Personal and confidential information: what never goes in
- Video
- 4 min
- Reading
- 13 min
- Exercise
- 12 min
- Total
- 28 min
Educational content, not legal advice.
In this lesson you will
- Recognize personal information and confidential business information in everyday work material.
- Apply the never list, and tell an approved tool from an unapproved one.
- Reduce what you paste by replacing names and cutting details the task does not need.
- Respond correctly when information went into a tool by mistake.
Video lesson
Read the transcript
Welcome back to AI Literacy at Work. Lesson three: personal and confidential information, and what never goes into an AI tool.
Two kinds of information need protecting. The first is personal information: anything about a person who can be identified. A name, a phone number, a unit number with a complaint, a photo, a voice recording. Even a photo of a kitchen can be personal, if it shows who lives there. The second is confidential business information: contracts, prices, plans, passwords and legal advice. Both can leave your organization the moment you paste them.
Some things never go into any AI tool, even an approved one, unless a process built for it says so. Passwords, access codes and keys. Government ID numbers, like a social insurance number. Bank and card numbers. Health information. And full files about a person, like a rental application or a disciplinary file. Keep this list close. It covers most of the damage. Your organization’s list may be longer. If it is, theirs wins.
For everything else, ask two questions before you paste. First, is it on the never list? If yes, stop. Second, is the tool approved by your organization for this kind of work? A free app, a browser extension, or a meeting bot you added yourself is not approved, however useful it is. The same chat assistant on a personal account is not approved either. If both answers are fine, paste only what the task needs.
That last step matters most. Replace names with Tenant A and Tenant B. Leave out the unit number if the tool does not need it. Cut the parts of the document that are not part of the question. But take care: a few small details together can still point to one person. When in doubt, describe the situation instead of pasting the document.
Here is why the law cares. Where PIPEDA applies, personal information has to be protected by safeguards that suit how sensitive it is. PIPEDA Sch.1 cl. 4.7 Where it applies, the organization also stays responsible for personal information it hands to a third party for processing. PIPEDA Sch.1 cl. 4.1, 4.1.3 And in Quebec, a business has to take security measures that are reasonable given how sensitive the information is, and given its purpose, quantity, spread and medium. P-39.1 s.10 When you paste into an unapproved tool, nobody has checked how that third party protects it.
Recap: never paste the never list. Use approved tools, and paste only what the task needs. If you made a mistake, report it fast. Lesson five shows how. In the next lesson: using AI responsibly with colleagues and customers.
Two kinds of information to protect
Most of the serious mistakes people make with AI at work are not about wrong answers. They are about what went in. Two kinds of information need care.
Personal information is information about a person who can be identified. In PIPEDA, Canada’s federal private sector privacy law, personal information means information about an identifiable individual. PIPEDA s.2(1) (personal information, commercial activity, organization, federal work, undertaking or business) Quebec’s private sector law describes it as any information that relates to a natural person and directly or indirectly allows that person to be identified. P-39.1 s.2 In plain words: if it is about a person, and someone could work out who, treat it as personal.
That covers more than names. A unit number with a noise complaint, a photo of a damaged kitchen with family pictures on the wall, a voice message, a work schedule, an email address: all can be personal information.
Confidential business information is anything your organization, a client or a partner expects to stay private: contracts, prices and bids, financial results, plans that are not public, security details, legal advice and passwords.
The moment you paste either kind into an AI tool, it leaves your hands. It goes to the tool’s provider, under the terms of your account (lesson 1).
The never list
Some information never goes into any AI tool, approved or not, unless your organization has built a specific process for it and told you so.
| Never paste | Why |
|---|---|
| Passwords, access codes, keys, alarm and door codes | Anyone who sees them can get in |
| Government ID numbers: social insurance, health card, driver’s licence, passport | Used for identity theft, very hard to undo |
| Bank account and payment card numbers | Direct financial harm |
| Health information, including sick notes and accommodation requests | Highly sensitive, harm if exposed |
| Full files about a person: a rental or job application, a credit report, a disciplinary or legal file | Many sensitive details at once |
| Information about children | Extra sensitive |
| Legal advice and documents about a dispute | Highly sensitive; ask the person handling the dispute first |
| Anything a contract says to keep confidential | You may breach the contract |
Your organization’s list may be longer. Theirs wins.
