CYBER WELFARE

Protect your Digital Privacy

AI assistant privacy controls: what they do, and when to use them

Many artificial intelligence assistants already have, in their settings, the tools you need to decide what they keep, what they remember and what they can reach. You do not need to install anything. You need to know what each control does, what it does not do, and when it is worth using.

This post describes AI assistant privacy controls in general terms: history and auto-delete, memory, the training setting, temporary chats, connected apps, shared links and data export. Item names vary between services and versions, and often change with updates; here we describe them by what they do, so that you can recognise them on the service you use.

It is the technical side of the recommendation to check your AI assistant’s data settings. The practical sequence for using them, in the right order, is in the guide on how to delete AI chat history and memory.

How to read this list

The controls are split into four groups, like the phases of security:

  • prevention: they limit what the assistant keeps, remembers or reaches from now on;
  • detection: they show you what it already has;
  • response: they let you remove what you do not need;
  • governance: they govern how data leaves the assistant, who decides on accounts and which rights you can use.

For each control you will find the risk it reduces, its strengths, its limits, an example of use and its complexity. One principle applies to all of them: history, memory, training and connected apps are separate functions. A control usually acts on only one of them.

1. Prevention technologies

Training setting

This is the switch, available on many services, that decides whether your conversations may be used to improve the provider’s models. One large provider’s documentation explains that, with it turned off, new chats are not used for future training, while data already included in training that has started stays there.

  • Risk reduced: your conversations being put to a further purpose beyond the answer.
  • Strengths: usually a single step, and it does not change how you use the assistant.
  • Limits: it works going forward; on some services a thumbs rating can still share the whole conversation; on others the same setting also governs the history.
  • Example of use: you turn it off on the assistant where you paste work emails, and leave it on, deliberately, on the one you use for recipes.
  • Complexity: low.

Memory

This is the feature that lets the assistant remember details about you from one conversation to the next. On some services memory updates itself, on others it saves only what you ask; it can often be turned off altogether.

  • Risk reduced: personal details, and other people’s, building up into a profile that travels between chats.
  • Strengths: you decide the balance between convenience (not repeating yourself) and privacy.
  • Limits: it is separate from the history: on some services turning it off does not delete memories already saved, and deleting a chat does not necessarily delete what was remembered from it.
  • Example of use: you keep it on for work, but reread the memories every few months.
  • Complexity: low.

History auto-delete

Many services let you choose after how long conversations are deleted automatically, or not saved at all. The periods available are given in each service’s help pages.

  • Risk reduced: an archive that grows for years.
  • Strengths: it runs by itself, with nothing to remember.
  • Limits: it may delete a conversation you needed; it does not cover the memory.
  • Example of use: you choose a short period and copy elsewhere whatever you want to keep.
  • Complexity: low.

Temporary chat

This is a conversation that stays out of your history and, on many services, neither updates the memory nor is used for training.

  • Risk reduced: occasional or sensitive questions building up.
  • Strengths: it leaves your general settings alone; you use it only when needed.
  • Limits: the service still keeps it for a short period, given in its help pages, and receives it like any other chat: it does not make you anonymous.
  • Example of use: a question about a family matter you would rather not find again in a month.
  • Complexity: low.

Minimum permissions for connected apps

When you connect email, calendar or file storage, many services ask which permissions to grant: read, edit, send.

  • Risk reduced: actions in your name on other services, including on the instructions of outside content.
  • Strengths: the assistant does what you need, and nothing more.
  • Limits: permissions cannot always be separated; data shared with the other service is also subject to that service’s rules.
  • Example of use: you connect your calendar read-only so that it can suggest times.
  • Complexity: low to medium.

Two-step verification for the account

It is not a privacy control in the strict sense, but it protects everything else: it asks for a second proof, beyond the password, when someone signs in from a new device. The recommendation on protecting accounts with a second factor explains how to choose it.

  • Risk reduced: other people reaching your history, memory and connected apps.
  • Strengths: a single protection for all the account’s data.
  • Limits: not every service offers it.
  • Example of use: you turn it on the day you connect the assistant to your email.
  • Complexity: low.

