Helen works in an architecture practice in Milan. For a few months she has been using two artificial intelligence assistants with her personal accounts: one on her phone, to translate emails to a German supplier and outline a talk for a conference; the other on her laptop, connected to her email and calendar to plan a weekend in Lisbon with her daughter, who is fourteen.
One evening, as she asks the second assistant to suggest a flight time, the answer begins: “Since you take your daughter swimming on Saturday mornings…”. Helen does not remember telling it; perhaps she did, weeks earlier, in another chat. What does that assistant remember about her, and who decided? Nobody, really: the settings she found on day one were still in charge.
This recommendation, R38 of the Cyber Welfare Framework, invites you to do what Helen did that evening: check your AI assistant’s data settings and make, one by one, four decisions that the service keeps separate. It does not ask you to stop using AI. It asks for a fifteen-minute check when you start using an assistant, and then at least once a year, or when the service announces a change.
What this recommendation says
Recommendation R38 states that what your AI assistant keeps, remembers, uses to improve itself and reaches outside the chat should match choices you have made.
A few terms to clear up straight away:
- an AI assistant (or chatbot) is a service you write or speak to in everyday language and that replies by generating text, images or actions;
- chat history is the list of past conversations saved in your account so that you can pick them up again;
- memory is the feature, available on many services, that lets the assistant remember details about you from one conversation to the next, even when you do not repeat them;
- training is the process by which an AI model learns from very large amounts of text; some services may also use users’ conversations to improve future versions;
- connected apps (also called connectors) are the permissions you give the assistant to read or use other services of yours, such as email, calendar or online file storage.
In practice, it comes down to four decisions, taken one at a time:
- history: how long to keep your conversations, and which ones to delete;
- memory: what the assistant may remember about you, and what it should forget;
- training: whether your chats may help improve the model, often via a dedicated switch;
- connected apps: which ones stay connected, and with which permissions.
There is one extra tool: the temporary chat, offered by several services, which stays out of your history and does not update the memory. It suits sensitive questions, but it does not make you anonymous: the service still receives what you write and your connection.
What it is not. It is not a call to stop using AI, nor a guide to what to type into a chat: for that there is the resource on what not to share with AI chatbots, and how far you can stay unrecognised is covered by the resource on how to use AI chatbots anonymously. The general review of what your accounts keep is the subject of the recommendation on reviewing what your accounts keep about you: R38 is its specific case for assistants. “Sign in with…” buttons are covered by the recommendation on limiting social login, and app permissions on your phone by the recommendation on checking what your apps can do.
Where it applies. To the assistants you use with a personal account: in the browser, in phone apps, inside other programs. For work or school accounts, the organisation decides the settings: if in doubt, ask them.
Why it matters
An assistant is useful because you give it context, and that context gradually becomes an archive: a handy notebook full of things you do not remember writing.
The four functions run independently, often with settings on different pages:
- history keeps your chats until you delete them or the period you chose, if any, runs out;
- memory holds details separately from the chats: on at least one of the main services, according to its documentation, deleting a conversation does not necessarily delete what the assistant remembered from it;
- training, on many services, can be turned off, but the choice works going forward: one provider states explicitly that data already included in training that has started stays there;
- connected apps stay active until you disconnect them, and what the assistant shares with them is also subject to those services’ own rules, as the documentation of another large provider points out.
There is also a little-known exception: on some services, if you rate an answer with a thumbs up or down, the rating can carry the conversation with it, even when training is turned off.
In the EU and the UK, the GDPR (the UK GDPR in the UK), the data protection regulation, includes the right to object (Article 21): when an organisation uses your data on the basis of its own “legitimate interests”, meaning a purpose of its own that does not require your consent, you can ask it to stop. The European Data Protection Board (EDPB), in its Opinion 28/2024 of 17 December 2024 on AI models, notes that in these cases the right to object applies and should be ensured. The UK regulator, the Information Commissioner’s Office (ICO), commenting in September 2024 on a large platform’s use of user data for AI training, said the public must be able to trust that their privacy rights will be respected from the outset.