Approved and unapproved tools
An approved tool is one your organization has reviewed and agreed to use for certain work, usually on a business contract. The review typically looks at whether the provider stores your text and for how long, whether it uses your text to train its models, who at the provider can see it, where it is processed, and whether your organization can control and delete it.
An unapproved tool is anything nobody has reviewed for work use, even if it is popular and useful:
- free or personal accounts of chat assistants, including the same product your organization uses on a business plan;
- browser extensions that "help you write" and can read every page you open;
- meeting bots that join calls and record them;
- phone apps that transcribe or translate;
- free trials of new AI features.
If a tool is not on your organization’s approved list, treat it as unapproved. If you think a new tool would help, ask for it to be reviewed (lesson 6). Do not start using it while you wait for the answer.
Approved does not mean anything goes
An approved tool lowers the risk. It does not remove it. Even in an approved tool:
- the never list still applies;
- paste only what the task needs;
- the output still needs checking (lesson 2);
- the use has to fit the purpose the information was collected for.
That last point matters. Where PIPEDA applies, when an organization wants to use personal information for a purpose it did not identify before, it identifies the new purpose before using it and, unless the law requires that use, gets the person’s consent. PIPEDA Sch.1 cl. 4.2, 4.2.3, 4.2.4 Using tenant emails to build a marketing profile, for example, would be a new purpose unless it was identified at collection, not a quick AI task.
Paste less: minimize before you paste
Most tasks need far less information than people paste. Before you paste:
- Replace names with placeholders: Tenant A, Applicant 2, Contractor X.
- Remove numbers the task does not need: unit, phone, account, date of birth.
- Cut the parts of the document that the question is not about.
- Describe instead of paste: "a tenant with a long history of late payments asks for a payment plan" is often enough.
Take care: removing names is not the same as making something anonymous. A few small details together, such as "the only wheelchair user on the third floor of the Spring Street building", can still point to one person.
Why the law cares
In Canada, where a privacy law applies, it also applies when AI tools touch personal information. OPC and FPT privacy authorities, Principles for responsible, trustworthy and privacy-protective generative AI technologies (December 7, 2023) Three rules matter most to everyday users.
- Where PIPEDA applies, personal information has to be protected by security safeguards that suit how sensitive it is. PIPEDA Sch.1 cl. 4.7
- Where PIPEDA applies, the organization stays responsible for personal information it transfers to a third party for processing, such as an AI tool’s provider. PIPEDA Sch.1 cl. 4.1, 4.1.3
- In Quebec, a business has to take the security measures needed to protect personal information that are reasonable given its sensitivity, its purpose, its quantity, its distribution and the medium it is stored on. P-39.1 s.10
PIPEDA’s safeguards principle also says organizations shall make their employees aware of the importance of keeping personal information confidential. PIPEDA Sch.1 cl. 4.7.1 to 4.7.5 This lesson is part of that.
Your organization cannot protect information that staff have already sent to an unreviewed provider. That is why the never list and the approved list exist.
If you pasted something you should not have
It happens. What matters is what you do next.
- Stop. Do not paste more, and do not share the output.
- Do not try to hide it. Do not delete the chat yet: the person handling your report may need it, and deleting does not undo what the provider already received.
- Report it quickly to your manager, privacy officer or IT, as your organization’s rules say. Lesson 5 has an incident report form.
- Write down the facts: which tool, which account, what information, when.
Where PIPEDA applies, a breach that creates a real risk of significant harm is reported as soon as feasible after the organization determines it occurred. PIPEDA s.10.1(1)-(3), (6) It can only do that once it knows. A fast, honest report gives it the best chance to limit the harm.
A rental application in a free chatbot
Fictional organization. Quillfield Property Management Inc. is invented for this course. Any resemblance to a real company is a coincidence.
On a busy Friday, a leasing agent at Quillfield had six rental applications to assess. To save time, she opened a free chat assistant on her phone, on her personal account, and pasted in one full application: name, date of birth, social insurance number, employer, salary, two bank statements and the names of two references. She asked: "Does this person earn enough for a 1,650 dollar apartment?"
