Working from sources, not toward them

The strongest position is to make the AI work from material you have already verified. You find the report, the data, the ruling. You give it to the tool. The tool summarises, extracts, compares, reformats. Everything it produces is anchored to something real that you can point to.

This reverses the usual flow, and it is slower at the start. But it cuts the fabrication problem sharply, because the model is not recalling anything, and every claim can be checked against the text in front of you. It is transforming what is in front of it, which is the thing it does most reliably.

For any high-stakes work, this is the method: sources first, AI second. The AI saves you the hours of reading and structuring. You keep the responsibility for what is true, and you can show where every claim came from.

In practice. A policy analyst stopped asking AI for facts about legislation and started pasting the legislation in. "Summarise section 12 in plain English." "What does this schedule require of employers?" Every answer could be checked against the text on screen. Her error rate on specifics, which had been the reason she was nervous about AI, went effectively to zero.

At your desk. Take a task you would normally ask AI to answer from memory. Find the primary source yourself first. Give it to the AI and ask for what you need from that document. Compare the result with what you would have got the other way.

Write it down.

• The task I reversed:
• The source I found first:
• The difference in the output: