EP41: MyHeritage Announces Scribe AI, Dave Vance Talks AI at FamilyTreeDNA, Full Text Search at Fold3, Wikipedia's AI Ban

[MUSIC] Welcome to episode 41 of the Family History AI Show. My name is Mark Thompson. I'm here with my co-host, Steve Little. We're excited to have you join us for a weekly discussion of the most interesting topics in artificial intelligence and in particular, how they're affecting the family history community. Good morning, Steve. How are things going with you today? >> Awesome. It is Monday, March 30th, and March can be quite variable here in the Mid-Atlantic. It can be 30 degrees one day, and it's expected to be near 90 degrees later this week. Today's a good day. Today's a comfortable place in the middle. Very good day. >> That's wonderful. Although I don't feel like I'm actually going to see much of the outside today. On top of podcasting today, I've got a whole bunch of syllabus to create. Grip is coming up and around the corner, and as usual, I'm behind in my deadlines. How are you doing on ears? >> Yes. Spring is the season when I think about flying kites and riding syllabi. As you mentioned, we are both coordinating grip courses. Grip Genealogy Institute is a part of the National Genealogical Society. It was formerly known as the Genealogical Research Institute of Pittsburgh. It's been Grip Genealogy Institute for some time, and Mark and I are very glad to be coordinating two courses this year. I will be for the third year coordinating the virtual beginners course in June. That's going to be from June 22 to June 26th. It's called Practical AI for Genealogists, foundations and first steps. This is going to be for experienced genealogists who may not be experienced with artificial intelligence. Mark and I will be teaching the bulk of the courses, but we'll also be joined by our friends, Blaine Betinger, Nicole Dyer, and Angela Packer. Mark, you're coordinating the in-person course in Pittsburgh and July. You want to tell us a little about that? Yeah, sure. That one starts on July 17th and runs for that week. It's in-person in Pittsburgh, as you said. This one is called Advanced AI Techniques for Genealogists. This is for people who probably have six months to a year of experience using AI, maybe a hundred or so hours of hands-on keyboard practice. I really try and improve the research and writing skills in using AI to actually help them with their hands-on genealogy work. If you've not taken an institute course before, it's a special way to pack a lot of information into four and a half days into a week. We would love to have you join us, and there's lots of other courses available, not just AI. You can learn more about this at grip.ngsgenialogy.org. Wonderful. I'm really looking forward to it. This will be my first time that I've actually been to Pittsburgh period, and let alone grip in Pittsburgh. It's going to be a lot of fun. Need to. Let's stop into the show, Steve. We've got a few really interesting things today, and we're also carrying on with our interviews that we recorded in-person at RootsTuck this year. We're going to kick off today's episode with a discussion of how my heritage is doubling down with its UCA AI, with a new tool that they introduced while we were at RootsTuck, called Scribe AI. Then in our interview this week, which we recorded in Salt Lake City during RootsTuck with David Vance. He is the general manager of Family Treaty and A. Stephen, I had a wonderful discussion with Dave, which we'll go into at length in a few moments. We had a great wide-ranging conversation about both artificial intelligence, as well as Family Treaty and A's new genetic genealogy tests that are based upon next-generation sequencing. Great conversation. Then in RapidFire, we have three interesting topics to jump into. First is a huge upgrade with at Fold 3, where they bring in full-text search capabilities for revolutionary war pension files. Then we talk about Cloud for Excel, something that I've actually been looking forward to for months. We finally got AI with an Excel that is usable. Then finally, we'll wrap up today's episode with a very interesting announcement from Wikipedia about how they want to approach the use of AI or not in their articles in the future. Well, that's a full plate. I'm glad to listen again to the interview with David Vance. But before we get there, I know you were very excited to take a deeper dive into what my heritage has done with script AI. Really some groundbreaking stuff. We're seeing some of the major vendors starting to boldly go into new places and implement some new features. So, that's what you discovered, Mark. I was really excited to hear when my heritage announced the release of Scribe AI. It actually sounded a lot like a lot of the things that you and I actually do, Steve, where we, you know, analyze records and analyze photographs. I was really interested to see what their implementation looked like. I've done a fairly deep dive since we got back from its tech because I didn't have much time while we were there. We barely had time to breathe this year. We had so much going on. But I'm very impressed. What they've essentially done is they've put a button adjacent to records and photographs and images that are already available or that you even upload yourself. And then when you click the Scribe AI button, it gives you both a description of what it is that it's seeing as well as a whole list of great research and analysis that it performs on whatever record that is that you feed to it and really do a great job of highlighting genealogically relevant information that might be available and either that record or that photograph that you show to it. So, where they've applied this is every time that you look up a record, it's already available on the site. So, like a census record or a birth marriage or a death record of any form, when you open up the record and you take a look at it, there's a big button that shows up there called Scribe AI and you click the button. You can almost take a look at the information they provide and reverse engineer the prompt that they used to create it. It's clear that what's going on is there's a prompt that's being created and sent off to a large language model somewhere. It's not clear whether or not it's actually one that they've like an open source one that they've got installed internally or they're actually sending it off to one of the public models. I can't actually find anywhere that actually describes how