2. Detection technologies

Memory page

This is the page, where the service offers one, that shows what the assistant remembers about you.

  • Risk reduced: wrong memories, memories about other people, memories that came from outside content.
  • Strengths: you see at a glance the profile the assistant is using.
  • Limits: not every service shows it in full; some also use past chats without listing them as memories.
  • Example of use: after an answer that “knew too much”, you open the page and look for the line responsible.
  • Complexity: low.

List of connected apps

It shows which services are connected to the assistant and, often, with which permissions.

  • Risk reduced: forgotten connections, permissions broader than needed.
  • Strengths: you know what the assistant can reach outside the chat.
  • Limits: it does not always show since when an app has been connected or what it has done.
  • Example of use: you find a file storage service connected for a trial and remove it.
  • Complexity: low.

List of shared links

Some services gather on one page the conversations you have shared through a public link.

  • Risk reduced: conversations readable by anyone with the link.
  • Strengths: you can revoke in one place everything you have shared over time.
  • Limits: not every service has one; it does not tell you who opened the link.
  • Example of use: you remove the link you sent a friend months ago to show an itinerary.
  • Complexity: low.

Sessions, devices and sign-in alerts

Many accounts show which devices they are signed in on and can send an alert when someone signs in from a new one.

  • Risk reduced: someone using the account without you noticing.
  • Strengths: it lets you act early.
  • Limits: it works only if you read the alerts; not every service sends them.
  • Example of use: you see a session from a computer you do not recognise and sign it out.
  • Complexity: low.

3. Response technologies

Deleting chats and memories

These are two separate commands: one removes conversations from the history, the other deletes details from the memory.

  • Risk reduced: conversations and details you do not want kept.
  • Strengths: immediate, from your point of view.
  • Limits: on many services deleted chats are removed from the provider’s systems after a certain time; on at least one of the main services, conversations already reviewed by human reviewers are kept even when you delete your activity.
  • Example of use: you delete the memory about your flat-hunting first, then the chat where you talked about it.
  • Complexity: low.

Disconnecting an app or reducing its permissions

It removes the assistant’s access to another service.

  • Risk reduced: actions and reading on email, calendar or files.
  • Strengths: reversible: you can reconnect it when needed.
  • Limits: it does not delete data already shared with the other service, which follows its own rules.
  • Example of use: after an action on your email that you did not ask for, you disconnect the app straight away.
  • Complexity: low.

Revoking a shared link

It makes a conversation you had shared private again.

  • Risk reduced: reading by whoever received or found the link.
  • Strengths: immediate.
  • Limits: it does not remove copies or screenshots others made in the meantime.
  • Example of use: you remove the link to a chat posted by mistake in a group.
  • Complexity: low.

Deleting the account

It deletes the account and, with it, history, memory and connections.

  • Risk reduced: everything the account keeps.
  • Strengths: you start again from scratch, or stop using the service.
  • Limits: on many services it does not touch what has already been used for training; you also lose what was useful.
  • Example of use: an account opened for a trial years ago and never used again.
  • Complexity: medium.

4. Governance technologies

Data export

Many services let you download a copy of your conversations and other account data.

  • Risk reduced: losing what you need when you delete chats or the account.
  • Strengths: you can see in black and white what the service keeps.
  • Limits: the file contains everything you have written: look after it like an important document and delete it when you no longer need it.
  • Example of use: you export before closing an account you no longer use.
  • Complexity: low to medium.

Work and school accounts

On these accounts the organisation decides the data settings, often under agreements that differ from personal accounts.

  • Risk reduced: the organisation’s data being handled under unsuitable rules.
  • Strengths: the choices are made by people who know the organisation’s risks.
  • Limits: they do not cover personal assistants, even if you use them for work.
  • Example of use: for a question about the settings of the practice’s assistant, you ask the IT team.
  • Complexity: managed by the organisation.