The answer was useful. Ten minutes later, during a quick refresher on this course, she realized what she had done.
What she did next.
- She stopped and did not paste the other five applications.
- She called the leasing team lead and the privacy officer, and filled in the incident report form from lesson 5: which app, which account, what information, when.
- She did not delete the chat before talking to the privacy officer, who asked her to take a screenshot of the account’s privacy settings first, and then delete it.
What the privacy officer did. He recorded the incident, checked the app’s terms (chats kept, and used for training unless switched off), and assessed the risk to the applicant with the compliance manager. The social insurance number and bank details made the information highly sensitive. Quillfield followed its breach procedure, which is covered in lesson 6.
What Quillfield changed.
- A short "never paste" card went on every leasing desk. It is the template for this lesson.
- The approved chat assistant got a saved instruction for income checks: "Enter monthly income and monthly rent only. No names, no ID numbers, no bank details."
- Leasing agents were reminded that an affordability check needs two numbers, not a whole file.
The lesson. The agent needed two figures: monthly income and monthly rent. She pasted fifty pieces of personal information to get them. Minimizing first would have made the task both safe and faster.
Sort what can go in
You work in tenant services at a property management company. Your organization has approved one chat assistant on a business account. For each item, decide: never (no AI tool), approved tool only, minimized (and say what you would remove), or any tool (public, nothing sensitive).
- A tenant’s email asking for a payment plan, which mentions she lost her job and is seeing a doctor for stress.
- The building’s published visitor parking rules.
- The door entry code for the building at 40 Harbour Road.
- A list of 30 tenants with their phone numbers, to draft a reminder text.
- A draft newsletter about the new gym, with no names.
- A lawyer’s letter about a dispute with a former tenant.
Show the model answer
- Approved tool only, minimized, or no AI at all. Remove her name and unit. Remove the job loss details and everything about her health: health information is on the never list. "A tenant asks for a three month payment plan after a drop in income" is enough to draft a reply. A person reads the reply carefully before it goes out.
- Any tool. It is published and contains nothing personal.
- Never. An access code lets anyone into the building.
- Approved tool, minimized. The tool does not need the list at all. Ask it to draft the message, then send it from your texting system to the list. Phone numbers stay out.
- Any tool, though the approved one is best. No names, nothing confidential. Still check facts like opening hours.
- Never. Legal documents about a dispute are on the never list. Ask the person handling the dispute.
Why this is the strongest answer
The pattern behind the strongest answers is that the task almost never needs the sensitive part. A payment plan reply needs the request, not the medical details. A reminder text needs the wording, not the phone list. Once you ask "what does the tool actually need?", most items become safe to handle in the approved tool, and the few that are left (access codes, legal documents, health details) are clear cases for the never list.
What never to paste: a desk card
A one page card for every desk: the never list, the two questions before you paste, how to minimize, and what to do in the first hour if something went in by mistake.
Check your understanding
Sources cited in this lesson
- PIPEDA s.2(1) (personal information, commercial activity, organization, federal work, undertaking or business) official text current to 2026-09-30, checked on 2026-10-29
- P-39.1 s.2 official text current to 2026-10-15, checked on 2026-10-29
- PIPEDA Sch.1 cl. 4.2, 4.2.3, 4.2.4 official text current to 2026-09-30, checked on 2026-10-27
- OPC and FPT privacy authorities, Principles for responsible, trustworthy and privacy-protective generative AI technologies (December 7, 2023) official text current to 2025-05-06, checked on 2026-11-01
- PIPEDA Sch.1 cl. 4.7 official text current to 2026-09-30, checked on 2026-10-24
- PIPEDA Sch.1 cl. 4.1, 4.1.3 official text current to 2026-09-30, checked on 2026-10-24
- P-39.1 s.10 official text current to 2026-10-15, checked on 2026-11-01
- PIPEDA Sch.1 cl. 4.7.1 to 4.7.5 official text current to 2026-09-30, checked on 2026-10-29
- PIPEDA s.10.1(1)-(3), (6) official text current to 2026-09-30, checked on 2026-11-03