they're doing this. I really makes me want to dig into it and figure out how they're doing it. But what's clear what's going on is they send off whatever picture or record that is that you're looking at to some large language model and then runs for maybe 20 seconds, 30 seconds. I saw some run as long as a minute and then it comes back and it gives you a nicely laid out report that describes what is it you're seeing, conclusions that it can draw from whatever it is that you showed to it. Most importantly, and I think this is the thing that a lot of genealogists are going to love, particularly those beginner to intermediate level genealogists. It gives you a list of potential research efforts that you could undertake based upon what it is that it's found. So for example, one of the great ones I saw was I fed it a death certificate. And on the death certificate, it mentioned the informant and the spouse of the person who had died. And so the research follow-up section showed, well, follow-up on the informant because informants are usually family members, even if they have a last name that's different. So these types of things that you learn when you study every one of those different records that's deeply, that information is just presented front and center either to help teach the beginner or the intermediate level genealogist or to remind everybody no matter how advanced you are of the basic elements that are in that record set that are worthy of follow-up. So I just think they did a really nice job of this. One of the things that you demonstrated for me that really impressed me demonstrating that my heritage is paying attention to how genealogists actually work and use these things is the copy button. Exactly. I thought it was great. They realized they understand and accept that this valuable information they're getting that genealogist are actually going to want to use this information. And just the polish of providing a copy button so that it saves you the control lay control C just instantly puts the information into your clipboard. That demonstrates to me that they are genealogist focused that they want this to be useful for working genealogists. Yeah, I thought that was such a nice little quality of life issue that they put into the user interface. because any time that anybody sees them like that, they want to copy it and get it into either their research log or into their research report or descended an email to a cousin about a thing that they found or to ask a question about something that popped up in the analysis that scribe I did. They tested it with real genealogists to make it work the way that a real genealogist does work. I just thought that was a really nice thing. And I also really appreciate that my heritage continues to experiment and maybe even stick its neck out a little bit farther than a lot of the other genealogy companies do in terms of the things that they're willing to try and the information that they're willing to provide because we all know that AI's can hallucinate and they can make mistakes and I did find, you know, I found little mistakes when I tried it with a bunch of different records when I was doing my testing and even with my own photographs. So this is worth pointing out. I said it works with photographs. If you upload your own photographs into my heritage and you go to your photo album, you just open up the image. Scribe AI shows up on the side. Click the Scribe AI button and it 30 seconds later you get back a quite a nice actually analysis of what it saw, possible interpretations of what it saw, particularly things that are genealogically relevant. So it gives you back its best guess at the five W's, the who and when, why of the image that you can use as the basis for clues to come to your own conclusion about what is it you're looking at? I've, you know, could as to my heritage. They continue to do good work with their implementation of artificial intelligence. Yes, deserves to be acknowledged and appreciated that as you often say when we talk about more AI-focused stuff, competition is a good thing. Our industry genealogy is fortunate to have several big players. And my heritage has made it one of their selling points is that they're innovators and the responsible and ethical use of artificial intelligence for genealogy. Always pushing the envelope a little bit and discovering new things. And even if this wasn't a discovery, this is stuff that you and I have been doing and teaching for two years, they're bringing it to the masses in a way that you don't have to know how to write a three-page long structured prompt to be able to benefit from that. We talked back in our prediction episode for 2026 that, you know, one of the big things that we're expecting to see in 2026 is there's a button for that in many, many apps. And this is the this is the big AI button. And you know, at my heritage now, the big AI button is called scribe AI. So whether you're going to a record or you're going to an image or you're going to a coat of arms, there's this big button that pops up beside of this as scribe AI. And so I predict that as they get more and more comfortable with this particular feature and how they build their prompts and also, you know, the technical challenges in the back office, that scribe AI button is going to show up in more and more places. You know, there's a couple of things that worth pointing out that you can see where their the places where they are being careful. You know, if their documentation states pretty clearly that you're not allowed to do any research on living people. So they don't want it to be used as sort of a fine people button. They just want to be able to use it as a, you know, research my ancestors button, which is a really, really good guardrail. And they it's also worth noting that it's even available on a free account in a limited fashion much like we're seeing with all of the others AI tools out there. So you get to click it a few times to see how it works if you've got a free account. But if you actually want to use it a lot. And in fact, the documentation even says an unlimited amount you actually need to have their complete subscription package, which I don't know what it is in US dollars, but in Canadians about 500. So my heritage is not the cheapest genealogy tool in the toolkit, but this gets you access to, as you say, the big button that helps you do analysis on almost any record that you can find at my heritage, particularly ones that