AI in Genealogy: interview with expert Mark Thompson
Mark Thompson is an AI in Genealogy expert who founded the website Making Family History, and he co-hosts a podcast called The Family History AI Show with Steve Little. Mark sat down with us to help us understand what AI can and can't do for our genealogy research, and how to best use it to get the most helpful and accurate results.
From Collecting Family History Stories to Technology and AI
Legacy Tree:
We'd love to know more about your background and what brought you into the genealogy space. Based on your history and where you're at right now, how does AI overlay with that?
Mark:
I came to it naturally from the technological side. I've always been a bit of a geek, and spending time at the library was always a fun thing to do when I was a kid. I got into computers in my teens, so I was always looking for some way to use them. My whole career was in tech and high tech, and I've always found myself as the guy who was in charge of trying to figure out how to introduce new things, usually in a big corporate environment. I was always trying to figure out how to make things bigger, better, faster, stronger, and cheaper.
Sometimes, technology would come along that a department head would want to take advantage of, or sometimes a department was in trouble and was running out of money or running out of people or running out of skills or whatever. They needed to introduce technology to improve things. Thankfully, I found myself in that place where I was always looking at what was new. I'd evaluate how it works, what it's good for, and what it's not good for. Then we would look at how the company could improve our business with the use of those technologies.
Genealogy became important to me in my early twenties. The family genealogist before me was my great-uncle, my paternal grandfather's brother. He had been the family genealogist for almost two generations. He was a Catholic priest, and he traveled around from kitchen table to kitchen table and parish to parish, collecting stories and facts and writing them down on huge scrolls. When I was in my twenties, he visited me and started telling me stories about his genealogy work, and I was sold. That day, I wondered how I could get his work into a computer. I remember that it all clicked for me at the same time.
How AI Is Similar to the .Com Era
Legacy Tree:
How is AI perceived today?
Mark:
Genealogy is currently in the crazy position of trying to figure out the best way to apply a new technology. Every single industry on the planet is going through this right now in a way that's not at all dissimilar to how I remember the .com era. We're in the same situation as in 1991, '92, '93, and '94, the early days of .com where everybody was going, «This internet thing, it's cool. It freaks me out a little bit. I don't like all this information sharing and hackers and bad guys.»
All of the movies coming up at the time said all the worst things that could happen. But I see possibilities. That's where we are right now. It's very early days when people are still figuring out what we can do with AI. People are still trying to figure out what can be done and what's good to do in genealogy.
How to Successfully Ask LLMs to Perform a Task
Legacy Tree
Why is it important to use the right language when asking an LLM to perform a task?
Mark:
When you use certain words, you will find that large language models apparently “understand” them. If you know the right words, it will understand you better. Some people are good at choosing words that clearly describe their intent, and others are not. AI is like the revenge of the English majors. They are people who are good at putting words together and clearly describing what they mean. They will get better results from the current crop of large language models.
If you know the words to describe what you want, you'll get better results than if you can't describe it well. If you know how to ask something from a large language model, you can get it, particularly if you're good at breaking complex problems down into small problems. Large language models are essentially trained on every word that has ever been publicly stated. If you're trying to figure out how to do something that has been stated many times on the public internet, you can probably get that information out if you know how to ask for it.
If you know how to describe a genetic cluster and have the words to describe what that is, you can get a large language model to help you. If you know how to explain how you want something summarized, you can get it to summarize properly. For example, do you want it summarized by topic? Do you want it summarized by date? If you can describe what you want, you can get the LLM to do it. But if you just say summarize, you'll get something that is summarized, but not necessarily the way that you intended.
Legacy Tree:
What are ways to get the output from an LLM that you expect?
Mark:
To get the output you expect, you must clearly understand what you want in your head and then communicate that with words to the chatbot. That is the definition of prompt engineering. If you have a hard time saying what it is that you want, your hallucination rate will go up. If it isn't clear, succinct, and understandable, the LLM will take a swing at what it thinks you asked for and it will give you something. However, it might be different from what you intended, but if you give it an open question without really clear expectations, it will go and take its best swing. That is both the beauty and the bane of prompting. You'll always get something back, but it might not have anything to do with it you actually wanted.
Steer Clear of Genealogy Research Questions
Legacy Tree:
Accuracy in LLM responses can be a problem. How can we improve accuracy?
Mark:
For the people new to genealogy and artificial intelligence genealogy, my recommendation is to steer clear of genealogy research questions.
For example, where was great-grandpa Fred born? Asking the type of question you would type into Google is actually the worst-case scenario. That doesn't mean that you can't do it, but it's one of the riskiest endeavors in terms of getting a hallucination. It assumes that the fact is out there to be found, and it might only be if great-grandpa Fred is famous and written about more frequently than other people with the name great-grandpa Fred. You might get a great-grandpa, Fred, but you might not get yours. That's a common type of hallucination.
