What Not to Share with AI Chatbots: Keep, Generalise, Share

Additional resource for the lesson “What to Share (and What Not To) with a Chatbot” — Online Privacy course

A chatbot answers better when you give it context, and that is exactly why it is worth pausing for a moment before you type. This resource is about what not to share with AI chatbots, and about what you can share comfortably: a simple rule of keep, generalise, share, and a short exercise to turn one real request into one that still works without giving away more than it needs.

A. Why this matters

A conversation with a chatbot feels private. You type in your own words, on your own screen, and the reply seems written just for you. That feeling is part of what makes these tools useful, and it is also a reason to slow down for a second.

Where your words go afterwards, and how long they are kept, is covered by another resource in this course. This one looks at the moment before: deciding what not to share with AI chatbots, and what to share, in the first place. That is the part you fully control.

The good news is that a careful request is not necessarily a worse one. Much of the help you want often depends on the kind of problem you have, not on who exactly is involved. Learning to tell the two apart is a skill, and it gets quicker with practice.

What the sources say

The UK’s National Cyber Security Centre (NCSC) explained, when chatbots first became widely used, that a query “will be visible to the organisation providing” the service. Its recommendation was “not to include sensitive information in queries to public LLMs”, meaning the large language models behind public chatbots. It added that stored queries could be leaked or, more likely, made public by accident.

The Office of the Privacy Commissioner of Canada gives individuals a short list of tips. It suggests you “limit the personal information that you share about yourself”, do not share sensitive or identifiable details, and do not share personal information about other people. It also suggests changing details, “for example, use aliases for names”. Get Safe Online, a UK public awareness service, puts it in similar terms: never reveal unnecessary personal or confidential information about yourself, your family, friends or finances.

The providers themselves say something similar. In its Gemini Apps Privacy Hub, Google explains that human reviewers look at some conversations, and asks: “Please don’t enter confidential information that you wouldn’t want a reviewer to see.”

At work, the question also belongs to your employer. The Government of Canada’s guidance for its own staff, for example, tells them to “follow your departmental guidance on generative AI” and never to put personal or protected information into public tools.

B. Key concepts

Five ideas that help you decide what goes into a chatbot message.

Keep, generalise, share

A simple rule with three outcomes for every detail in your request. Keep it: it stays with you and never goes in. Generalise it: you replace it with something less specific that still does the job, such as “a colleague” instead of a name. Share it: it is safe and useful, so in it goes.

Why it matters to you: Much of the work happens in the middle option. Generalising lets you get the same help without putting a real person, account or place into the conversation.

Personal data, including other people’s

Personal data is any information that relates to an identified or identifiable living person. The European Commission adds that separate pieces of information can count too, when together they point to one particular person. In a chatbot request, that includes the people you are writing about, not only you.

Why it matters to you: Other people have not agreed to be part of your conversation. When you generalise their details, you protect their privacy as well as yours.

Details that need extra care

UK data protection law treats some information as special category data. The ICO’s list includes health, ethnic origin, religious or philosophical beliefs, political opinions, trade union membership, sex life and sexual orientation. Its guidance explains that this type of data “needs to be treated with greater care”.

Why it matters to you: The ICO explains that using this kind of data is more likely to expose someone to discrimination, so it matters more if it travels further than intended. If your request needs them at all, describe the situation in general terms, and keep other people’s out completely.

The identifying combination

A request can identify someone even with no name in it. “My colleague, the only music teacher at a small school in a named town” describes one person. Generalising means looking at the whole message, not word by word.

Why it matters to you: The European Commission notes that data which has been pseudonymised, with names replaced by codes or aliases, is still personal data if it can be used to re-identify someone. Only details that no longer point to anyone are truly anonymous.

Secrets that do not belong in any message

Some things have no place in a chatbot at all: passwords, one-time codes, recovery codes, bank card and account numbers, passport or national insurance numbers. No answer you need depends on them.

Why it matters to you: A chatbot can explain how to reset a password without ever seeing yours. About to paste a secret? That is the moment to stop.

C. A practical example: Sara’s planning evening

Sara is 28 and teaches at a secondary school. On a Sunday evening she opens a chatbot to get through her week faster. She has four requests in mind.

• A set of quiz questions on the water cycle for Monday’s lesson.
• A reply to a parent who has written, quite upset, about a homework policy. Her first idea is to paste the whole email, with the parent’s name, the class and the school’s name.
• A note to cover a colleague’s lessons, explaining that the colleague is off for a medical procedure.
• Help with the school’s new online reporting form, which keeps rejecting her login. She is about to paste her username and password so that the chatbot can “see what is wrong”.

What changed when Sara applied the rule

Sara stopped before the second request and sorted her details into keep, generalise and share.

• The quiz questions she shared as they were. Nothing in them is about anyone.
• For the parent’s email she generalised: “A parent is upset that homework is set over the weekend. Help me reply politely, explaining the policy and offering a meeting.” She removed the name, the class and the school, and summarised the email instead of pasting it. The reply she got was just as useful.
• For the cover note she kept the medical detail entirely. “A colleague is absent this week” was all the chatbot needed. The reason for the absence is her colleague’s information, not hers to pass on.
• The login she kept. She asked instead, “What are common reasons a school web form rejects a valid login?”, and then contacted the school’s IT support, the people who can actually look at her account.

Before the next school week, she also checked one thing she had never looked at: whether her school has a policy on using AI tools for work. It does, and it names the tool staff are allowed to use. The habit now takes her a few seconds per request.

D. Try it yourself: rewrite one real request (10 minutes)

Four short steps, on paper or in a note on your phone. Use a real request you have made, or are about to make, to a chatbot.

