HOW AI TURNED ME INTO AN ACCIDENTAL MEDIEVALIST

What chasing a medieval family through a thousand years of fragmented evidence taught me about working with artificial intelligence.

In the summer of my sabbatical, I did not set out to become an amateur medievalist.

It started, as these things often do, with a trip.

Not that kind of trip.

I was planning some much-needed time in France and became curious about an old family story that traced my paternal surname from the hills of Appalachia back through Scotland and England to ancient Aquitaine—and, more specifically, Blaye, a small town on the Gironde just north of Bordeaux.

So I did what every great researcher apparently does in 2026.

I opened ChatGPT.

My first question was considerably less sophisticated than where the research eventually ended up: Find me the places, towns, castles and archives in France that might be relevant to this old Riddle family story.

I assumed this would produce a mildly interesting detour on my trip. Maybe I’d spend an afternoon poking around an archive, find a charming footnote, visit a ruined castle, take a photo, and move on with my life.

Instead, one of those searches led me inland from Blaye toward Baignes-Sainte-Radegonde, a village in the Charente, and to a château called Montausier—a place that appeared in the old family tradition I was beginning to investigate.

Then things got strange.

I searched for the château.

It was on Airbnb.

(This is not a sponsored post.)

The current property—a nineteenth-century manor built on the remains of the older Château de Montausier—had been converted by its owners into a vacation home. Its documented history stretched back through the medieval lords of the region, the Sainte-Maure family and, eventually, its reconstruction in the nineteenth century.

Naturally, I went down another rabbit hole and started looking at property in the area.

The idea of a vacation home in the French countryside sounded nice, and I had some cash I could theoretically invest.

That’s when I discovered the château wasn’t just available on Airbnb.

It was for sale.

There is a particular difference between reading the name of a place connected to your family in an old pedigree and discovering that you can sleep there on Airbnb.

There is an even greater difference between discovering that you can sleep there and realizing that, theoretically, you could buy it and make it your own.

Suddenly, my passing inquiry on ChatGPT became considerably less abstract.

A family story that had existed in books, names and half-remembered genealogies now pointed toward an actual landscape—a house, a surviving medieval tower, a village, fields, roads and archives that I could physically visit.

And a château I could theoretically own.

To be clear: finding Château de Montausier through a search on ChatGPT did not prove that I descended directly from the medieval people who lived there. Neither did discovering that it was for sale and somehow within the realm of affordability.

In fact, separating those things became one of the most important lessons of the entire project. What the discovery did do was supercharge my ability to investigate two related hypotheses:

Did my family actually descend from the medieval cast of charachters associated with Angoulême, Blaye and Montausier?

and

Could I use AI to work through the diligence required to determine whether purchasing the property and operating it as a boutique hospitality venture was actually within the realm of possiblity?

One question pointed backward almost a thousand years. The other pointed very practically toward the future.

Both began as possibilities. Neither was true simply because ChatGPT could make a convincing case for it.

A discovery can make a hypothesis more interesting without making it more true.

My trip to France had suddenly become a lot more interesting.

Within a few days, instead of casually preparing for a holiday in Europe, I found myself following genealogical rabbit trails, comparing eleventh-century cartularies, exploring papal records, discovering monastic charters, tracing witness lists and property transfers, reading nineteenth-century pedigrees and trying to make sense of obscure Latin variations of the same handful of names.

Rudel. Rudellus. Ridel. Ridellus. Riddell. Riddle.

The inquiry began moving outward from southwestern France through networks touching Capetian France, Norman southern Italy, England and Scotland, and eventually toward Ulster and the American branch of the family I actually knew.

What had started as “where should I go in France?” was becoming a much more difficult question:

What, if anything, could actually be demonstrated about how these people were connected?

At some point, I looked at the growing pile of sources, hypotheses, timelines and contradictions and realized I had accidentally built a medieval research project using methods that would have been nearly impossible for me to attempt alone even a few years earlier.

The machine was very good at getting lost

Truth be told, AI made this kind of inquiry possible for me in a way that would have been considerably more difficult even a few years ago. The medieval record is fragmented almost by definition.

A person might appear under a territorial name in one charter, a family name in another, a Latinized form somewhere else and only as a witness in a fourth document. Sources are dispersed across French regional histories, English antiquarian volumes, digitized manuscripts, archaeological surveys, scholarly databases and books published over hundreds of years.

When it comes to the work of genealogy, the challenge is not simply finding information. It is noticing relationships between fragments that were never designed to be read together.