There are many examples of chatbots giving back information that makes sense. It's just more frequently found information that sounds like yours. A chatbot shows you the most frequently found words in that part of the internet that match the prompt. It's just a statistical engine that comes back and says, here are the words most likely to relate to what you asked for, even if they're not the words you thought you were asking for.
ASK LLMs What You Already Know
Legacy Tree:
Why is it important to ask an LLM something you already know?
Mark:
LLMs are best at helping you do things you already know how to do faster and more efficiently. If you know how to summarize a document, you can use it to summarize the document quickly and accurately. One of my favorite tasks is obituary analysis. There are lots of loops and turns through an obituary. There can be dozens of people referenced. Anybody who's ever tried to analyze an obituary and turn it into a family tree learns that many are poorly written. I would print out the obituary, go through each name, confirm it, and then add it to the tree. I know how to do it by hand. So I created a prompt that helped me do all those steps and turned a 15-minute task into a two- or three-minute task instead.
Legacy Tree:
What prompts are available to assist genealogists and how to create a prompt?
Mark:
I've created and made available my custom GPT for obituary analysis to anybody who wants it. And it gets used all the time. People look it up and just run it themselves, copy-paste the obituary into it, and press the enter key.
Most people don't think about this, but you can get a chatbot to help you create a prompt for your chatbot as well. You can say, «I'm having a tough time coming up with a prompt. Act in the role of a prompt engineer. Come up with a prompt for reliably analyzing an obituary so that it's easier for me to get it into my family tree. Can you help me with that?» Then you can review the response and then see where you went wrong or what you forgot to include. Essentially, it's teaching you how to be a prompt engineer through a real-world genealogy example.
Assigning LLMs a Role
Legacy Tree:
What does it mean to assign the LLM a role?
Mark:
The simplest prompt style is to give it a role and a task. For example, «Act in the role of a marketing person. Please review my copy for this ad and provide me feedback.»
That's the simplest prompt style; some call it a role and a goal or a role and a task. In genealogy, I recommend three steps as the most basic prompt. A role, a task, and a format that helps analyze or create a piece of genealogy information. In my obituary example,
- ROLE: Act in the role of an obituary analyst.
- TASK: Please review the following obituary and tell me how everybody mentioned in the article relates to the decedent.
If I stop there, I still have a narrative to go through and double-check. All I've done is reformat the information, but I haven’t necessarily made my task any faster.
- FORMAT: Define the format of the information.
In the obituary example, I always get it formatted as a table.
Formatting LLM Answers as a Table
Continuing the obituary example, I tell the chatbot to give me the information in a table. Then, it creates one line per person: Mary is Joe's daughter, Barney is Joe's brother, and Gertrude is Joe's mom. Now I've got a list that's ordered that I can go through, and now I don't have to print it out anymore because it's all in one row. It's easier for me to eyeball it on the screen. Picking a format that helps you with your next steps with the information is a great way to improve both the speed as well as the accuracy of whatever it is you’re working on.
Taking the formatting to the next level, you could say, «Column one is the person's name. Column two is the person's relationship to the decedent. Column three is the rationale that you used to determine the relationship. Show me the text in column three that you used from the obituary that helped you come to that conclusion.»
Now I've not only gotten the information, but I've also got a hallucination check. I can read the text that the chatbot used to conclude that Mary was Joe's daughter and confirm it without having to search through the obituary.
I can look at the text in the obituary to double-check. I'm not relying on the chatbot to make my work product. I'm relying on a chatbot to help me make my work product efficient and reliable to use. It will help me get a little faster or help me get a little bit more reliable, reduce my human error rate, help me be more efficient, and help me do things consistently. Some of our family trees have thousands of people. So we do these processes, like obituary reviews or looking at a census record, hundreds and thousands of times. If we can find one step in that process to simplify, automate, make more efficient, make more reliable, that's the kind of gift that keeps on giving.
Analyze Your Manual Processes and Create Automation with AI
Mark:
Before jumping into AI automation, ask yourself a few questions. How do I do that task manually, or, what are the steps I follow to do that task myself? I encourage you to ask yourself how you do the task today. Then, we see if the process can be done faster with AI. Or we focus on the benefit that you get from using the tool. Ask yourself, where does bringing a tool in better improve things? It all starts with the analysis of the manual way. If the old way worked fine, you wouldn't bother bringing a webtool into the mix.
Legacy Tree:
Tell us a bit about Clustering and AI.