Step 1 — Write the request as you would naturally type it (2 minutes)

• Choose something ordinary: an email to draft, a problem at work, a letter, a plan.
• Write it the way you would if you were not thinking about privacy at all, with all the details.
• This is your starting point, not a test.

Step 2 — Mark every detail K, G or S (3 minutes)

• Underline each name, place, date, number, organisation and personal circumstance.
• Next to each one, write K (keep), G (generalise) or S (share).
• Ask two questions for each detail: does the answer depend on it, and is it mine to share?
• Anything that is a password, code, or account or ID number is always K.

Step 3 — Rewrite the request (3 minutes)

• Replace each G with a general term: “a colleague”, “a client”, “a city in the north”, “last month”.
• Take out every K completely. If it seemed essential, describe the problem instead of the detail.
• Read the new version as a whole. Could someone who knows you work out who or what it is about? If so, generalise one more detail.

Step 4 — Check the context and decide (2 minutes)

• If the request is about work, ask yourself whether your employer has a policy on AI tools, and whether this tool is the one it allows. If you do not know, find out before sending anything work-related.
• Compare the two versions. Would the rewritten one still get you a useful answer? You will often be able to tell straight away.
• Decide: send the rewritten version, adjust it, or keep this one away from the chatbot altogether. All three are good outcomes, because you chose.

E. Videos, articles and further resources

Institutional sources first, then a public awareness service and official platform documentation. All in English.

Institutional sources

National Cyber Security Centre (UK) — ChatGPT and large language models: what’s the risk?
Why queries are visible to the provider, and the recommendation not to include sensitive information in them.
https://www.ncsc.gov.uk/blog-post/chatgpt-and-large-language-models-whats-the-risk

Office of the Privacy Commissioner of Canada — Your privacy and AI chatbots
Short tips for individuals: limit what you share, change identifying details, leave other people’s information out.
https://www.priv.gc.ca/en/privacy-topics/ai-technology-and-innovation/artificial-intelligence/ai-chatbots_ind/

Information Commissioner’s Office (UK) — What is special category data?
The kinds of personal data, such as health, that UK law says need greater care.
https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/what-is-special-category-data

European Commission — Data protection explained
What counts as personal data, and why pseudonymised data still counts while truly anonymous data does not.
https://commission.europa.eu/law/law-topic/data-protection/data-protection-explained_en

Using AI at work

Government of Canada — Generative AI in your daily work
An example of an employer’s guidance for staff: what never to put into public tools, and following internal rules.
https://www.canada.ca/en/government/system/digital-government/digital-government-innovations/responsible-use-ai/generative-ai/generative-ai-your-daily-work.html

Public awareness and platform documentation

Get Safe Online — AI: Artificial Intelligence
Everyday advice on using AI tools, including not revealing unnecessary personal or confidential information.
https://www.getsafeonline.org/personal/articles/artificial-intelligence/

Gemini Apps Help — Gemini Apps Privacy Hub
A provider’s own explanation of human review, and its request not to enter confidential information.
https://support.google.com/gemini/answer/13594961?hl=en-GB

Links checked in October 2026. If an address changes, searching for the title on the organisation’s site usually finds it.

F. The Cyber Welfare Framework: Skills, Awareness, Secure Behaviour

This resource sits on the Secure Behaviour pillar at level FL2 – Beginner, because its aim is a small habit repeated before every request, not a setting changed once.

Skills

• Sorting the details of a request into keep, generalise and share.
• Rewriting a request so that it keeps its meaning without names, numbers or identifying combinations.
• Recognising the kinds of data that need greater care.

Awareness

• Understanding that a chatbot conversation is not a private notebook, even when it feels like one.
• Recognising that other people’s information is theirs, not yours to share.
• Knowing that at work, your employer’s policy is part of the decision.

Secure Behaviour

• Pausing before you paste: reading the whole message, not only the question.
• Never putting a password, code or ID number into a chatbot.
• Checking your workplace’s rules before using an AI tool for work tasks.

G. Questions to sit with

• Which details do you tend to include out of habit, even when the answer does not depend on them?
• If the person you are writing about read your request, would they be comfortable with how you described them?
• Is there a kind of question you would rather keep away from chatbots altogether, and why?

H. What to do now

The recommendations (R) and security measures (MS) from the Cyber Welfare database that apply most directly here.

1. Before you type

The work here is a habit, not a setting, so no single recommendation covers it: read the whole request once before sending, generalise names and places, and leave other people’s sensitive details out. At work, find your organisation’s AI policy before using a chatbot for work tasks.

Minimum commitment: before your next chatbot request, take out one detail that the answer does not depend on.

2. Secrets stay out of the conversation

• R2 — Store your passwords safely, in a password manager: that is where a password belongs, never in a chat window, even when you want help with a login.
• R1 — Use a different password for each account: if one ever does end up somewhere it should not, it opens one door, not many.

Minimum commitment: if you have ever pasted a password into a chatbot, change that password the next time you have a few calm minutes.

In short

• A chatbot conversation feels private, but what you type is no longer only yours to manage.
• For every detail, choose: keep it, generalise it, or share it. Many requests need the shape of the situation, not the names.
• Other people’s information, and anything health-related, deserves extra care. Passwords, codes and ID numbers never go in.
• Knowing what not to share with AI chatbots is not about using them less. It is about deciding, before you type.

Related resources in this course

• AI Chatbot Privacy: Where Your Conversations End Up
• Opt Out of AI Training: Check Your Chatbot Settings
• Family AI Rules: Agreeing How AI Is Used at Home and at School

Discover more companion resources from the online courses of the Protect Your Digital Privacy programme.