I asked AI about my great-grandfather's war. It found my great-grandmother's.

I started with a narrow question.

I asked AI to help me understand one part of my great-grandfather's wartime record. The expected path was obvious. Find the formal records. Follow the male line. Build the chronology. Summarise the evidence.

That would have been a normal desk review.

It was also too narrow.

The more interesting result came when the workflow started to follow the context around the record. It surfaced my great-grandmother's war as well. Not as a footnote, but as part of the same system of displacement, labour, family survival, bureaucracy, and memory.

That is the methodological point.

Good AI-assisted research should not only answer the question you asked. It should help you see when the question is too small.

The pattern matters beyond family history

Humanitarian desk reviews have the same failure mode.

We ask for the obvious actor, the obvious dataset, the obvious intervention, or the obvious institutional history. Then we mistake the answer for the whole picture because the answer is coherent.

AI can make that problem worse. It can produce a clean synthesis that hides the frame it inherited from the prompt.

It can also help solve the problem if the workflow is structured properly.

The technique is not magic. It is discipline.

Ask the first question. Extract the claims. Identify the named sources. Ask what is missing. Search for adjacent actors. Test the timeline. Check whether the records privilege formal institutions over household experience, local networks, women, informal systems, or people whose names changed across documents.

Then separate what is known from what is inferred.

What the workflow did

The useful move was not a clever prompt. It was sequencing.

The workflow treated the first answer as a map of possible leads, not as a finished output. It asked which entities appeared repeatedly. It looked for gaps in the record. It followed contextual clues. It compared claims across sources. It kept uncertainty visible.

That is also how crisis analysis should work.

A model can review documents quickly. It can suggest source trails. It can compare timelines. It can spot tensions between claims. It can draft a first synthesis. But the analyst has to decide what counts as evidence, where the frame is too narrow, and which omissions matter.

In my family example, the omission was personal.

In humanitarian work, the omission can change a programme decision.

The lesson

AI-assisted desk review is strongest when it does three things.

First, it speeds up source discovery without pretending discovery is verification.

Second, it widens the frame by asking who and what the original question leaves out.

Third, it preserves the audit trail so the reader can see the difference between evidence, interpretation, and speculation.

That is why I keep coming back to verification discipline in AidGPT.

The aim is not to make people faster at producing plausible text. The aim is to make them faster at finding the right question, testing the evidence, and knowing when the answer is not enough.