Mark:
As genealogists, we cluster all day, every day, and we don't even think about it. I'm doing surname clustering. I'm clustering first names. I'm finding a pattern. Clustering is just pattern recognition. And we see those names that flow through our families. That's a form of clustering. I wish I could go back in time and tell my Thompson ancestors that there were more names than John, Charles, James, and George. Family names are a form of clustering. I feel better if I see one of those names because Thompson is a very common last name. But if I see a John Thompson or a James Thompson or a George Thompson, I create this cluster in my head that suggests these people might be one of my people.
Clustering shows up in so much of genealogy. We look at migration patterns. We look at immigration patterns. We look at naming patterns. We look at clothing styles and haircuts in old photos.
AI is amazing at seeing patterns. As long as you can describe the pattern you want it to see, you can back off and ask a very, very open-ended question and say, do you see any patterns here? One caution with this type of question is that it can be prone to hallucinations, so you need to check the pattern the LLM returns to see if it is legitimate.
Large language models catalog, group, and organize patterns and based on the prompt you give them. That's their secret sauce. So if you can figure out how to describe the pattern you want the LLM can help you look for it.
Legacy Tree:
If you know how to do a thing, you can get AI to help you do the thing. What does this mean?
Mark:
If you can't verbalize what you want, your hallucination rate will be higher. In the worst case, you don’t have the words to ask the chatbot how to do the task. The best thing about large language models is the actual tool itself. The technical hurdle with using a chatbot is quite low. There has never been a technology tool with this much power where the effort to learn how to use it is this low. That is an amazing superpower. We just have to use words, but we have to use the right words. The more accurate the words, the less hallucination you get and the more that they will support the work that you're trying to do.
You no longer need to wonder how you're going to get these ideas out of your head and into this crazy user interface with all the clicks on the menus and stuff. That hurdle's gone now.
Legacy Tree:
How do we get the data into an LLM for analysis?
Mark:
Part of the challenge of getting chat GPT to help you with a problem is getting the information into the prompt. For example, it could be a simple prompt to summarize a document. Copy and paste the 500 words into the prompt, and then give it your prompt for how you want it to be summarized. Then, click enter.
A year ago, even that was hard, and you couldn't get a spreadsheet in. It was groundbreaking when you could get a CSV file in. Now, you can actually upload an Excel-formatted XLSX file straight off your desktop and put it into most of the chatbots. So the barriers to entry regarding how I get my information is changing rapidly.
Legacy Tree:
How do AI companies help get data into an LLM?
Mark:
Currently, the three big chatbots are Claude, Chat GPT, and Gemini. They're all built into Amazon, Microsoft, and Google. Because these companies store your data, they are working very hard to solve the problem of getting your data into their chatbot. The chatbot environment itself can access the places where your data is stored. For example, when you get a Google Gemini account, it's part of your Google workspace. Your Google Workspace has access to all your files. So you can say, “look through all of the reports that I wrote last year, and please find me the paragraph that talks about great-grandpa Fred.”
It finds the paragraph and pulls it out. You can direct it to find all the paragraphs I wrote in my 49 documents about great-grandpa Fred in the last ten years, summarize them into a timeline for me, and Gemini pulls it all up. Google and Microsoft both announced recently how they will be able to do this.
Microsoft Copilot for the enterprise does this today. I can search through my Microsoft OneDrive today and extract data from files. But in five years, my Ancestry account, MyHeritage account, and DNA match list will be securely accessible from the chatbot because I authorize it just like you do when you sit down at Zoom and give Zoom authority to access your Google Calendar so that it can update. So I believe we will be able to authorize ChatGPT to access my Ancestry account and so it can analyze my DNA match list.
With the new technology we just discussed, there'll be many privacy concerns. There are societal and cultural issues to overcome before this will be widely available. I would not be surprised if all that stuff is taken care of in the next five years. Some people will gladly make their information available by choice by opting in to consolidate information from different sources. Then they could analyze, massage, summarize, consolidate, assimilate, and output it from their chatbot. Then rather than going through the hassle of transferring information between systems for analysis, we could just connect to the data directly.
This takes away the need for so many of the tools, not to mention the effort of learning how to use them. That is what the future looks like for AI. It's not one big system to rule them all. It will connect to all those systems and be able to read from them, understand the nuances between them, and reason through what it finds there. Then give it back in a prompt organized and formatted as you instructed. That's what the future of information management looks like. And it's an incredibly exciting place.
Getting Started with AI for Beginners: Chatbots are Not Like Google
Legacy Tree:
For people who are new to AI, there are a few conceptual hurdles that everybody goes through.
Mark:
- I ask a simple question. I want a simple answer. That's different from what an LLM is designed to do. «Tell me what the temperature is today in Victoria.» That's not what they're designed to do. That's what Google's designed to do. But LLMs behave differently than we have learned to interact with Google. Google has trained us to think in a specific way. To put information into that little white box, ask an incredibly terse, keyword-rich, verb-light question.
