Responsible AI Use at the Workplace
Administrative staff handle the most sensitive information in any organisation. This part is the difference between AI making you faster and AI making you the subject of an investigation.
Confidentiality: what never goes into an AI tool
Assume anything you paste into a public AI tool has left your control. Free consumer accounts may use your input to improve their models, and a chat history is one shared screen away from being visible.
Never paste these
- Identity data: IC numbers, passport numbers, bank accounts, home addresses, phone numbers.
- Staff and student records: salaries, appraisals, medical information, disciplinary files, results.
- Unpublished decisions: appointments, promotions, terminations, tender outcomes, exam papers.
- Contracts, quotations and procurement documents under evaluation.
- Anything marked Confidential, Restricted, Rahsia or Terhad, no exceptions for convenience.
| Not safe | Safer alternative | |
|---|---|---|
| Drafting a letter about a specific person | Pasting the full name, IC and case details | Using placeholders: [NAME], [ID], [DATE], then filling them in offline |
| Summarising a confidential report | Uploading the file to a public chatbot | Using an organisation-approved tool, or summarising a redacted extract |
| Checking a staff list | Pasting the whole spreadsheet | Pasting only the column structure and asking for the formula or format |
| Recording a sensitive meeting | Letting a notetaker auto-join | Manual join only, with announced consent and a retention limit |
Practical safeguards
- Redact first, prompt second. Replace real details with placeholders before you paste anything.
- Turn off chat history / model training in your account settings where the tool allows it.
- Prefer the tool your organisation has approved over the one you like best.
- Never sign in to a work AI tool on a shared or public computer, and log out of the counter PC.
- If in doubt, ask your supervisor in writing once, then you have a documented position.
Consent and transparency
- Tell people when AI was used to record them, and let it be minuted.
- Do not present AI-generated text as another person's words.
- For external correspondence, the responsible officer's name still owns the content, review it as if you wrote every word, because officially you did.
- If your organisation has an AI policy, read it once and keep a copy in your own notes.
Recognising AI mistakes
AI errors are not random noise, they follow patterns. Once you know the patterns you catch most of them in seconds.
| What it looks like | How to catch it | |
|---|---|---|
| Hallucination | A confident, specific claim, a circular number, a clause, a statistic, that does not exist | Ask 'what is your source?' then verify in the actual document. No source, no use. |
| Outdated information | Old rates, superseded procedures, a policy replaced last year | Check the effective date on the real document, not the AI's summary |
| Plausible arithmetic | Totals and percentages that look right and are not | Recalculate in a spreadsheet. Always. |
| Over-agreement | It agrees with whatever you suggest, including wrong assumptions | Ask 'what is wrong with my assumption here?' and 'argue the opposite' |
| Bias | Gendered job assumptions, Western defaults, one-sided framing of an issue | Read the output as a stranger from your community would read it |
| Fabricated formatting authority | Official-looking reference numbers and letterhead lines it invented | Every reference number is typed by you, from the real record |
Ask the AI to check itself
Review your previous answer critically. 1. List every factual claim you made. 2. For each one, state whether it came from the document I provided, from general knowledge, or from inference. 3. Mark anything you are not confident about. 4. List what I should verify independently before using this.
This does not make the AI reliable. It does make its guesses visible, which is enough to protect you.
Your office AI checklist
Tick through this before anything AI-assisted leaves your desk. Print it and keep it beside your monitor.
0 of 9 ticked
Group activity: build your unit's rule
- 1In pairs, list the three most sensitive document types your unit handles.
- 2For each one, write a single sentence rule: what may and may not be given to an AI tool.
- 3Write down one safer workaround for each, redaction, placeholders, or an approved tool.
- 4Agree on one sentence your unit will add to minutes whenever an AI notetaker is used.
- 5Combine the results into a one-page 'AI use note' for your unit's shared folder.
You should end up with: A one-page AI use note written in your own office's language, not a generic policy.
Do
- Redact before you prompt.
- Verify every fact against the original document.
- Disclose AI-assisted recording to meeting participants.
- Keep a written record of the permission you were given.
Don't
- Paste personal data 'just this once' to save time.
- Cite an AI answer as the authority in official correspondence.
- Use a personal AI account for confidential organisational work.
- Assume silence from management means approval.