The practical effects show up in four areas:
- an archive that says a lot: projects, journeys, people, worries, all in one place;
- a profile you did not choose: memory pieces together scattered details, sometimes wrongly;
- uses that go beyond the answer: training puts conversations to a further purpose;
- open doors to other services: a forgotten connected app gives access to your email or your files.
| Benefit of the four decisions | Why it matters |
|---|---|
| A history that lasts as long as you choose | The archive stays in proportion to what you need |
| A memory you know | You know what the assistant carries from one chat to the next, and you can correct it |
| A deliberate choice on training | Your conversations help improve the model only if you decide so |
| Only the connected apps you use | Fewer services within reach, fewer permissions to remember |
| A date for the next check | Settings change with updates: a regular check keeps them in line |
A concrete example
Back to Helen. The next day, in fifteen minutes of her lunch break, she opens the settings of both assistants.
In the history of the first she finds more than two hundred conversations, some from a year ago. She deletes the ones she no longer needs and, where the service allows it, chooses an auto-delete period shorter than the default.
Then she opens the second assistant’s memory. There are about twenty lines: the name of her daughter’s school, the Saturday swimming, the fact that she is looking for a flat in another neighbourhood, and one thing that is wrong, that she works “in property”. She deletes the lines about her daughter and the wrong one, and keeps the ones about her job, which help with translations.
On training she makes two different choices: she turns it off on the first assistant, where she pastes work emails, and leaves it on for the second, which she uses for trips and ideas, because she is happy to contribute. She knows the choice covers future chats.
Finally she looks at the connected apps. As well as email and calendar, she finds a file storage service connected months earlier for a trial and then forgotten. She disconnects it. For her more sensitive messages, she will use the temporary chat.
She leaves the practice’s assistant alone: the organisation manages it. The way she uses her assistants has not changed; now she knows what they keep about her, and why.
When to apply it
There are moments when the check makes an obvious difference.
- When you start using an assistant. Before the first important question: the default settings apply from day one.
- At least once a year, on a set date. A day that is easy to remember turns the check into a habit.
- When the service announces a change. An email about new terms or a new feature, such as a wider memory, is the right reminder.
- When you connect a new app. Ask yourself which permissions you are granting, and for how long.
- After a sensitive conversation. Delete what you would rather not keep, and check the memory.
How to apply it
You do not need to know every service’s menus, which change often: you need a compass and a short sequence to repeat.
- List the assistants you use with a personal account. Leave aside any work or school ones.
- Open the settings and look for the right words. Names vary between services and versions: look for items such as “data”, “privacy”, “activity”, “memory”, “personalisation” or “improve”. If something is missing, search the official help pages.
- Decide on history. Delete the conversations you do not need and, if available, choose the auto-delete period.
- Read the memory. If the service shows what it remembers about you, go through it line by line: correct what is wrong, delete what you do not want kept, or turn the feature off if you prefer.
- Decide on training. Leaving it on, deliberately, is as legitimate a choice as turning it off. Remember that it usually covers future conversations, and that on some services a thumbs rating can share the whole chat.
- Check the connected apps. Disconnect the ones you no longer use and, for those that stay, look at their permissions: read, write, send.
- Try the temporary chat. If the service offers it, check in the help pages how long it is still kept, and decide which questions to use it for.
- Note your choices and the date. One line per service. A decision written down is easy to revisit at the next check.
If a service has no training setting, or you are not sure what it covers, you still have the right to object: you can write to the service and state that you object. Where the main services keep the setting today, with its current names, is explained in the resource on how to opt out of AI training. How to delete conversations and memories in the right order is in the guide on how to delete AI chat history and memory; the controls available, with their strengths and limits, are described in the post on AI assistant privacy controls.