That turns out to be something large language models are very good at. While researching, I could ask the agent I created in ChatGPT to search for variant spellings of a person’s name across different collections.

I could ask it to compare witness groups.

I could move sideways: if this person appears here, who else appears with him? Where else do those people appear? Which monasteries did they patronize? Which lord did they serve? Did the chronology even allow these two records to describe the same person?

I could ask it to produce ten possible explanations for a gap in the record and then spend an afternoon trying to destroy all ten.That last part became increasingly important. AI is very good at producing plausible connections, but it is not naturally good at being embarrassed by them when they turn out to be horribly wrong.

Give a language model two men called Geoffrey Ridel separated by a few decades and several hundred miles, and it will happily help you imagine a beautifully coherent story about how one became the other. Sometimes that story might even be right.

But coherence is not evidence. And the more powerful the research system became, the more disciplined I had to become about that critical difference.

Then I stopped asking the machine for answers

Eventually, one methodological shift changed the nature of my research project more than anything else: I stopped asking ChatGPT questions like Who was Geoffrey Ridel? and started asking, What does this document actually tell us about Geoffrey Ridel?

That sounds like a small distinction, but it changed everything.

Instead of treating AI like an oracle capable of reconstructing a thousand years of family history, I began treating it more like an extraordinarily fast research partner with an occasionally dangerous imagination.

I turned my string of chat context windows into a dedicated research agent and gave it a fairly strict job: find sources, extract names, dates, places and relationships, compare them with what we already knew, point out contradictions and suggest possible connections. Most importantly, it had to show me where the information came from and avoid quietly turning a possibility into a fact.

Over time, I found myself sorting my research into three simple questions: What do we actually know? What can we reasonably infer? And what could we reasonibly imagine to fill in the gaps?

My agent might identify a medieval charter where a man named Warin witnessed a land transfer alongside three other people. That is something we know, since it is evidenced by a stated fact.

If those same four people repeatedly appear together across several documents over twenty years, it becomes reasonable to infer that they belonged to the same social or political network. If I then propose that Warin was the father of another Ridel who appears later in the region, that might be a promising explanation, but it is still a hypothesis.

AI tends to blur those distinctions because language models are exceptionally good at producing coherent narratives, and coherence can feel remarkably similar to truth when you are reading it on a screen.

So I started deliberately adding friction to the research.

Whenever the agent found something exciting, I would ask it to attack the finding. What evidence contradicts this? Who else could this person be? Are we relying on a nineteenth-century historian who was himself repeating an older claim? Can we find the original document? What would have to be true for this theory to fall apart?

In other words, I stopped using AI only to build arguments and started using it to try to destroy them.

That turned out to be considerably more useful.

The research environment grew with the project. There was a corpus of sources, timelines, people, places, competing identity theories, unresolved questions and claims that needed to be checked against original material. As I dug into the archives, some connections that initially looked thrilling were downgraded. Others became stronger as independent pieces of evidence began to converge.

Because the machine could hold and compare far more of that landscape than I comfortably could, I could move through the material in a way that felt almost spatial. If this man appears here, who appears around him? Where else do those people appear? What happens fifty years later? What do the Italian sources say? What is the strongest reason not to believe this connection?

That was where the technology began to feel genuinely new to me.

AI had not solved medieval genealogy. It had changed the spatial dynamics and economics of inquiry.

Questions that once might have required days of preliminary searching by a highly trained specialist could be explained, explored and debunked within minutes. Weak hypotheses could be discarded faster. Promising ones could be pursued much more deeply. Sources written in languages that I do not fluently read suddenly became navigable. Connections across communities, continents and centuries became easier to notice.

The cost of asking another question had collapsed. And that meant that I could ask far more of them.

The breakthrough was iteration, not automation

This distinction began to matter to me far beyond genealogy. Most conversations about AI at work still gravitate toward automation. What tasks can the machine perform? What jobs can we eliminate? How quickly can we generate an output?

Those are real questions, but the research project that I stumbled into during my sabbatical made me more interested in something else: what happens when AI makes it dramatically cheaper to think through a problem, its potential solutions and their implications?

The value of using machine learning in a research process was never that it could write a family history while I went to lunch. The value was that I could explore an idea, interrogate it, reject it, reformulate it and try again at a speed that changed what kinds of questions were practical to pursue.

Learning real estate with an actual estate

While I was busy working with Chat to discover which relatives were connected to my family line, the same thing was happening with the château. One line of inquiry was moving backward through medieval France while the was moving very much toward the future. Could I actually buy Montausier?

