Most of us tapped "Accept" on an AI app once and never looked again.
I have set Claude up on a lot of phones and laptops now.
Almost nobody I meet has opened its privacy screen.
So these seventeen short articles walk through what Anthropic's own privacy policy says.
No hype and no fear, just what is actually written down.
Who is this for?
Anyone who uses an AI assistant most days and has never checked what it keeps.
It reads just as well if a son or a niece goes through it with you.
By the end you will know three separate things.
What is collected, what is yours to control, and what no setting can change.
That last one stops you worrying about the wrong things.
Does my plan change anything?
Yes, and it is settled before you touch a single setting.
Free, Pro and Max sit under Anthropic's Consumer Terms.
There, your chats improve the models only if you opt in.
Team, Enterprise and the API sit under the Commercial Terms.
There, training is off by contract rather than by a toggle you have to find.
So the trap is doing office work on a personal account.
The protections your employer assumes are in place simply are not there.
If it is work, use a work plan.
What does Claude actually collect?
Your messages, and Claude's replies alongside them.
Files and screenshots count as messages too, under the same rules.
A screenshot of a bank statement is a bank statement.
Then the ordinary account bits: your name, email and device details.
Card details are handled by a third-party payment provider, not by Anthropic.
Location is collected only if you choose to share it.
It is the one genuinely optional item on the list.
Anthropic states that it does not sell your data.
Claude carries no advertising either.
Does training keep a copy of my chats?
No, and this is where most of the fear lives.
A trained model is not a filing cabinet with your name on it.
It learns patterns across enormous amounts of text.
Once trained, it cannot pull your original conversation back out.
Your chats do still sit in your account, though.
That is a separate question, and it is the one worth managing.
Where do the controls live?
On Free, Pro and Max, go to Settings, then Privacy, then Model Improvement.
You may see it labelled "Help Improve Claude" instead.
OFF means your chats are not used for training and are kept about thirty days.
ON means they help improve Claude and are kept up to five years.
Memory sits under Capabilities rather than Privacy, which is why nobody finds it.
Deleting a chat clears it from your history at once.
It then leaves back-end systems within thirty days, whichever way your toggle points.
For a one-off sensitive question, an Incognito chat is cleaner still.
Incognito chats are never used for training.
A training run that has already started cannot be pulled back by a toggle or a delete.
What can no setting change?
That your message is read by the system answering it.
Your words travel to Anthropic's servers and go through the model.
That is the only way a reply can exist.
The controls decide storage and training, not processing.
So keep passwords, OTPs, card numbers and Aadhaar numbers out of the box.
There is also one honest exception to turning training off.
Chats flagged by Anthropic's safety classifiers as possibly breaking the Usage Policy can still be used.
The purpose is narrow, which is catching abuse and enforcing the rules.
Those are kept longer, with inputs and outputs up to two years.
Trust and safety scores are kept up to seven.
Stay inside the Usage Policy and it never touches you.
An app that says "off means off" with no footnote is the one hiding something.
Is anyone reading my chats?
By default, no.
Automated systems run the model and check for abuse, with nobody reading along.
Your chats are encrypted in transit and at rest.
Staff can look in limited cases, such as a conversation flagged for breaking the rules.
The other case is when you give consent yourself.
Tapping thumbs up on a reply is exactly that consent.
Rating a reply shares the whole conversation, even with training switched off.
That copy can be kept up to five years.
Your email and account ID are separated from it first.
So give feedback freely, but glance at the chat before you tap.
Does this help if I use ChatGPT or Gemini?
Yes, although every menu path here is Claude's.
The names differ from app to app, but the shape does not.
Something is collected automatically, something is your decision, and something is unavoidable.
Sorting any assistant's settings into those three piles is the actual skill.
It travels with you to whatever you use next.
Things to Remember
- Check which plan you are on before you check a single setting.
- On Free, Pro and Max, training happens only if you opt in.
- Treat every file and screenshot exactly like text you typed by hand.
- Rating a reply shares that whole conversation, even with training turned off.
- No setting stops your message being processed, so keep account numbers out of it.
- Delete a chat the day you regret it, not next year.
17 parts, in order. A new one goes up every morning at 10, so this list grows until the series is complete.
- Everything You Type Into an AI: A Five-Part GuideThe start of a five-part plain-English series on what Claude collects, how it uses it, and which of it you can actually control.
- Your Plan Decides Whether AI Trains on Your DataConsumer and commercial Claude plans follow two different rulebooks, and using a personal account for work quietly changes which one applies.
- What Happens to Your Chats, and How to Delete ThemYour messages are saved to your account until you remove them, trained on only if you opt in, and cleared from back-end systems in about 30 days.
- Dragging a PDF Into Claude Counts as a MessageFiles and screenshots you upload are handled exactly like text you typed, with the same training rules and the same retention.
- What Does Claude Know About You Outside Your Chats?Beyond your conversations, an AI account holds ordinary sign-up and billing details. Anthropic says it sells none of it.
- Location Is the One Piece of AI Data You ChooseDevice details, usage logs and cookies are collected automatically around your chats. Location is the piece you hand over yourself.
- What You Trade When You Connect Drive or Slack to AILinking Google Drive or Slack to Claude shares some data with it, but that content stays out of training unless you paste it in.
- Why No Privacy Setting Stops Claude Reading Your MessagePrivacy controls decide how long an AI stores your data and whether it trains on it, not whether your words get processed.
- What Training on Your Data Actually MeansA trained model is not a filing cabinet of your conversations. Here is what happens to your words when an AI learns from them.
- What Does Model Improvement Mean in Claude?One setting on Free, Pro and Max decides whether your chats train Claude, and whether they live 30 days or five years. Here is what each side does.
- Before You Tap Thumbs Up on a Claude ReplyRating a Claude reply shares that whole conversation, even if you have turned model training off. Here is what the thumbs buttons really do.
- The One Exception to Turning Training OffEven with training off, a chat flagged for breaking Anthropic's Usage Policy can still be used to catch abuse. Here is what that means in practice.
- Is Anyone at Anthropic Reading Your Chats?By default, no human sits reading your Claude conversations. Here is what automated systems do, and the limited cases where staff can look.
- Where Claude Keeps Its Privacy ControlsA simple map of Claude's Settings screen on the web, showing exactly which tab holds training, memory, data export and account deletion.
- Turn Off Claude Training in Four TapsStopping Claude from training on your chats comes down to one toggle under Privacy. Here are the exact four taps, and why the timing matters.
- How Long Does Claude Keep Your Chats?One setting decides whether your Claude conversations live for about 30 days or up to five years. Here is what changes either way.
- What Happens When You Delete a Claude Chat?Deleting a Claude conversation removes it from your history at once and from Anthropic's servers within 30 days. There is one catch.