Lesson 4

Using AI at work responsibly

Video
3 min
Reading
13 min
Exercise
12 min
Total
28 min

Educational content, not legal advice.

In this lesson you will

  1. Decide when to tell colleagues and customers that AI was used, and what to tell them.
  2. Recognize decisions about people where a real human review is needed, and what makes a review real.
  3. Describe the main legal disclosure and review rules that apply to AI in Canada and the EU, at a user’s level.
  4. Avoid common intellectual property and likeness mistakes with AI generated content.

Video lesson

Length: 2:52
Read the transcript

Welcome back to AI Literacy at Work. Lesson four: using AI at work responsibly. Three habits: say when AI was used, keep a real person in decisions about people, and take care with other people’s work.

Start with one idea. You are the author of everything you send, file or decide, whoever drafted it. If an AI tool wrote the first version, the mistakes in it are still yours once you send it. That is not a reason to avoid AI. It is a reason to read what you send. Read it as if your name is on it, because it is.

Next, saying when AI was used. With colleagues, tell them when AI did a real part of the work, like an analysis or a summary they will rely on, and tell them what you checked. With customers, be open whenever they could be misled. A tenant chatting with a bot should know it is a bot. An image or a voice made by AI should not pass as real.

The law is moving the same way. In the European Union, the AI Act’s transparency duties in Article fifty apply from August second, twenty twenty six. AI Act Art. 50, Art. 99(4)(g), Art. 111(4), Art. 113, second paragraph And in Ontario, since January first, twenty twenty six, an employer with twenty five or more employees that uses AI to screen, assess or select applicants for a publicly advertised job has to say so in the posting. ESA 2000 s.8.4; O. Reg. 476/24 ss.1, 2(1)

Now, decisions about people. Renting an apartment, hiring, discipline, collections. Here a person stays in charge, and the review has to be real. Ask two questions. First, does a person really review the output, with the time, the information and the authority to change it? Second, can the person affected be heard? In Quebec, when a decision is based exclusively on automated processing of personal information, the person has to be told, can ask for the reasons, and can submit observations to a staff member who is able to review the decision. P-39.1 s.12.1 A warning sign to remember: if reviewers never disagree with the tool, the review is probably not real.

Last, other people’s work. Do not paste someone else’s confidential documents. Do not ask a tool to copy a competitor’s ad, or to imitate a real person’s voice or face. Check images for logos and look alikes. Before you use AI output in anything public, check the tool’s terms and your own policy. When in doubt about who owns something, ask.

Recap: you are the author. Say when AI was used, keep people in charge of decisions about people, and respect other people’s work. In the next lesson: AI and security.

Notes

You are the author

One idea runs through this lesson: you are the author of everything you send, file or decide, whoever or whatever drafted it. If an AI tool wrote the first version, its mistakes become yours once you use it. That is not a reason to avoid AI. It is a reason to read, check and own what you send.

Responsible use comes down to three habits:

  1. Say when AI was used, to the people who need to know.
  2. Keep a real person in charge of decisions about people.
  3. Take care with other people’s work and likeness.

Saying when AI was used: colleagues

Inside your organization, tell colleagues when AI did a real part of the work they will rely on.

SituationTell them?What to say
You used AI to fix the grammar in an emailNot needed
AI drafted a report or analysis, and you edited itYes"First draft by the approved AI assistant. I checked the figures against the ledger."
AI summarized a long document for your managerYes"AI summary. I checked the dates and the exceptions against the original."
AI suggested a ranking or a score that someone will act onYes, alwaysSay what the tool was, what you checked, and what you did not

The point is not to confess. It is to tell the next person how much checking has been done, so they can decide how much more to do.

Saying when AI was used: customers and the public

With customers, tenants and the public, the test is simple: be open whenever someone could be misled.

  • A person chatting with a bot should know it is a bot, not a staff member.
  • An image, a voice or a video made by AI should not pass as a real photo, a real person or a real event.
  • A letter or decision about someone should not hide that an automated tool decided it.

Your organization’s policy sets the details. The law is also moving this way.

  • In the European Union, providers of AI systems that interact directly with people design them so that people are informed they are interacting with an AI system, unless that is obvious. AI Act Art. 50(1)
  • In the European Union, deployers of an AI system that generates or manipulates image, audio or video content that is a deep fake disclose that the content was artificially generated or manipulated. AI Act Art. 50(3) and (4)
  • These Article 50 transparency duties apply from August 2, 2026. AI Act Art. 50, Art. 99(4)(g), Art. 111(4), Art. 113, second paragraph
  • In Ontario, since January 1, 2026, an employer that publicly advertises a job and uses AI to screen, assess or select applicants includes a statement in the posting that discloses it, unless the employer has fewer than 25 employees on the day the posting goes up or another exclusion applies. ESA 2000 s.8.4; O. Reg. 476/24 ss.1, 2(1)

If you are not sure whether to disclose, disclose, or ask.