That’s when I created another agent specialized in real estate and boutique hospitality to help me work through that question too. What would French property diligence require? What did the available reports suggest about the roof, windows, sanitation and energy performance of the current house? What kinds of hospitality concepts might make sense in this part of the Charente? What might the nightly rental economics look like? What legal, financial, development and operating questions would I need to answer before the idea stopped being romantic and started becoming investable?

Again, AI could not tell me whether buying a château was a good idea. God help us if it could.

What it could do was help me turn a vague fantasy into a structured set of questions that I could ask a realtor and property owner very quickly.

In both the genealogy and the property project, AI helped move me from “wouldn’t it be interesting if…” toward a much more useful question: What would we need to know to find out?

In reflection, this feels like one of the most valuable things that machine learning can do.

This started to look a lot like design

The more I found myself collaborating with AI in this way, the more familiar the process felt. Looking back at my core skills in design thinking, I was essentially engaging in rapid prototyping in highly specialized arenas.

Design thinking teaches us to notice something. Then you form a hypothesis. You make that hypothesis concrete enough to examine. You develop concepts to solve for the problems you uncover, then test those against reality. You learn something. Then you go around again.

That same pattern shows up constantly in the world of design, product and tech.

A simple prompt can now help us synthesize hundreds of customer reviews and surface patterns that a team might otherwise miss. It can generate several explanations for why a product is underperforming in a specific region. It can compare a concept against a set of highly nuanced design principles.

It can turn a rough service concept into something tangible enough for people to react to before anyone commits significant engineering resources. It can scan a marketplace or cultural landscape and surface signals that would be difficult for one person to hold in their head.

All of that creates extraordinary leverage. But the moment a possibility becomes a decision, something else enters the room.

Judgment.

There is a meaningful difference between saying the evidence shows this, the model sees this pattern, we think this might mean something and we are going to dedicate resources to build this.

AI makes it dangerously easy to glide from the first statement to the last.

I think one of the defining design challenges of the next decade will therefore be less about whether AI can perform a task and more about whether we can preserve visibility into the foggy middle ground where machine reasoning ends and human judgment begins.

I had already begun thinking about this while building AI-assisted curation tools at Airbnb. A curator might be reviewing an experience in a city they did not know intimately. AI could help surface relevant context, compare an experience against shared quality principles, identify reasons something appeared strong or weak and find patterns across far more supply than one person could ever review manually.

But the interesting question was never simply, “can the model score this?”

It was, “can the prompt help a person make a better decision?”

That requires showing reasoning. Showing relevant evidence. Making uncertainty visible. Allowing disagreement. Creating feedback loops. And being clear about who ultimately has the final authority to say yes or no.

The machine does not need to disappear for AI to be useful. Neither does the human.

The most interesting AI-powered systems sit somewhere in between. They let machines do what machines are increasingly good at—searching, comparing, generating, summarizing, detecting patterns, holding enormous amounts of information and exploring possibilities quickly—while preserving for people the things we remain uniquely capable of: context, values, taste, feeling, consequence and the final decision about what deserves to become real.

As making gets cheaper, judgment gets more valuable.

AI can already generate more copy, images, interfaces, concepts, strategies and research summaries than most teams could ever reasonably use. That changes the bottleneck.

Increasingly, the question will not be whether we can make something. It will be whether we know what is worth making.

What is signal and what is noise? What is merely polished and what is actually good? What belongs in this place, for these people, at this moment? What are we assuming? What evidence would cause us to change our minds? What should we leave alone?Oddly enough, these are almost exactly the questions I found myself asking while staring at eleventh-century charters trying to determine whether two dead Frenchmen named Geoffrey were related.

AI did not make the research process less human. It made the distinctly human parts harder to ignore: curiosity, skepticism, taste, context, the ability to tolerate uncertainty, the willingness to change your mind, the discipline to abandon a beautiful theory when reality refuses to cooperate, and the responsibility to know when you simply do not know.

I still do not know exactly where the Montausier story ends. I am going to France to visit the château for a proper round of diligence on the property. The business case is still being tested. Simultaneously, my medieval research continues. Some genealogical connections have become much stronger. Others that looked exciting at first have fallen apart completely.

Which, increasingly, feels like the point.

AI helped me travel much farther into these fantasies than I could have gone alone. It helped turn an old family story into a research project and a château listing into something concrete enough to investigate as a possible future.

The machine expanded what I could explore, but it did not tell me what to believe.

I still have to decide what survives contact with reality.