are about deceased persons or for very little effort. And I hope that the other genealogy companies follow suit because this is this is almost exactly the feature that we were talking about during our interview with Nicole and Diana last week. Where we were talking about the I want a big button that actually, you know, submits to the AI, the things about this particular round sestory, this particular record, and just tells me what it sees and what it means in the context of my own genealogy research. This is a highly accessible approach that's easy to use for anybody who doesn't actually want to learn how to use AI. So with that, let's let's kick over to our interview that we did with David Vance, the general manager from Family Tree DNA. Well, we were on the show floor at RootsTech. Stephen, I are very happy to be joined today by Dave Vance, the Senior Vice President and General Manager of Family Tree DNA. We're coming to live from the RootsTech show floor. We're so excited to get together with Dave today. One of the exciting parts about your own story recently is you recently moved to Family Tree DNA and took over the top spot as the general manager. What's that been like in the last year, moving into a new job? The stereotype is that you're drinking from a fire hose and it really has felt like I'm not only drinking out, but I'm lying on the floor while they hose me continually. It's been interesting because I've been in this business for, you know, 20 years or more. And of course, I've known Family Tree DNA because it's been around for even longer than I was in genetic genealogy. But you see it, of course, as a consumer, as a customer, as, you know, somebody's helping to lead the field. And it's interesting to be part of it now. There's a lot that goes beyond the scenes that never comes out because they think of great things and they try them out and then they don't work out or they're not ready for it or so on. But just the innovation and the thought process and how they come up with things is really fun to watch because that is an inventive process and it still is. I remember in the early days of genetic genealogy, we were doing a lot more inventing of methods and tools and now it's more about practice and that's all great. But the innovation still goes on and you don't see as much of that now out in the field because it's about practice more than is innovation. Yeah, well, there's been a lot of maturity in genetic genealogy in the last five or six years. And if you're really close to it, it feels like it's kind of been like you say, just maturing. But the last few years, all of a sudden, it's actually gotten really exciting again. Like there's lots of new stuff coming out and, you know, you've just recently, both you and my heritage announced the introduction of next generation sequencing. I think something that a lot of genetic genealogists have been dreaming about for the last decade, what's it been like bringing this to market, something that's been so long on people's minds? That's been part of the fun too because we've been practicing our craft, so to speak, for the last, say five to ten years. But we've been practicing the techniques that were invented and we've, you know, getting better at doing what we can with the data we have. But it's been a while since we've had us a step function in capability. We're planning to get so much data that we're not even sure what we can do with it. We know what we want to do with it. We're not sure how, because we've not had this amount of data before, nobody has. So we're not sure if we can prove the things that we think we can prove, if we can develop the products that we think we can develop. So we are being a little, you know, quiet about what things we can do with it, because we're not positive that everything we promise will come to fruition. However, we have enough that we can think about that I think we'll come out with some really cool things. So that's just kind of a teaser. But I do think we are all, and that's not just family treaty and A, but the whole industry is on the cusp of getting a lot more data available and the methods and tools that we have to develop will follow it. For our listeners who aren't, or who aren't yet familiar with next generation sequencing and what that means, would you mind giving us like the two minute version of why this is a step change in the genealogy? Well, yeah, I mean, about 20 years ago or so, when all the zonal testing first sort of, you know, I broke onto the market. It's always been using what are called microarray geno typing chips, which are repeatable cheap chips that you make. And the best part about it was that they always were the cheapest way of doing a repeatable testing of some genome. But they only collect about 700,000 of your data points from your DNA. And if you know anything about DNA, that's like 0.02 percent of your genome. It was enough to get shared, sent to Morgan's, it was enough to do matching, it was enough to do ethnicity, it's enough to do a lot of things. But it never changed. I mean, the chips have changed in small ways. But basically, that's the amount of data we've had throughout his zonal testing for many years, for many decades. That just recently changed because next generation sequencing allows you to test more of the genome and collect a certain amount, the cheaper ones collect less and the more expensive ones collect more, but you can vary that. And we can come out now with a test that's as cheap as the microarrays, but they collect more like 9 or 10 percent of your genome at least. Now, they have tests that can collect more, but they're hundreds of dollars still. So we found a test that we can offer in the same way as our family finder that gives people, that basically instead of collecting 0.02 percent of your genome, we're collecting about 9 percent. And that's 400 times more data that we can use to learn about your genome and where your ancestry came from and match you with other people. We can get a lot more precise, we can get a lot more data, we can do more things with that data than we ever could before. Do you use the nice analogy between skimming a book and reading a book and we're only still going to be able would read 9 or 10% of the book. But what are some of the first things that we might look for and what would be the surprises? - One thing that we can read in your genome is more of the differences between you and other people. And those differences are important because