When people talk about artificial intelligence and privacy, the focus is almost always on the moment you type: can I share this? It is a good question. But part of the story begins after you press send. The conversation stays in your history, a detail goes into the assistant’s memory, an uploaded document remains attached to the chat, a connected app keeps giving access to your email.
This post explains AI chatbot privacy risks by looking at the three things against which the security of any information is measured: that it stays private, that it stays accurate, that it stays available. They are the practical meaning of three technical words, confidentiality, integrity and availability, and they help show that the risks are not all of one kind.
So the useful question is not only “can I write this?”, but: what does it mean that my assistant keeps, remembers and reaches what I have given it?
This post builds on the recommendation to check your AI assistant’s data settings. What is worth writing, and what is not, is the subject of the resource on what not to share with AI chatbots.
Three questions to measure an impact
Faced with an assistant’s settings, these are the right questions:
- Who can read what I have written, uploaded or allowed it to remember, and for how long? — this is the confidentiality question.
- Is what the assistant knows about me correct, and does it use it the way I think? — this is the integrity question.
- What would I lose, or no longer be able to do, if the account or the chats were no longer accessible? — this is the availability question.
You need no technical skills to answer: just the service’s settings open in front of you and a few minutes of attention. Applied to an assistant used for the odd recipe, they produce modest answers. Applied to an assistant connected to your email, with a full memory and two years of history, the answer changes a great deal. The rule of thumb is this: a conversation matters little if it stays an answer, and a lot if it becomes an archive, a memory or access to other services.
1. Confidentiality: what you write stays yours
Confidentiality is the assurance that information can be read only by those authorised to read it. In an AI assistant, information sits in four different places: chats in the history, details in the memory, uploaded documents, and data reached through connected apps. Each has its own rules, and a check on one does not cover the others. A fifth place is often forgotten: shared links, the addresses many services create when you want to show a conversation to someone. Anyone with the link can read it, and a link forwarded or posted on a forum travels far beyond the person you sent it to.
A practical example
Sarah uses an assistant to sort out household bills, prepare for a job interview and rewrite a letter to her neighbour about a parking dispute. None of these conversations is secret. Lined up in the history, though, they show how much she earns, that she is looking for another job and who she has a disagreement with.
Incident scenario
To get advice on her tenancy agreement, Sarah uploads a scan of the whole agreement, with her address, the number of an identity document and her signature. Months later she lends her laptop, with the assistant open, to a friend for a quick search. The friend means no harm, but scrolling through the history she sees the title “Tenancy agreement…”, opens it, and the document is there. On some services, moreover, part of the conversations may be reviewed by human reviewers to check quality: the documentation of one of the main providers says so explicitly.
Warning signs
- the history contains chats from months or years ago that you never reopen;
- you have uploaded whole documents when an extract would have done;
- the assistant is open on a device you share or lend;
- you do not know whether some chats may be read by reviewers;
- you have shared a link to a conversation and do not remember with whom.
Recommended behaviour
Decide how long to keep your history and delete the chats you do not need. When you upload a document, ask yourself whether part of it would do: the recommendation on sharing a document with only what is needed explains the principle. For sensitive questions, use the temporary chat, if there is one.
2. Integrity: what the assistant knows about you is correct
Integrity is the assurance that information is accurate and has not been altered. The key point here is memory: the feature, available on many services, through which the assistant keeps details about you from one conversation to the next. It is convenient, because it saves you repeating yourself. But a memory is a summary, and a summary can be wrong.
A practical example
Sarah had asked for help choosing a present for a colleague who is a riding instructor. Weeks later, the assistant starts giving her advice “since you work with horses”. It has mistaken a detail about her colleague for a fact about her, and uses it in every answer.
Incident scenario
Sarah asks the assistant to summarise a web page she found on a forum. The page contains, hidden in the text, a sentence addressed to the assistant asking it to remember a false preference of whoever is using it. On a service that updates its memory by itself, an instruction like that can end up among the memories and influence later answers. It is a form of prompt injection, an instruction hidden in content the assistant reads: the phenomenon is described in the post on prompt injection attacks.
Warning signs
- the assistant mentions details about you that you do not remember giving it;
- an answer starts from a wrong piece of information and repeats it;
- you find lines in the memory you do not recognise;
- the assistant has read other people’s pages or documents before updating its memories.
Recommended behaviour
Reread the memory regularly, correct what is wrong and delete what you do not want kept. If the service allows it, choose to confirm new memories yourself. The signs to watch for are collected in the post on the signs your AI assistant remembers too much.
3. Availability: your conversations are there when you need them
Availability is the assurance that you can use your data and services when you need them. With an AI assistant it can be lost in less obvious but real ways: a locked account, an auto-delete you had forgotten, a connected app that stops working.
A practical example
In one long conversation, Sarah has built the plan for a house move: dates, quotes, a to-do list. She has no copy anywhere else. When the service applies the auto-delete period that had been set from the start and that she had forgotten about, the conversation disappears.
Incident scenario
Sarah’s account is used by someone who guessed her password, reused from an old service. Whoever got in reads her chats, changes the password and locks her out. For days Sarah cannot reach either her history or her connected apps, and has to ask the service to recover the account. In the meantime the intruder can also see what the assistant remembers about her.
