A Gentle Introduction to AI for Family Historians

How modern language models actually work and what they mean for anyone researching or writing family stories

Artificial intelligence gets talked about as if it’s powerful, unpredictable storm rolling in.

But what if it’s more like a tool sitting on the desk, waiting to be understood?

For family historians, AI isn’t about replacing us as researchers or writers. It’s about working with language in a new ways, and using it to turn raw information into meaningful connections and thoughtful stories.

This is a gentle walk through what’s actually happening behind the scenes when a Large Language Model — the kind of AI that chats with us such as ChatGPT — generates words on a screen.

What “Artificial Intelligence” really means

Artificial intelligence simply refers to computer programs designed to act like humans do. It’s “artificial” because it is not human, it just responds to us in a human way.

When that intelligence in the computer program is focused on language, it’s called a Large Language Model, or LLM.

Think of an LLM as a system that has read more text than any single person ever could (books, articles, diaries, news archives, and web pages) and from that reading, it has learned patterns in how language works.

But an LLM doesn’t memorize everything it reads.

It learns relationships: which words usually follow others, how tone shifts, how sentences fit together to form meaning.

That’s how an LLM can compose an original paragraph that sounds human without having feelings or thoughts of its own.

Inside the computer program, this happens through a network of mathematical layers called a neural network, loosely inspired by the human brain.

A special design called a transformer helps the LLM focus on the most important words in a sentence — a process called attention.

And through training the LLM — reading and adjusting again and again — the model develops an intuition for language.

It’s pattern recognition, and each new LLM model gets better at it.

This is why everyone gets excited when ChatGPT or Claude or Google’s Gemini releases a new model.

Internet searching vs. AI generating

When an internet search engine, such as Google, is used, it looks for existing material in a database of webpages.

Type Montgomery County dairy 1950 into a search bar and it hunts for those exact words in the database, returning a list of pages already written.

That’s searching. Think of it as a librarian handing over books with the pages marked for you.

A language model does something entirely different.

It doesn’t look through databases. It writes.

Given a few words or a question (called a prompt), the AI predicts what piece of text should come next, one tiny fragment at a time. Those fragments are called tokens — sometimes whole words, sometimes parts of words.

If that sounds abstract, think of Mad Libs.

There’s a blank in the sentence: “My great-grandmother was born in ___.”

Dozens of possible words could fit: “Ireland,” “1882,” “Spring.”

Each choice changes the meaning, but all are plausible. (How it picks which one is correct we’ll cover in a minute.)

That’s how the model builds language. It fills in the blanks based on patterns it learned from all that reading.

Search retrieves what exists.

An LLM generates what fits.

The long training journey for AI models

Training an LLM isn’t done overnight.

It takes weeks or even months of nonstop computing on thousands of specialized processors. During that time, trillions of tokens — bits of text — move through the system.

Every moment, billions of internal “settings,” called parameters, adjust themselves to improve predictions.

It’s like teaching a person to write by showing them every book in the library and asking, “What comes next?” until the guesses are right most of the time.

Once training ends, the model is fine-tuned for tasks like summarizing or polite conversation, then released for public use. That’s when we interact with it, typing prompts and seeing replies appear instantly.

Behind that instant response sits months of learning and staggering amounts of computing power.

Each model also has a knowledge cutoff, a date when its learning stopped (For ChatGPT 5 that is September 2024, over a year ago.).

It’s like working with a well-read historian who’s brilliant on everything up to a certain year but hasn’t seen the latest newspaper.

The chat: from prompt to response

A prompt is simply text that gives the model direction. It can be a question, a phrase, a request.

When the model receives it, it begins something called inference, using what it already knows to predict the next token.

It evaluates countless possibilities, ranks them, and picks the one most likely to follow. Then it repeats the process again and again, building a sentence word by word until the thought feels complete. It does all this in fractions of a second.

It’s a bit like storytelling improv.

Someone offers the first line, and the storyteller continues naturally from there, following rhythm and tone.

That’s what the model does, not by thinking, but by calculating.

Tiny changes in a prompt, such as the punctuation, phrasing, or even mood of the user, can shift the entire answer.

That’s why two identical questions sometimes return different results.

The model isn’t fickle; it’s probabilistic. It sees many possible paths forward and takes one of them.

When AI sounds sure but isn’t right

Sometimes an LLM speaks with absolute confidence and still gets things wrong.

That’s part of how prediction works.

We often call this a hallucination (some people refer to it as lying) and it happens when it invents a detail that sounds true but isn’t.

It doesn’t lie; it fills a gap in a pattern.

Because language online is full of repetition, bias can sneak in too.

And occasionally, a model becomes a little too flattering or agreeable, and this is called sycophancy. This was an issue for ChatGPT 4.

And that’s where our good judgment as humans steps in.

Facts can be checked. Records can be verified. Editing writing brings balance back to tone and meaning.

LLM prediction can’t replace human verification or our lived experience (and we do not want it to).

AI revolutionizes genealogy

AI is easy to misunderstand because it acts human and looks like magic. But it is really a giant pattern matching machine.

For family historians, that means less time wrestling with records and more time connecting the dots that make our ancestors’ lives feel real.

AI is a fantastic tool for both researching and writing when we go beyond the chat box and use its full power. I’ll be covering more of that this fall.

— Denyse

P.S. Would it be helpful to see this turned into a free mini-course with visuals—something that walks through each idea step-by-step?

Comment and let me know.