Decisions about people: keep a real person in charge

Some uses of AI are routine: drafting, summarizing, translating general notices. Others shape what happens to a person: who gets an apartment, who gets an interview, whose account goes to collections, who is disciplined. For those, a person stays in charge, and the review has to be real.

A real review has three parts:

  • Time. The reviewer has enough time to look at each case, not seconds per file.
  • Information. The reviewer can see why the tool scored or flagged the person, and the person’s own file.
  • Authority. The reviewer can disagree with the tool, and does when the facts say so.

A warning sign: if reviewers almost never disagree with the tool, the review may be a rubber stamp.

The law adds its own rules for decisions about people.

  • In Quebec, when a business makes a decision based exclusively on automated processing of personal information, it informs the person no later than when it informs them of the decision. P-39.1 s.12.1 On request, it tells them the personal information used, the reasons and the main factors and parameters behind the decision, and their right to have the information corrected. P-39.1 s.12.1 It also gives them the chance to submit observations to a staff member who is in a position to review the decision. P-39.1 s.12.1
  • Where the GDPR applies, a person has the right not to be subject to a decision based solely on automated processing that produces legal effects for them or similarly significantly affects them, subject to listed exceptions. GDPR Art. 22

You do not need to apply these rules yourself. You need to spot when AI is shaping a decision about a person, and make sure the process your organization set up is followed.

Respect for colleagues

Use AI on colleagues only through processes your organization has approved: no private tools to summarize a coworker’s messages, rate their performance or guess their mood. In the European Union, AI systems that infer the emotions of people in the workplace are banned, except for medical or safety reasons. AI Act Art. 5(1)(a) to (h); Art. 113, third paragraph, point (a) The Commission’s non binding guidelines limit that ban to biometric data, such as a voice or keystrokes. Commission Guidelines on prohibited AI practices, C(2025) 5052 final (July 29, 2025), paras. 5, 245, 251, 265

Other people’s work and likeness

Intellectual property questions can be complex, and this course does not give legal advice on them. A few habits keep you out of most trouble.

  • Do not paste other people’s confidential work into an AI tool: a client’s report, a partner’s plans, a supplier’s price list marked confidential.
  • Do not ask a tool to copy a competitor’s ad, a photographer’s picture, a song or a well known character, or to write "in the exact style of" a living creator for commercial use.
  • Do not imitate a real person. No cloned voices, faces or signatures of real people, colleagues included, without their clear permission and your organization’s approval.
  • Check generated images for logos, trademarks, text and faces that look like real people.
  • Check the tool’s terms and your organization’s policy on who can use what the tool produces, before you use it in anything public.
  • Credit your sources when the content came from somewhere you can name.

When in doubt about who owns something, or whether you can use it, ask your manager or the person who handles legal or marketing questions.

Worked example

The applicant score nobody questioned

Fictional organization. Quillfield Property Management Inc. is invented for this course. Any resemblance to a real company is a coincidence.

Quillfield’s tenant screening service gives each rental applicant a score from 1 to 100. For two years, leasing agents treated any score under 60 as a "no". Nobody had written that rule down. It had just become habit.

What surfaced it. A Montreal applicant with a steady job and good references was turned down. She asked why. The leasing agent could only say "the system scored you low". The applicant asked to speak to someone who could look at her file again. The agent did not know who that was.

What the team found when it looked.

QuestionFinding
Did a person really review?No. Agents spent seconds per file and had never gone against the score.
Did agents know why the score was low?No. The service showed the number only.
Was the applicant told a tool was involved?No. The letter said "after careful review".
Could she be heard?Not in practice. There was no named person to review the decision.

The privacy officer flagged one more point. For applicants in Quebec, a decision based exclusively on automated processing of personal information carries specific duties: inform the person, explain on request, and let them submit observations to a staff member who can review the decision. P-39.1 s.12.1 Treating the score as the decision came close to that line.

What Quillfield changed.

  • The score became one input, not the decision. A leasing agent now reviews every application below 60 against the file, and can approve it.
  • The screening service was asked to show the main reasons behind each score. Until it does, agents note their own reasons for each decision.
  • The rejection letter now says that a screening tool is used, and names a person who can review the decision on request.
  • Every quarter, the leasing team lead checks how often agents disagree with the score. "Never" is treated as a warning sign.

