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

AI disclosure: This piece predates MarketImpact's per-article disclosure standard, introduced June 2026. During this period AI tools assisted with research synthesis, multilingual data collection, and supported drafting under our verification discipline; all analysis, conclusions, and editorial judgement are the author's own. How we use and verify AI:marketimpact.org/how-we-use-ai.

Last week I knew almost nothing about my great-grandfather. I knew his name was Michael Burns. I knew he was from Ballyvaughan, a small village on the south shore of Galway Bay in County Clare. I knew the family said he had been "involved" in the Irish War of Independence.

That was about it.

So that is the question I asked my AI .

Help me find out about Michael Burns of Ballyvaughan, who was involved in the War of Independence.

The question was natural. It was also incomplete in a way I did not recognise at the time.

For readers outside Ireland, a short note is needed here. The conflict I am writing about is the Irish War of Independence, fought between 1919 and 1921, and the local republican organisations of that period. When historical records use "IRA" in this context, they mean the Irish Republican Army of the War of Independence era, not the later organisation most international readers associate with the Northern Ireland Troubles. This was my great-grandfather's generation: rural west of Ireland, British rule, local ambushes, reprisals, family memory and state pension files decades later.

I began with the most obvious question: what did Michael do?

That question reflected my own assumptions. I was looking for the man with the rifle, the ambush, the military unit, the named engagement, the service record. I was not looking for his wife. I was not looking for Cumann na mBan, the women's republican organisation. I was not looking for the women who carried messages, fed men on the run, stored documents, watched the movements of British Marines, used schools and post offices as intelligence nodes, nursed wounded men, handled money, and kept the local republican infrastructure working.

I did not ask about them because I did not know enough to ask.

The AI did not "know" either, in any human sense. But it followed the evidence into a part of the story my original question had not made visible.

That is why this is not just a family-history story. It is a desk-review story.

In humanitarian work, we make the same mistake all the time. We start with the formal actor: the agency, the programme, the donor, the sector, the coordination mechanism. But the work often depends on people and systems that sit just outside the original frame: local staff, women's groups, volunteers, community health workers, teachers, faith networks, informal referral pathways, translators, drivers, traders and local authorities whose names rarely appear in the executive summary.

A good AI-assisted research process does not just search faster. It helps you notice when the question is too narrow.

That is what happened here.

Twenty-four hours later I had a sourced narrative of Michael's wartime activity. But I also had something I had not expected: the outline of my great-grandmother Margaret Burns née Keane's republican service, the Cumann na mBan branch she belonged to, the women around her, and the support network that made local operations possible.

I had Michael named in a witness statement by his battalion commander. I had his home confirmed as burned in reprisal by Crown forces. I had his comrades identified. But I also had Margaret's pension files downloaded and analysed. I had her own words, written decades later, saying that her three brothers and her husband had been active against the enemy. I had those three brothers identified as Volunteers in the same local battalion. I had her commanding officer traced. I had the Ballyvaughan Cumann na mBan branch mapped. I had a network of women whose work was much more than "catering and dispatches".

I was working from a kitchen table in Cyprus. I did not visit an archive. I did not hire a genealogist. I used public databases, an AI assistant, and a structured research method.

The question I thought I was asking

My starting assumption was a common one. Family history about conflicts often begins with the men: military service, ambushes, arrests, medals, pensions, armed units, barracks attacks.

That was the frame I brought to the search. Michael Burns was the known name. Michael was the family story. Michael was the person I asked about.

The AI began there. It searched the Military Archives sitewide database, the Bureau of Military History witness statements, and the Military Service Pensions Collection. It found him named in BMH WS1072, a 1955 witness statement by Seán McNamara, his local battalion commander. It cross-referenced that against Mid-Clare Brigade nominal rolls, Brigade Activity Reports, and other Military Archives material. It suggested searching under Burns, Byrne, and Byrnes because Irish records of this period are inconsistent.

That was all useful. It was exactly what I had asked for.

Then the research widened.

The witness statement placed Michael in the Ballyvaughan Ambush of 21 May 1921. It also named the intelligence chain around the operation. Kathleen McNamara heard information about the movements of British Marines. Miss Lyne at Corkscrew Hill was part of the communication route. Miss Grant at Ballyvaughan Post Office was involved in the warning system. A bicycle scout carried the signal.

The Post Office was not just a building. The school was not just a school. These were nodes in a local intelligence and communications network.

The AI did not stop at "Michael was there". It asked: who were the women in the chain?

That was the turn I would probably not have made on my own, at least not so quickly.

The story I did not know to ask for

Once the research followed the women, the picture changed.