they tell us you're lineage and how you're connected to others. And so just as an example, we do triangulated groups today in our zonal testing, but we put them all together and then we try to figure out how they're related to each other. With a little more data about the differences that those groups share or differ between each other, you can come out with a better mapping of how their common ancestors are connected and which ones share closer common ancestors with each other than others. So we could, assuming that we can get a level of detail out of this that is repeatable, we could do a lot more with triangulated groups than anybody else has ever been able to do and give you not only a group that matches each other with a common ancestor, but the order of common ancestors that they connect to, which would be really powerful for people exploring their genealogy trying to match those up with known ancestors. So, that's the kind of thing we're hoping to do with this kid. Wow. That's good. Everything's getting a bit simpler. Yes, but more complicated because we have a lot more data to manage as well. Yeah, it's easy to get lost in the numbers, but because people always hear like, the SNP chips were only some very, very small decimal point percentage of the amount of available information, but it was a really important stuff. And family treaty DNA has gone that extra step. They're going to look deeply at about 9% or 10% of the genome, which sounds like some people might walk away from that going only 9% or 10%. But like you said, it's what factor more than we were had before? It's about 400 times more data. About 400 times more data. And I think about the genetic genealogist that I know, they feel like they're swimming in data today. And it's one of the reasons why they like AI, right? Is it helps them find the patterns that maybe it's hard for them to, you can spend five or 10 years learning the fundamentals of genetic genealogy. When we increase that by a factor of 400, like how are, because it's not just about the tests, it's about making it possible for the people who use the test to understand the information that you're showing to them. And I know that's going to take years to work out the tools and the techniques to present that. But what role do you think that AI or tools like AI to help people visualize stuff could play and help me make sense of all of this information that we're going to be swimming in? - That's a good point, because up until now, people have used AI mostly in the visible sense, I think, to bounce their genealogies off and have conversations with the LLMs about instruction narratives that relate to their ancestors. Tell me what my ancestors, life would have been like in the 1850s and that sort of thing, right? Which is all good stuff. I think we have a lot more use for behind the scenes AI and this gets into different kinds of AI, like predictive AI and agente AI and things that we don't see as much, but that help us analyze the data, make some conclusions from the data. And it's hard enough to ask people to understand the basis of 700,000 SNPs that they have in the shared sentom organs and how they match together. It's going to be impossible to do that with 280 million pieces of data. Let alone whatever else we do beyond that, right? For the next wave of innovation. I think that's an opportunity for AI really to step in and say, okay, I will give you the conclusions that you should have drawn from this because there's no way you're going to be able to analyze this constructively. And I can do it hopefully more reliably that'll be the key, but more repeatedly and faster than humans could do themselves. - We teach some sentom organs, which most people don't understand. We teach the megabase pairs, which most people don't understand. But they still learn enough of the lingo and how it matters. And now we're going to add on new tests and people are going to spend the next two years and you can already see it starting, trying to figure out how to map the relationships between tests, right? Because they're going to be, well, how does this compare to my old one? So there's another layer of abstraction that's going to get created for the next little while while the dust all settles. So there's a lot of room in here for, what does this really mean to me? - There's a lot of things that we hope to do back to your question, Steve, which was, we hope to get better than to get more granular than the seven centimorgan limit that we have now for usefully analyzing shared segments. We hope to be able to better differentiate between pile up regions identical by chance and real matches, right? Because if you can get under the covers of your shared centimorgan, you can get better resolution as to them. And you can know that something is identical by descent and not by chance just because of the patterns of the mutations that you see. And if that breaks through the five to eight generation limit that our disomal has, that would be really powerful. Now, obviously, if you didn't inherit some DNA from a particular ancestor, any DNA from a particular ancestor is nothing we can do. But we can push the limits of going, small segments going back in time, hopefully. But if we could bring those kinds of tools to market, that would be really powerful too. - But you're talking about a different type of artificial intelligence happening behind the scenes, a genetic where you're doing math and calculations and using old school artificial intelligence perhaps mixed in with a little bit of generative, but it's not guessing. It's helping you figure out how to do the math and the biology and the coding better behind the scenes. Do you get deep enough in the weeds to know a little bit about that? - I do know that it's happening and to admire it. I'll say, how much of it I understand is depends on the particular subject. But I think you're right. I think the analytics will happen faster now going forward. I think the generative AI, I'm still a little skeptical of its ability to do the internal semantic networks that are necessary to truly understand what it is that it's talking about. I mean, we've had a lot of experience with generative AI and people coming to us saying, well, here's what Chats GPT tells me about my ancestry. And 90% of it may be spot on and then 10% of it is wholly, completely off the rails. And unfortunately, people believe as much as, because they have no reason to doubt it, but we look at that and