Policy drafting with AI: what to never paste

Three ways to see it

1. 1

   Reduce a noting to its first principles and you find three layers. The first is public knowledge: laws on the gazette, judgments online, IMF country reports, World Bank statistics. The second is internal-but-non-sensitive: the ministry's organogram, the standard SOP for a summary, the format of an Executive Committee of the National Economic Council submission. The third is sensitive: cabinet positions, intelligence-fed assessments, individual taxpayer data, security force movements. The discipline of policy drafting with AI is to know which layer you are in for every paragraph you paste.

2. 2

   Layer one is fair game. Paste the gazette text. Ask the model to summarise the Trade Policy 2024 and compare it to the previous policy. Ask it to give you ten arguments for and against a tariff cut. The model is well-suited to ingest publicly available text and produce a structured comparison. Anything that is already in a Lahore High Court PDF, an SBP circular, or an IMF mission concluding statement is fine to feed in. The risk is near zero, the speed-up is real.

3. 3

   Layer two is conditional. Internal templates and SOPs may be safe to paste only if the AI provider's terms forbid training on your inputs and your ministry has a written approval for this category. Many federal ministries now have an AI-use circular. Read yours. If it is silent, treat layer two as unsafe by default. The cost of caution is one extra hour. The cost of carelessness is a notice from the Cabinet Division and a possible departmental inquiry.

The why-tree

Try this with Claude

AI-edge prompt: 'I will paste the public text of Pakistan's Finance Act 2024 sections 8 to 14. I will then paste a public stakeholder consultation summary from the FPCCI. Produce a noting for the Federal Board of Revenue chairman that compares the two, identifies three points of friction, and proposes a public-facing FAQ. Do not invent any numbers. Mark anywhere you would normally need internal data with [INTERNAL: please supply].'