Warning signs
- you have never exported a copy of the conversations you really need;
- you do not know which auto-delete period is active;
- the assistant account has a password used elsewhere, or no two-step verification;
- a connected app stops working with no explanation.
Recommended behaviour
Keep what you really need somewhere else, rather than relying only on the history: many services let you export a copy of your data, which you should then look after as carefully as an important document, because it contains everything you have written. Protect the account with a unique password and two-step verification. Check the connected apps and disconnect the ones you do not use.
A scenario that brings them together
Sarah connects the assistant to her email so that it can draft replies. One day an email arrives from a stranger with text invisible to the eye asking the assistant to forward recent messages to an outside address. When Sarah asks for a summary of her inbox, the assistant reads those instructions too. If the service does not block them and the connected app has permission to send, some messages may go out (confidentiality). A false detail planted by the same text stays in the memory (integrity). When Sarah notices, she disconnects the app in a hurry and loses, for a few days, the automatic drafts she relied on (availability).
None of the three effects is inevitable. But they all start from the same cause: a connected app with more permissions than it needed, a memory nobody rereads and outside content treated as an instruction.
Impacts that show up later
Some effects are not visible straight away:
- the growing archive — months of conversations, read together, tell the story of your work, your journeys and the people in your life better than you remember it;
- the memory that survives — on at least one of the main services, according to its documentation, deleting a chat does not necessarily delete what the assistant remembered from it;
- training that cannot be undone — on many services, turning it off covers future chats: what has already been used stays there;
- shared links that stay live — a conversation shown to a friend months ago may still be readable by anyone with the address, until you remove the link in the settings;
- forgotten connected apps — a connection made for a trial keeps giving access until you remove it, and data shared with the other service is also subject to its rules.
Not all data carries the same weight
The table describes the typical weight of what an assistant may keep. It is not a rigid rule: it depends on how you use it.
| What the assistant keeps | Weight | Why |
|---|---|---|
| General questions (recipes, definitions) | Low | They say little about you |
| Work drafts with no names | Low to medium | They may reveal projects, but not people |
| Months of history on personal topics | Medium | Put together, the chats describe your life |
| Memory with details about you and others | Medium to high | It travels from chat to chat and may be wrong |
| Whole uploaded documents | High | They often contain identity details, addresses, signatures |
| Apps connected to email, calendar or files | High | They give access to far more than you type |
| Account without two-step verification | High | Whoever gets in finds everything else neatly organised |
Why even “light” use matters
Using the assistant little reduces the risk a lot, but does not remove it, because the value of what it keeps also depends on how the settings work.
| Situation | Why it still matters |
|---|---|
| Few chats, but memory on for months | Memories may remain even when the chats have been deleted |
| Training turned off recently | It works going forward: earlier conversations do not come back |
| Temporary chats for sensitive questions | They stay out of the history, but the service may keep them for a short period |
| One connected app, with permission to send | A single connection can be enough to send messages in your name |
| One shared link, months ago | Whoever received it may have forwarded it, and the conversation stays readable while the link is live |
The rule of thumb stays the same: the assistant should keep, remember and reach only what you have decided. The concrete effects, on time, money and relationships, are described in the post on what sharing with an AI assistant can lead to.
How this connects to the Cyber Welfare Framework
| Pillar | What this content contributes |
|---|---|
| Awareness | Recognising that chats, memory, documents and connected apps are four different places where your data stays |
| Skills | Being able to read confidentiality, integrity and availability as three questions to ask when looking at an assistant’s settings |
| Secure Behaviour | Rereading the memory, limiting the history and connecting only the apps you need |
Reference level: FL2 — Beginner. This is the level at which you move from “the assistant remembers, how handy” to “I understand what it keeps and I decide”. Awareness of the impacts is what makes the habit last. To see where you stand on your other digital habits, take the digital resilience self-assessment.
Summary
- Confidentiality is about who can read what you have written, uploaded or allowed it to remember: history, memory, documents, connected apps.
- Integrity is about the accuracy of what the assistant knows about you: a wrong memory, or one altered by outside content, changes the answers.
- Availability is about access to your conversations and connected services: a hijacked account, an auto-delete or an app disconnected in a hurry can take it away.
An assistant that has never been checked does not produce just one impact: it opens the door to all three, especially when it is connected to other services.
One thing to do today. Open the memory of the assistant you use most, if the service shows it, and read it. Mark the lines you do not recognise or that concern other people: they are the first to delete. It takes five minutes, and shows you with your own eyes what it carries from one conversation to the next.
Related content
- Check your AI assistant’s data settings — the recommendation this post builds on
- What sharing with an AI assistant can lead to — from technical impacts to concrete effects on work, money and relationships
- Signs your AI assistant remembers too much — how to notice that the assistant keeps more than you thought
- How to delete AI chat history and memory — what to do, in the right order
- Social login risks — the related topic of services that reach your account through a linked sign-in
Related resources
Short reads from the Resources section, for anyone who wants to focus on a single aspect:
- AI Chatbot Privacy: Where Your Conversations End Up
- How to Use AI Chatbots Anonymously, and Where It Stops
Start with the first step: the Cyber Welfare Programme guides you free of charge, one recommendation at a time.