go, okay, we can see where it's right and wrong. I think that's a semantic failure for now. And that will improve, but I think the expert system side that generative AI relies on has to improve a little bit first before we can truly rely on that to be 100% reliable. Yeah, because there's two pieces there. It's a really interesting point to grab on to for a second. There's the two pieces there in terms of making the results good. One is really strong starting data. And you're going to make the starting point 400 times better for a good hunk of the genome, the important part of the genome for genealogists right off the hop. So that's going to get way better. So you're going to better signals to make smarter models. But then there's also the model. Like you have to have some way to vet what good looks like and to pass it through to the user or to your guys back in system that draw some conclusions on our behalf, right? So that's, I assume that's going to be, there's people who are hard at work thinking about how they're going to take advantage of all of this information. One of the things I'm a bit worried about is if the explosion in the analytic AI side gets wider adoption within the industry, it actually reduces the ability of generative AI to make conclusions and draw inferences because they don't have access to that level of detail. All they see is the outputs and the reports, but they don't know how that was come up with necessarily. So that separation will almost work against the ability of large language models out in the public sector to be able to use the data effectively because all they'll see is the conclusions and they'll have to rely on that. So there's a little bit of leapfrogging and capability that'll probably happen between the AI's before we catch up to that. Steven, I have noted a lot that there's a big overlap between the people who have a great interest in genetic genealogy and the people have a great interest in artificial intelligence because they're both sort of analytical pursuit set of level. This is where the genealogy geeks hang out is in sort of this corner of the genealogy room. The information that is going to be made available, do you think that we're going to be able to take that information and be able to analyze it ourselves at some level? How much do you foresee making available to us to be able to use in the coming six months? - Yeah, that's an interesting question because I think making it available is not the problem. I think the problem's going to be the volume of data and the ability to make the ability to make conclusions yourself manually, I think will actually decrease. So unfortunately that works against the citizen scientist and it's going to get harder. Now that's actually most industries have grown up like that, you have the people who might, if you think about the invention of the car, people made kit cars long before they had companies that made cars and people made them in their garage and drove them all the time. I mean, we've gotten beyond that stage in genealogy but it's still kind of a, oh I thought of a good tool, I'm going to make it public kind of a field and that's or an interesting AI, I mean AI prompts are the new version of that, right? I created a great AI prompt to do X. I put it out there on social media, people copy it, people do that stuff and that's where AI I think has really revitalized the citizen scientist feeling among the practitioners in genealogy. I don't know how long that survives in a world where you have to feed something to AI just to get the conclusions how to chew out. And this is because just the amount of data we're talking about is it 400 times as much when you go from 0.2% to 9% or 10%. So it's just prohibitive to be able to give consumers all that information and expect that they would have the computational power to be able to process. - Right, I mean, it's their data, they have a right to it. So it's not about not sharing it, it's about them being able to use it for some purpose. I mean, there are plenty of people who use band viewers to look at their detailed results now, and that will continue, that won't change. And, you know, someday we'll be able to get 100% of the genome in the same way, for as cheap as we do, what is almost testing today. I mean, that'll happen, maybe not in my lifetime. (laughing) Oh, you got lots of time left. - That's right, but, (laughing) but, if we get that far, of course, then we're talking in order of magnitude. We're basically talking about six billion pieces of information for every individual. You may be able to analyze it yourself, but being able to compare it to your matches, usefully and do conclusions based on it in groups of people that you have to do to learn something about your knowledge. - That's the thing that most people, like, I, I, again, watching people try to do this, they forget that it's not just their DNA, it's an exponential problem, 'cause when you get into the matching algorithm, when you start to build your own matching approach, you're, call it 100 data points, right? You're 100 data points and Steve's 100 data points and Dave's 100 data points, all of a sudden, it's not 300 comparisons, right? - Right, you know, it's 100 times, 100 times, 100. - But as more people test that volume of data grows from that alone, you know, let alone the testing capabilities. I mean, I always tell the story in my wide DNA studies, I have 20,000 Y12 matches. I cannot usefully analyze 20,000 matches, so I haven't looked at that level in years. And I can look at the other levels, I have fewer matches because it gets more precise, as you look at higher levels of wide DNA. But the Y12 level, there are people who usefully analyze that and I can't, because there's nothing I can do with 20,000 matches. And that's just a matter of the fact that a lot of people have tested and I happen to be in a very common Y12, you know, group. But, but, went, but that problem will get worse for people as more people test, let alone how much data you have accessed. - Even if tools were to become available, I know folks that still like to use paper and pencil to do some of the aspects of this work. But other folks just want to put a token in the machine, pull the lever and get the answer. Is a path starting to become available, or can you glimpse where how far out we may be to it's putting the tree together for you? - I think we'll get to a point, we're already at the point where a lot of tools suggest your tree. And