The rules: policy, confidentiality and your responsibility

Three layers of rules

When you use AI at work, three layers apply at once:

1. National policy: the National AI Policy 2025, which sets direction for the whole country.
2. Existing law and service rules: the rules on official secrets, the handling of classified documents, conduct and the protection of personal information already apply. Using a new tool does not suspend them.
3. Your department's instructions: which tools are approved, for what and by whom. These are the most specific and the most binding on you day to day.

What the national policy says, in brief

The federal cabinet approved the National AI Policy 2025 on 30 July 2025. For officers, its most relevant points are:

• a specialised curriculum for public servants on AI basics, ethical use, data governance and personal data protection, developed with the Establishment Division (section 2.2)
• disclosure of where AI is used in the public sector, including a public register of public sector AI systems (section 3.2)
• human oversight of critical AI operations, particularly in high-risk scenarios, with impact assessments and audits for high-impact systems (section 3.2)

The policy is a policy, not a law: it sets direction and targets rather than legal duties on individual officers. But it tells you clearly where government practice is heading. Working that way now is sensible.

The red lines

Never put these into a public AI tool (the free or personal versions of chatbots):

• Classified material: anything marked Secret, Confidential or Restricted, or that should be
• Citizens' personal data: names with CNIC numbers, addresses, phone numbers, health, financial or family details
• Personnel information: ACRs and PERs, disciplinary cases, medical records of staff
• Pre-decisional and sensitive material: draft policy not yet approved, procurement evaluations, legal advice, anything market-sensitive

Why? Text you type into a public tool goes to the provider's servers, usually outside Pakistan, under the provider's terms rather than the government's. See data leakage and data sovereignty.

When you are unsure

Ask yourself: would I be comfortable explaining to my secretary, an auditor or a court why this text left the government's systems? If not, do not paste it. Remove or replace the sensitive parts, use a department-approved tool or do the task without AI.

Responsibility does not move

Whatever tool you use, the responsibility for the output stays with the officer who uses it and signs it. "The AI wrote it" is not a defence for a wrong figure in a summary, a misquoted rule in a letter or a biased recommendation. This is what accountability means in practice.

Being open about AI use

Good practice, like the direction of national policy, is to be transparent. When AI helped to prepare a substantial document, consider saying so in your note to file, for example: "First draft prepared with an AI assistant; facts and references checked by the undersigned." It costs one line and protects you.

Try it

Take a recent letter or note you wrote. Mark every piece of information in it that would cross a red line above. Then rewrite the request you would give an AI assistant to help draft it, with all of those items replaced by placeholders such as [Applicant] or [Amount].

Check yourself

1. Name three types of information that must never go into a public chatbot.
2. Does the National AI Policy 2025 create legal penalties for individual officers?
3. If an AI-drafted summary contains a wrong figure and you sign it, who is responsible?

Answers: (1) Any three of: classified material, citizens' personal data, personnel information, pre-decisional or sensitive material. (2) No. It is a policy that sets direction and targets; it calls for a legal framework but is not one itself. (3) You are.