The lesson. No one at Quillfield decided to let a tool decide. It happened because nobody asked who was really deciding.

Scenario exercise · Allow 12 min

Three requests on a Monday

You are the team lead in tenant services. Three requests arrive on Monday morning. For each, say what you would do and why, in a few lines.

  1. A colleague wants to send tenants a newsletter with an AI generated photo of "our friendly maintenance team". The people in the photo do not exist.
  2. The accounts team wants an AI tool to pick which overdue accounts go to the collections agency, and send them automatically every Friday.
  3. A colleague wants to use an AI tool to read team chat messages and tell you who seems unhappy at work, "to help with retention".
Show the model answer
  1. Do not present invented people as the real team. Tenants would be misled into thinking these are the people who will come to their homes. Use a real photo with the team’s permission, or an illustration that is clearly not a photo. If an AI image is used for a general scene, label it.
  2. Keep a person in charge. Sending someone to collections is a decision about a person with real consequences. The tool can suggest a list, but a person reviews each account with time, the file and the authority to remove it. Tell affected tenants how to ask a person to look again. If any tenants are in Quebec and the decision would be based exclusively on automated processing, your organization’s process has to give them the information, and the chance to submit observations to a staff member who can review the decision, that Quebec’s law describes. P-39.1 s.12.1
  3. Decline and raise it. Using a private AI tool to judge colleagues’ moods from their messages is outside any approved process, and it reads private conversations for a purpose staff never agreed to. Talk to HR if retention is a concern. In the European Union, AI that infers emotions in the workplace is banned except for medical or safety reasons. AI Act Art. 5(1)(a) to (h); Art. 113, third paragraph, point (a) The Commission’s non binding guidelines limit that ban to biometric data, such as a voice or keystrokes, and list emotion inference from written text without biometric data as out of scope. Commission Guidelines on prohibited AI practices, C(2025) 5052 final (July 29, 2025), paras. 5, 245, 251, 265 So the reasons to decline here are privacy and your organization’s rules.

Why this is the strongest answer

Each request tests a different habit from this lesson. The photo is about not misleading people. The collections list is about keeping a real person in charge of decisions about people. The mood scanner is about respect for colleagues and approved processes. In all three, the strongest answer does not simply say no to AI. It keeps the useful part (an image, a suggested list, attention to staff wellbeing) and removes the part that misleads, decides without a person, or intrudes.

Template to download

Disclosure and human review checklist

A one page checklist for AI work that reaches other people: when and how to say AI was used, the three tests of a real human review, and the image, voice and copying checks before anything goes public.

Check your understanding

Scenario questions. Answer them all, then check: you will see why each option is right or wrong.

Question 1 of 5

You used an AI tool only to fix spelling and grammar in a routine email to a contractor. Do you need to tell the contractor AI was used?

Question 2 of 5

Which of these is the clearest sign that a human review of AI decisions is only a rubber stamp?

Question 3 of 5

A company in Ontario with 200 employees posts a public job ad for an Ontario based role. An AI tool screens the applications. What does Ontario’s employment standards law expect?

Question 4 of 5

A colleague wants to make a short video in which an AI copy of a coworker’s face and voice explains the new parking rules. The coworker said "sure, whatever" in the hallway. What is the best next step?

Question 5 of 5

Your team wants to publish an AI generated image of a busy street with shops for a newsletter. What is the most useful check before it goes out?

Sources cited in this lesson

  1. AI Act Art. 50(1) official text current to 2026-07-27, checked on 2026-11-01
  2. AI Act Art. 50(3) and (4) official text current to 2026-07-27, checked on 2026-11-01
  3. AI Act Art. 50, Art. 99(4)(g), Art. 111(4), Art. 113, second paragraph official text current to 2026-07-27, checked on 2026-11-01
  4. ESA 2000 s.8.4; O. Reg. 476/24 ss.1, 2(1) official text current to 2026-01-01, checked on 2026-11-01
  5. P-39.1 s.12.1 official text current to 2026-10-15, checked on 2026-10-24
  6. GDPR Art. 22 official text current to 2016-05-04, checked on 2026-10-26
  7. AI Act Art. 5(1)(a) to (h); Art. 113, third paragraph, point (a) official text current to 2026-07-27, checked on 2026-11-01
  8. Commission Guidelines on prohibited AI practices, C(2025) 5052 final (July 29, 2025), paras. 5, 245, 251, 265 checked on 2026-11-01