We've been trained to do that for the last 25 years of Googling, but that's not how a chatbot interacts. With a chatbot, it’s better to give a clear, precise, nuanced description of what you want. When you do this, you tend to get back what you intended, which is awesome. And that's a very different way of thinking than Googling.
- The other genuine big conceptual hurdle for people new to AI is that if I use my software this way today and then I do precisely the same thing tomorrow, I'll get the same result. That's not the way AI works. It's specifically designed not to do that. If I give the same prompt two days in a row – how do I cluster my DNA – I will get a different answer. And that is very confusing because we have learned that if I click on the file menu and then I click the print button, the printer will work the same way today as it did yesterday. That's not the way AI works.
Chatbots pull things together slightly differently every time, with a certain amount of randomness built in. If you get a result you really like, you'll want to copy it and use it, because you will get a different answer tomorrow.
Because of these, and many other differences, it’s important to experiment with AI. You just have to try stuff because Chat bots are so fundamentally different from other technology that you’ve learned. Take time to learn how to use the tools. It's fun. Get comfortable with the fact that you'll be uncomfortable for the first 20 or 30 hours until it starts to click. By then, you'll be able to write prompts that deliver the results you’re looking for.
Legacy Tree:
What are the next steps for more experienced AI users?
Mark:
For people who have spent 40 or 50 hours practicing and testing with a chatbot and have found some things that work for them, the next step is writing a reusable prompt inside Chat GPT, what they call a CustomGPT. Creating custom GPTs is probably the most important reason to use Chat GPT for people who have spent 40 or 50 hours and have found the things that they do all the time.
Legacy Tree:
Why are Custom GPTs important?
Mark:
Custom GPTs are so important because they don't just allow you to reuse a prompt. They give you a place to go back and improve your prompt based on how you use it. Custom GPTs make your tools better and allow you to refine them constantly. Until recently, ChatGPT was the only game in town that allowed you to create custom prompts to save and reuse. Now, the other major AI players have entered the scene. Gemini is calling their reusable prompts Gemini Gems. Claude is calling their reusable prompts Projects.
The ability to create a prompt that you can reuse again is invaluable for the professional genealogist. It can create massive time savings—it's like having a template for your genealogy report.
Legacy Tree:
How are Custom GPTs shared?
Mark:
When you save a custom GPT in OpenAI, you can save it for yourself or create a shareable link to send to others. You can also make it publicly available. When you share the prompt, it describes it, but you don't see the actual prompt steps. In this way, a person's work on the prompt is somewhat protected.
If you type genealogy into the chat GPT library, you can see all the names and descriptions of custom GPTs made using the word genealogy. The system will also show you how often each GPT has been used.
It is like downloading an app. I don't know how the app works, and I don't need to know how to code it. I just want to use it. You hunt around and find one that you like. You make the call about what you like. But knowing how to make a GPT is not necessarily what everybody wants. Some people just want to use ones created by someone else.
AI and Privacy
Legacy Tree:
What about privacy?
Mark:
The good rule of thumb for any genealogist is what I call the water cooler test. If you wouldn't be comfortable taking that piece of information that you are about to upload to your chatbot and pinning it to the wall at the office above the water cooler, don't put it into your chatbot.
Whether that's for privacy reasons, legal reasons, or because your boss thinks that information is competitive intelligence, always use the water cooler test. If you are ever in doubt, don't put it in your chatbot.
Legacy Tree:
How can we make the most of AI today?
Mark:
If you love AI and using these tools, keep learning and trying new things. The AI you use today is the worst AI you will ever use. It's going to get better every day. Use it for what it can help you do today. Learn how to do with AI what you already do with your genealogy, whether making family trees, contacting your matches, sending messages to family, or organizing the family reunion. Whatever you already do, those are the best things to do with AI.
Learn More about AI and Mark's Resources
If you want to learn more about Mark Thompson and AI in genealogy, visit https://makingfamilyhistory.com/.
Or, listen to The Family History AI Show podcast that he cohosts with Steve Little.
The Family History AI Show Podcast at Apple Podcast
https://podcasts.apple.com/ca/podcast/the-family-history-ai-show/id1749873836
Mark’s personal website
https://makingfamilyhistory.com/
Mark’s Obituary Analysyt CustomGPT
https://chatgpt.com/g/g-XVcnp2TJ4-mark-s-obituary-analyst
Mark’s Locality Guide for Genealogical Research
https://chatgpt.com/g/g-TpLAIvCzD-locality-guide-for-genealogical-research
If you'd like help with any of your genealogy research projects, please reach out to us! We'd love to work with you:
www.LegacyTree.com/contact-us
Legg igjen en kommentar