Common mistakes to avoid
- Thinking one setting decides everything. History, memory, training and connected apps are separate controls: switching one off does not switch off the others.
- Believing that turning training off erases the past. On many services it covers new conversations; what has already been used stays where it is.
- Deleting the chat and forgetting the memory. A remembered detail can outlive the conversation you wrote it in.
- Treating the temporary chat as anonymous. It stays out of your history, but the service receives it and may keep it for a short period.
- Leaving trial apps connected. A connection made out of curiosity works until you remove it.
- Changing a work account’s settings yourself. The organisation decides there.
How this connects to the Cyber Welfare Framework
R38 brings privacy into a tool many people use daily without opening its settings. It separates what the assistant does for you from what it does with your data, and hands back to you the choice on each of the four functions.
| Pillar | How this contributes |
|---|---|
| Skills | Knowing how to find the data settings, read the memory, decide on training and disconnect an app |
| Awareness | Understanding that history, memory, training and connected apps are separate decisions, and that training is turned off going forward |
| Secure Behaviour | Repeating the check when you start using an assistant, at least once a year and after announced changes |
Digital maturity levels.
- FL1 — Basic. You use an assistant with the settings you found on day one, without ever having opened them. It is the most common starting point, not a fault.
- FL2 — Beginner. You have opened the data settings and made the four decisions for the assistant you use most.
- FL3 — Autonomous. You repeat the check on all your personal assistants, reread the memory regularly and use the temporary chat when it helps.
- FL4 — Skilled. You know the right to object and can use it where a setting is missing; you weigh up each connected app’s permissions before granting them.
- FL5 — Expert-Guide. You help family, colleagues or students do the same check, calmly and without judging different choices.
R38 is the key step from FL1 to FL2, and it remains an active habit at every level after that.
How to check you are applying it properly
Three questions, to be answered honestly.
- Could I say what the assistant I use most remembers about me?
- Do I know whether my conversations help train the model, and did I choose that?
- Which apps are connected to my assistant, and when did I connect them?
Quick checklist
- ☐ I have listed the assistants I use with a personal account
- ☐ I have decided how long to keep my history and deleted the chats I do not need
- ☐ I have reread the memory and deleted what I do not want kept
- ☐ I have chosen, deliberately, whether my chats may be used for training
- ☐ I have disconnected the apps I no longer use
- ☐ I have noted my choices and the date of the next check
If a box stays empty, you already have your next step. For a more structured measure of your level, take the digital resilience self-assessment.
In short
An AI assistant can keep your conversations, remember details about you, use chats to improve the model and reach the apps you connect: four separate functions, with settings that are often left unopened after day one.
The answer is a short check: decide how long to keep your history, reread the memory, choose on training knowing it works going forward, and disconnect unneeded apps. At the start, yearly, and when the rules change.
The assistant remains a useful tool. What it knows and keeps about you, though, is your decision.
Something to think about. If you could read today everything your assistant remembers about you, what would you want it to forget?
Explore this recommendation
This recommendation is the pivot of a content unit. Each post looks at a different aspect.
- AI chatbot privacy risks — what gets hit, in terms of the confidentiality, integrity and availability of conversations, memories and documents
- What sharing with an AI assistant can lead to — the concrete effects on the operational, financial, legal, reputational and personal levels
- How to delete AI chat history and memory — the mitigations, from the initial settings to an objection request
- Signs your AI assistant remembers too much — the indicators to check and what they mean
- AI assistant privacy controls — history, memory, training, temporary chats, connected apps, shared links and export, with their strengths and limits
- Prompt injection attacks — hidden instructions and the other ways chats, memories and connected apps are targeted
Related resources
Short reads from the Resources section, for anyone who wants to focus on a single aspect:
- Opt Out of AI Training: Check Your Chatbot Settings
- AI Chatbot Privacy: Where Your Conversations End Up
- What Not to Share with AI Chatbots: Keep, Generalise, Share
- 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.