Michael's wife, Margaret Mary Burns née Keane, turned out to have been a member of Cumann na mBan, the women's republican organisation founded in 1914. She belonged to A Company, Ballyvaughan Branch, 6th Battalion, Mid-Clare Brigade. Her pension application was refused, but her later Service Medal and Special Allowance confirmed that her service was recognised for medal and allowance purposes. In her own application she described training, catering, assisting the local republican forces, carrying messages, supporting men on the run, and doing whatever work was ordered locally.

Those phrases can sound small if you read them lazily. "Catering." "Dispatches." "Assisting."

The AI's value was that it did not leave them as generic labels. It cross-referenced them against other women's files and the local operational record.

That led to Eibhlis Lyne, the schoolteacher at Corkscrew Hill. Her teacher's residence was used as a local headquarters. Her pupils were used for sentry duty and to gather information on Marine movements. Children helped with roadblocks. Arms, ammunition, documents and money moved through her orbit. She carried messages, handled clerical work, helped with rates and money transfers, and continued into Civil War-era activity.

It led to Kathleen O'Connor née McNamara, Margaret's commanding officer. Her file confirmed that she was Captain of the Ballyvaughan Branch. It also placed her directly in the Ballyvaughan Ambush intelligence chain: she overheard that Marines were en route to the village, reported back to local Volunteers, and the Marines were subsequently ambushed and rifles captured.

It led to Miss Grant at Ballyvaughan Post Office, a civilian intelligence contact whose full identity remains unresolved. She was not proved to be a member of Cumann na mBan, and I have been careful not to claim that. But the Post Office warning chain shows how a woman in a communications role could become operationally central without appearing in a formal military roll.

It led to Jane Clancy and Nora McCormack in the wider 5th Battalion/Kilfenora support circuit. Their homes were not passive domestic spaces. They were safe houses, meeting places, billeting points, arms-related sites, nursing spaces and logistics hubs. Nora McCormack's record specifically says she catered for local Volunteers following the Ballyvaughan Ambush. Jane Clancy's house was used before and after major operations, including Ballyvaughan.

This was no longer just a story about my great-grandfather's wartime activity.

It was a story about a local republican system in a rural district during the Irish War of Independence. Men with rifles were only one part of it. Around them were women moving information, food, clothing, documents, money, medical care, shelter and warnings. Some were formal Cumann na mBan members. Some were civilian contacts. Some were teachers. Some were post office workers. Some were sisters, wives, neighbours and daughters whose work only becomes visible if you follow the network rather than the surname.

My original question would not have found all of that. Or at least, it would not have found it quickly.

What the AI actually did

The AI did not "discover" this in a magical sense. It did something more useful and more mundane.

It searched systematically. It logged what had been searched. It suggested spelling variants. It cross-referenced names across databases. It noticed when a male-centred source pointed toward women's files. It treated "wife", "sister", "commanding officer", "reference" and "witness" as research leads rather than background detail.

When a witness statement named Michael's comrades, the AI proposed checking their pension files for references to him. When Margaret's pension file named her brothers, it proposed searching for Keane in Clare. When her file named Mrs Kathleen O'Connor as commanding officer, it looked for O'Connor, McNamara, and the Cumann na mBan branch structure. When a Military Archives operation page linked four women's pension files to the Ballyvaughan Ambush, it treated those as core sources, not side notes.

That is the key point. The AI did not simply answer the question I asked. It helped identify the question I should also have asked.

Who made Michael's war possible?

What this means for desk reviews

The obvious use of AI is to search faster.

The more important use is to ask better questions.

In a humanitarian desk review, the equivalent might be a programme evaluation that starts with the implementing agency but ends up showing that community volunteers carried the real referral system. It might be a protection review that starts with formal case management but discovers that teachers and women's groups were the practical early-warning network. It might be a food security review that starts with market prices but ends up depending on traders, transporters and informal credit relationships.

A conventional desk review can miss those systems because the original question does not point there.

An AI-assisted workflow can help surface them if it is designed to follow relationships, not just documents. When a report names a local partner, search the partner. When a footnote names a women's association, trace it. When an evaluation mentions informal referral pathways, map them. When one source calls something "community mobilisation" and another calls it "women volunteers", treat that as a signal, not a synonym to flatten.

That is what happened in the family history project. I asked about one man's wartime activity. The research found the women who made the local system work.

The same discipline applies to humanitarian analysis.

What the AI could not do

The AI could not do everything.

It could not reliably handle every archive interface. It could not access Facebook posts that required a logged-in session. It could not OCR every image-only pension file. It could not read every piece of handwriting. It could not visit the National Archives in Kew to inspect British administrative records. It could not talk to living relatives or interpret family memory without consent and care.