I think the key there is you have to do the work yourself to be able to validate the suggestions. I don't know that we'll ever get away from that, just because records can lie and may not be right. And, you know, DNA doesn't lie, but the, but the conclusions you draw can be statistical, can be predictive and we'll get better at the predictions, but we'll never be 100% in some cases. So, and hopefully what will happen is that the computers will get smart enough to suggest what is reliable and what is not and be able to tell us where we need to focus our attention to prove out the suggestions that's giving us, which has always been true with genealogy. I mean, we all know the books that have been written by our ancestors that we think are reliable and turn out not to be, right? So, there's no, we're certainly no strangers to data that we have to take with a grain of salt. And I think that will be true in our field as well. - One of the things that I appreciate, you just said, DNA doesn't lie, but it can be misinterpreted. Language models, they had no compunction about lying. They will just put in the next word if it sounds right. But one of the advances we've seen over the past two years is we know how to mitigate that. And you can see a path where we're gonna be able to look at the records and where the records are wrong. We're starting to develop the generative tools to identify where there's mistakes in the records. So, you can see these two worlds kind of coming together where the DNA is you're able to analyze more and the AI's gonna get better at interpretation and analysis of the records. And it won't only be a human only task to identify mistakes in documentation. But it still seems a long way off. - Yeah, that's a very good point. I think the other point that you just said, which is also important is, you know, people look at large language models as, it's great because you can just converse with them and ask questions and they'll give you answers. The truth is, prompt generation is an art. It truly is. And we all have to get smarter at prompts because of the, not just hallucination, but the tendency for large language models to give you the answer that you were asking about, or not to give, or going off and giving you slightly different answers because you didn't give enough, you know, guardrails to your question and that sort of thing. And I think our ability to use AI has to get smarter too. So there's that downside is, I don't know that it'll ever get smart enough to overcome our own tendencies and that's going to be a difficult thing to overcome. - The same problem that genealogists suffer from the unintentionally inserted bias into your research plan. - Exactly. - AI can just lean into that on your behalf. (laughs) Although you bring up something that's really interesting, Steve, that is, might be worth a follow-up. I think the ability to provide, not just an interpretation of the information that comes back, but also, well, based upon the results that I've shown you, here are the other things that, how this could be interpreted. Like it goes a step beyond it being, you know, 65% likely to, well, given that it's 65% likely, you might want to check that and that as well as the other thing. Like a research plan that is connected to the results that are received. Like that's the why do I care part as opposed to what did the data tell me part? And that's a generative capability that I think all of genealogy would benefit from it, but genetic genealogy is screaming for it. - Yeah, no, I think you're right. We have predictive tools in use today that you feed in your family trees, you feed in your information, you feed in your DNA and it models how the people are connected and it gives you predictions as to which one is best. I think that's a perfect opportunity for AI to take over and give you those predictions and rank them and give you with knowledge of your family tree, with knowledge of any data sources that it's allowed to access, pulling all that together and giving you those conclusions without forcing you to do all the analysis manually. I think that's a great advance. And I would look for, I would think that would happen sooner than later. - Oh, people are going to be creating those prompts. Well, people are already creating those prompts, but people are going to be making projects out of those for every single genetic genealogy. I think that's the kind of place where citizen science is going to really come out. I mean, even look back at your history. I mean, you built a lot of the tools that a lot of genetic genealogists use today. That was one of your on-ramps into the family tree DNA world. - That's correct. I would correct you to say a few tools, but yes, you know. (laughing) - I appreciate that you are aware of a genetic AI. Probably most of our listeners aren't. They may mark and I've teased about it was the word of the year for 2025, but nobody ever used it. But now we're really starting to see a genetic AI. And so that folks understand what this next step increase in functionality, it's no longer just chat bots that's responding to a question and answer thing, but these tools do work for you. And so you can almost teach them the methodologies that a genetic genealogist would use. And what the mind blowing part is you don't just have one of these. You could have a team of AI genetic genealogists given a task and you could say you're on team A, you're on team B and you're on team C. Go look at this genetic genealogy problem, come back and tell me where you agree and disagree and why. I mean, that's closer than we realize. And I think there's some folks who are leaning in real far and they're going to help you be trailblazers and understand this. How do you strike the right balance of what do we make available? Is it just what the technology can do? Or do you also talk and listen to the customers and see where their comfort levels are and strike those balances? That seems like there'd be a real challenge. I think it's our job to offer the capability to people. But I do think that the increase in complexity and opt-ins and consent is part of our journey over the next last 10 years. And that will get more complicated. And I do think a genetic AI needs to have those guardrails, at least for the foreseeable future. I mean, honestly, we're not in an industry where the actions that the AI can take include things like pushing the button for the nuclear weapons to drop and things like that. So we're