeping system, and they don’t replace your safeguarding policy, your GDPR documentation or your DBS processes. Keep using your proper systems for anything that needs to be a formal record.

And here’s the one most people miss. These five settings cover your chats. They don’t cover what you connect, and they don’t cover what you upload. The minute you connect your inbox, or drop a spreadsheet into a Project, you’ve handed over whatever was in it.

It’s an easy one to miss, because uploading a file doesn’t feel like typing something out. So it’s worth a two second pause before you drag anything in: what’s actually in this file?

What you can absolutely put in

The rule sounds strict, and it is, but it rules out far less than people think. Nearly everything you actually want help with is the problem, not the people.

All of these are completely fine:

• “A parent has complained about a class change and I need to reply without making it worse.”
• “Rewrite this term’s newsletter so it’s shorter and warmer.”
• “I’ve got fourteen teachers and a rota that’s not working. Talk me through it.”
• “Summer bookings are down on last year. What should I be looking at?”
• “Draft me a waiting list email that doesn’t sound like a robot wrote it.”
• “A four year old in one of my classes is struggling to settle and I want to handle it well. What’s worked for other people?”

Not one of those needs a family’s name, an address, a date of birth or a photograph. That’s the whole trick. You’re not hiding things from Claude. You’re describing the situation instead of handing over the file.

What about your own team?

Fair question, and it’s a different one. The rule at the top is about children and parents. Your staff are adults you employ, and how you handle their information is your call as their employer.

Here’s what I do, and it’s worked well. I use first names only. Claude genuinely can’t help with a rota unless it knows that Sarah does the Tuesday pre-school class and Emma can’t do Saturdays, so it needs to be able to tell them apart. So it gets first names and roles.

What it doesn’t get: surnames, phone numbers, addresses, pay, or anything about anybody’s personal circumstances. Roles and first names, nothing more.

That’s enough to be properly useful on a rota, and it isn’t handing over your team.

One last reminder

The people stay out. The problem can come in.

Never type or upload a child’s name, a parent’s contact details, or anything that could identify a family. Describe the situation, not the person.

Do that, and you can use it for everything else without a second thought.

---

Straight from Anthropic, if you’d rather read it yourself:

• Is my data used for model training?
• How long do you store my data?
• Use incognito chats
• Chat search and memory

August 2026 edition. Anthropic change these settings from time to time, so I check them every quarter and update this page when they move.

Questions? I’m at hello@katemarkham.ai, always happy to help.