Every one of those steps still requires a human being in a specific place.

The AI got me to the point where I knew exactly which PDF to download, which Facebook page to search, which census form to inspect, which file to request at Kew, and which family questions still need to be asked privately. That is the value. Not replacing the researcher. Replacing the weeks of preliminary searching that would otherwise be needed before you know where to look.

It also made mistakes. In the week of work that followed the initial twenty-four hours, a possible 1922 incident at Ballyvaughan Post Office was first withdrawn as a misread, then reinstated when another secondary source pointed to what was likely the same incident under a different date. The withdrawal and reinstatement were both recorded in the source log. The mistake was not hidden. It was recoverable.

That is why the structure matters.

The method in brief

I will write more about the method in a follow-up, applying it to a humanitarian use case. The short version is this.

Start with what you know, but name your uncertainty. Search systematically across every relevant source and every plausible spelling variant. Follow relationships, not just names. Cross-reference every claim. Ask adversarial questions: what are we missing, what would contradict this, whose work is invisible because of how the question was framed? Build a repository as you go, with source IDs, search logs, profiles, open questions and an audit trail.

The repository is not bureaucracy. It is what makes the work recoverable.

A pension claim is not the same as pension-certified service. A medal file is not the same as a witness statement. A family story is not the same as a civil record. The AI is useful because it can keep these distinctions visible across a large body of material, if you force it to.

This is the method we teach in detail in the AidGPT cohort. The next one begins on 2 June.

The democratisation question

Most of the sources I used were already public. The Military Service Pensions Collection has been online since 2014. The Bureau of Military History witness statements have been public since 2003. The 1901 and 1911 census records have been online since 2009. Margaret's pension files were sitting on the Military Archives server, downloadable, for years.

As far as I know, nobody in my family had followed the trail. Not because they were not interested, but because they did not know where to start.

The archives were open. The barrier was expertise. AI lowers that barrier dramatically, not by replacing expertise but by making the first pass accessible to anyone who can describe what they are looking for in plain language. The gap between "I know nothing" and "I know enough to ask the right questions" has collapsed from months to hours.

There is a risk here. AI can reproduce the bias in the question. If I had asked only for Michael Burns and accepted only records naming him, I would have ended with a narrower, more masculine, less accurate account of the local war. The method matters because it forces the research outward: from person to network, from formal role to practical function, from the obvious actor to the hidden infrastructure.

For humanitarian professionals, the same shift is happening with programme data, evaluation archives and policy databases. The organisations that recognise this early will move faster. The ones that do not will keep paying consultants to do manual searches that an AI-assisted workflow could complete in an afternoon.

The first, middle and last mile steps are still human

The AI found the sources. It cross-referenced the databases. It helped draft the narrative.

But the moment that mattered most was not a database search.

It was reading my great-grandmother's handwriting in an early-1950s pension file. Margaret had written, decades after the war, that her three brothers and her husband had also been members and active against the enemy.

Her own words. In her own hand. Decades after the men in her family had been the only ones the family remembered.

No AI could experience that sentence for me.

The AI told me the file existed and suggested I download it. A human reader opened the PDF and read the handwriting. The AI then helped me understand what it meant, trace the brothers, and place the statement in context. The moment of recognition, when a woman's handwritten words connected to a family story that had never been told, was mine.

The hardest research steps still require a person in a specific place. Reading a faded census form. Sitting in a reading room at Kew. Asking living relatives careful questions in private. Standing in a graveyard at Gleninagh and reading a headstone that says nothing about the war.

The AI gets you to the point where you know which graveyard, which headstone, and which silence matters.

That is more than enough.

If you want to learn this method

Later this week I will publish a follow-on piece applying the same method to a live humanitarian use case, not a historical one. Same repository discipline. Same source log. Same audit trail. Different databases.

On 2 June the next AidGPT cohort begins. Six 90-minute live sessions over three weeks. EUR 350, or EUR 280 for staff from national NGOs, local organisations and aid workers between roles. By the end you have a working repository structure for your own desk reviews, a set of verification prompts, and a method you can apply to the next country review or evaluation that lands on your desk. Delivered to the Norwegian Refugee Council, Caritas Switzerland, the International Rescue Committee, Action Against Hunger and the Estonian Refugee Council.

Apply for the June cohort at aidgpt.org.

Tom Byrnes is CEO of MarketImpact Digital Solutions and publishes Aid & Dev Dispatches. He runs the AidGPT training programme. He is currently writing a local history of Ballyvaughan during the War of Independence, based on the research described above.