pretty safe on the actions that we can ask at a genetic AI to take. But I think we would offer the consumer the ability to say, "Okay, now show me what records you're changing before you do it. Show me what your conclusions are and let me oversee your proposed changes before you go into my REC, not DNA. Obviously, those will remain the same. But maybe your family tree, it'll update your family tree in different ways and you may not want it to without checking that cross-checking it either." But yet, people may decide that they trusted enough to say, "Okay, go ahead, maybe after they've had practice with it." So I think there's a level of opting in that will come out of that as well that says, "Okay, don't make the decisions for me. Show me what you're going to make and then I'll push the button to say, "Okay, go ahead and do it." Like over the course of the last three years since the AI revolution has hit the genealogy world, like I see some people that are getting, like they're more nervous about it than they've ever been. And I see some people going, you know, damn the torpedoes. Let's lean in. Are you getting any signals yet from your, from either across the greater genealogy community or from, you know, family treaty and his customer specifically about whether it's like we're ended up with two polls, like afraid, or we are actually starting to shift more one way or the other right now. I don't think I've seen a shift and this isn't something that we measure, you know, specifically, but it's, but it's about, you know, keeping sort of tabs on the industry and the discussions that are going on and things that are on people's mind at this. So it's anecdotal more than, you know, actually database, but I don't think I've seen a shift. I've seen it be a spectrum that continues to be a spectrum, right? And people come at it from different, pizza, from different points on that spectrum and they're more comfortable. Now people do change. They don't stay on the same point on the spectrum, but they still, you know, there are lots of people that are very afraid of where we might end up. I mean, again, the genealogy world is, is relatively hummus in that sense, but people don't want the AI making decisions about my ancestry either, you know, but, and then there are people like, oh, good, I can ask it to print out my family tree and I can send it to my family. And so that, but that's always been true in genealogy. And I mean, I have the guy that I talk about who, who leaked to my ancestors on ancestry and history goes back to Thor of Ezgard. And so, you know, it's just, there's always people that take it to a degree that you wouldn't or vice versa. Yeah. Well, and I think, you know, the family tree DNA community is largely regarded as sort of the geeky end of the genetic genealogy spectrum. I mean, you make so many different tools and so much data available through all of the various testing options. Like, I think, you know, you have a very special perspective on the community that I think is going to fit very well with, with next generation sequencing, like they're ready for it. And I think, you know, and I think I'm, I'm hopeful because we've seen so many great tools and methods come out of the greater community, or just around specifically related to genetic genealogy. I think we're going to see a lot of people try to figure out how to take advantage of that new information, even if it makes your computers catch on fire occasionally. What's the hardest thing you ever had to learn to do in genealogy? Before I got into DNA testing, I would have said that was to learn how to properly use a microfiche reader. The hardest thing I think was understanding the sequencing technologies, to be honest, I was not a geneticist to start. So that I forced myself to do that when I was a citizen scientist because I wanted to understand the basis for the things that I was looking at. I've never been much of a band file viewer. I know plenty of people who were really into the deep side of that and I respect them for that. I've gone as far enough as I want to to learn that, but I think that was my biggest journey was figuring out what the underlying science behind this was, enough to be able to be dangerous about it. Well, that probably brings up the most important question is the role that AI does or doesn't play because a lot of people get into genetic genealogy and genealogy in general because they want to learn about the family, not because they want to learn how to use tools. What are the things that you think AI and the technologies, even of genetic genealogy, are never going to replace in that journey? Well, that's a tough one. I think there's a couple of sides of genealogy that again go back to the practitioner and what they got into it for. And I don't think AI will ever or even the tools that we have to offer will ever replace what you take out of it and what you learn about it and what was most important to you because those are very personal questions. But the other side of that is I think all that this will ever be able to do is offer you as much evidence as can be found in it is out there. What you conclude from it and what you believe from it, you will never be able to push a button and get you or proven 100% ancestry back to whatever you'll get a lot of good clues, you'll get what the data tells you. Some of it will be inarruable and at some point the overwhelming weight of evidence says I have to accept what it says, but there's always room for subjectivity. There always has been a genealogy and there will continue even if we get better at collecting evidence and making conclusions from it. And I think that last mile so to speak or that last step, if you will, is where people will always have the ability to answer the question. Not only is what does this really tell me, but what does it mean to me? And I think those two questions are the fundamental questions of why most people get into genealogy in the first place and I don't think that's something that will ever get away from having the human factor. That's wonderful. I couldn't think of a better place for a wrap. Thank you very much, David. It was great to see you at Roots Tech and I look forward to seeing you again next year. Well, thank you, Steve. Thank you, Mark. Thank you. Thank you. Oh, I always love talking with David Vance. You know, there's going to be so many interesting things that come out of the new next generation sequencing tests at my heritage as well as at