Responsible Use of AI with Official Data
Saved in this browser only. See all revisions due
Key points
- Never upload confidential or personal data to public AI tools
- Deleting the chat does not undo the disclosure
- Use institution-approved tools and anonymise data where required
- Human in the loop: verify every rule number, figure and date
- The officer who signs remains accountable; AI said so is no defence
- Do not let AI decide recruitment, penalty or evaluation matters
- A good prompt gives role, context, task, format and constraints
What must not be uploaded
Classified, confidential and personal data must never be uploaded to a public or unapproved AI tool. This includes names, identification numbers, pay details, medical particulars and disciplinary case records.
Such data is governed by the Digital Personal Data Protection Act, 2023 and the rules made under it. Some government departments have issued internal instructions restricting the use of public AI tools on office devices, so the position must be checked against the latest instructions in force.
Key point
Deleting the chat afterwards does not undo the disclosure. The data has already reached an outside server the moment it is uploaded.
Core principles
- Approved tools only: use what the institution permits, and anonymise data where required.
- Human in the loop: check every rule number, figure, date and name against the source document.
- Accountability: the officer who approves or signs a note or letter remains fully responsible. AI said so is not a defence.
- Bias: recruitment, penalty and evaluation decisions must not be handed to an AI tool.
Example
To summarise a disciplinary inquiry report that names staff members, the right course is to use an institution-approved tool as per policy, with identifying details removed where required, or prepare the summary manually, and then verify it against the report. Uploading it to a free public chatbot is wrong even if the chat is deleted afterwards, and forwarding it to a friend with a paid subscription is worse, since it shares a confidential record with an unauthorised person.
Writing a good prompt
| Element | Example |
|---|---|
| Role | You are an office assistant in a central institute |
| Context | The background of the case, with no personal details |
| Task | Draft a reminder to departments on pending utilisation certificates |
| Format | Office Memorandum format, not more than 150 words |
| Constraints | Do not invent rule numbers; leave a blank where a reference is needed |
Do and do not
| Do | Do not |
|---|---|
| Use institution-approved tools | Upload personal or confidential data to public tools |
| Anonymise data before use | Copy an AI-supplied rule number without checking it |
| Verify against the original rule, OM or Act | Let AI decide recruitment or penalty matters |
| Keep your own judgement and signature | Cite an AI answer as authority in a note |
Exam tip
In scenario questions, discard the two extremes: complete reliance on AI, and a blanket ban on ever using it. The balanced, rule-compliant option is almost always correct.
Practice questions
Answer all, then check. Explanations appear after checking.
Finished this topic? Tick it off.
Saved in this browser only. See all revisions due