The right to object and other rights

It is not a technology, but it works like one. In the UK and the EU, the UK GDPR and the GDPR include the right to object (Article 21) when a service uses your data on the basis of its own legitimate interests; the EDPB’s Opinion 28/2024 notes that for AI models this right applies and should be ensured. You can also ask for access to your data and, in some cases, its erasure.

  • Risk reduced: uses of your data that the settings do not cover.
  • Strengths: it applies even where there is no switch.
  • Limits: it is best made in writing, so you keep a record, and it takes some patience: the reply is due within one calendar month, extendable in some cases; the service may refuse if it can show overriding grounds.
  • Example of use: you write to a service that offers no training setting.
  • Complexity: medium.

Shared rules for household devices

Not a technology, but a simple rule, agreed together, about who uses which account and when to use the temporary chat.

  • Risk reduced: history and memory read by others on a shared device.
  • Strengths: it needs no tools.
  • Limits: it has to be remembered, and revisited when devices change.
  • Example of use: on the family laptop, everyone uses their own account.
  • Complexity: low.

Comparison table

ControlGroupRisk reducedComplexityWhen to use it
Training settingPreventionFurther use of chatsLowWhen you start using an assistant
MemoryPreventionProfile travelling between chatsLowDecide at the start, review now and then
Auto-deletePreventionGrowing archiveLowAlways, with a chosen period
Temporary chatPreventionSensitive questions building upLowFor occasional questions
Minimum permissionsPreventionActions in your nameLow to mediumWith every new connection
Two-step verificationPreventionOthers signing inLowAlways, where available
Memory pageDetectionWrong or foreign memoriesLowAfter an answer that “knows too much”
Connected appsDetectionForgotten connectionsLowOnce a year and after trials
Shared linksDetectionChats readable by othersLowOnce a year
Sessions and alertsDetectionAccount used by othersLowWhen an alert arrives
Deleting chats and memoriesResponseData you do not want keptLowAfter a sensitive chat
Disconnecting an appResponseAccess to other servicesLowAs soon as it is not needed
Revoking a linkResponseReading by othersLowAs soon as it is not needed
Deleting the accountResponseThe whole archiveMediumFor abandoned accounts
ExportGovernanceLosing what you needLow to mediumBefore deleting
Organisation accountsGovernanceUnsuitable rulesOrganisationOn work and school accounts
GDPR rightsGovernanceUses the settings do not coverMediumWhere a switch is missing
Household rulesGovernanceShared devicesLowOn family devices

How to choose

  • If you use the assistant rarely: choose a short auto-delete period, decide on training and leave memory off if you do not need it.
  • If you use it every day for work: keep the memory, but reread it; connect only the apps you need, with minimum permissions; turn on two-step verification.
  • If you ask personal or sensitive questions: use the temporary chat, knowing that it does not make you anonymous.
  • If you share a device at home: use separate accounts, or the temporary chat, and agree a rule together.
  • If you use a work or school account: follow the organisation’s rules and ask them if in doubt.

The practical rule: the assistant keeps, remembers and reaches only what you have decided, and you check it when you start, once a year and when the service changes.

How this connects to the Cyber Welfare Framework

PillarWhat this content contributes
SkillsKnowing an assistant’s privacy controls and being able to recognise them on the service you use
AwarenessUnderstanding what each control does and what it does not, starting from the fact that history, memory and training are separate
Secure BehaviourChoosing the controls that suit the way you use the assistant and reviewing them regularly

Reference level: FL3 — Autonomous. This is the level at which you choose controls on purpose, rather than accepting the defaults, and extend them to every assistant you use. To measure your other digital habits, take the digital resilience self-assessment.

Conclusion

On many services, the controls you need to decide what an AI assistant keeps about you are already there: the training setting, memory, auto-delete, the temporary chat, the lists of connected apps and shared links. None of them solves everything on its own, and some, like export, move the risk rather than reducing it.

Used together, with this post’s practical rule, they hand back to you choices that the service, for convenience, had made on your behalf.

Related content

Related resources

Short reads from the Resources section, for anyone who wants to focus on a single aspect:

Start with the first step: the Cyber Welfare Programme guides you free of charge, one recommendation at a time.