Family Tree DNA. Just so much more information, so much more potential complications that come from all of that more information on genetic genealogists. Genetic genealogists already have to be data analysts, you know, by night in order to be genetic genealogists by day. And these new tools as exciting as they are, they create all kinds of new challenges that we're going to be struggling with in the genealogy community for the next few years, well, we kind of figure this out. So I know I've got my heritage test, my new my heritage test already sitting on the shelf to do their NGS test. And now I need to order a new Family Tree DNA test so that I can do the same thing there. So we can actually do some of our own real world testing with all of this new information that these tests are going to create for us. Excellent. Looking forward to hearing more about that too. So let's jump into rapid fire now. We've got three great stories today. And the first one is a big announcement from Fold 3. We are revolutionizing a revolutionary war pension files with full tech search. And I know you've dived into this one this week, Steve. What did you see? Yes. Well, for folks who are unfamiliar, Fold 3 specializes in military records. They are a part of ancestry. They're part of the ancestry family. And on March 10th, Fold 3 announced that they've used handwriting recognition technology to make 100% of the National Archives Revolutionary War pension collection fully searchable. Now that's 2.3 million pages across almost 3,000 roles of microfilm covering pension and bounty land warrant applications between 1800 and 1900. The big deal here is that you can now search any name mentioned anywhere in a pension file, not just the pensioner. If your ancestor appeared as a witness, provided an affidavit or was referenced in a fellow soldier's application, they are now findable. It's a brick wall breaker for a lot of people. Yeah, that's great. This is just like family searches, full tech search. The fact that they fed these large information, dense records through this OCR tool, it was such a game changer with full tech search. It's incredibly difficult to index those complicated records, but this must have been just amazing to look through. Yes. Anyone, any genealogist and there are many, many of us who have been amazed for the past two years now at family searches, full tech search. This is Fold 3's initial implementation of that. They just like when family search made that first batch of probate files, full tech searchable. The magic of this is it takes a document. Actually, it works one page at a time, but it takes an application that may be 30 pages or more. It does handwritten text recognition on those and then indexes that handwritten text recognition output so that it truly does make discoverable people who are just merely mentioned in the application, which is different than an initial indexing, may just mention the principles or one or two key people in the file. This really breaks open the searchability. You can see that this is just the first step in a wave of other documents will follow. The fact they actually announced that this was just the beginning. They will be doing more records in the coming weeks and months. This will spread not just from Fold 3, but through ancestors, greater document collection as well. Just for comparison sake, did you find the implementation, the way that they set it up was consistent with or remarkably different than the way that full tech search does it? Troz and Kahn's, if you've used family searches, full tech search, you will find this very familiar. But Fold 3 did implement some things that if family search or others haven't already done, I expect they will do. One of the nice things they've done when the initial transcripts are already done. So when you're looking at the image of a page from a revolutionary war pension application, you can click the more information button, the little red lowercase i, and that a tab, a sidebar, put flies out from the side, and you see a rough but good initial transcript. It maintains line breaks and even some misspellings will be preserved. What is what might be called a diplomatic, a literal transcript. It just looks like as if the handwriting were turned into tight text, mistakes and line breaks and all. But here's the innovate and that's we've seen that for a long time. The innovation that I really like is beside that there's another tab that says summary. And as soon as you click the summary tab, it takes that raw, rough initial transcript and turns it into something much more usable. It summarizes the information just on that page and then it extracts the names that are mentioned on that page, the dates that are mentioned on that page, and the places that are mentioned on that page. You can tell they've written a script that says, take this raw transcript and give me a 75 word summary and then a list of names, dates and places. That's very, very helpful. The one thing that was missing that I hope they'll do is that special polished smooth feature that my hair did, including the copy button. Oh yeah, the full three should acknowledge that people will want to actually use this information and just put the copy button right either at the top or the bottom of your transcript and summary boxes. But even with that, that just takes you from 99% to 100%. That's that last little polish. There's one other thing that I would highlight Mark, they do acknowledge that fold three acknowledges that this, especially the summary is using Ancestry AI. Ancestry is branded artificial intelligence. And when you click learn more, it actually takes you to their responsible use page. Ancestry is putting together a very good ancestry principles for responsible artificial intelligence. And so we'll take a deeper dive into their bigger philosophy about this. But needless to say, it's very well done. I'm so glad to see that this is becoming a more broadly available capability. I think within a few years, this is going to be table stakes for every single record set that gets brought to one of our genealogy records company. Oh yes, yes, this is going to become, well, as you said, table stakes. Just this is the basics of what's going to be expected that soon it will be the vendors who the companies that don't offer this function are going to be going, well, why not? Mark, there's something closer to home. I know that you've been working on something for our second rapid fire story. You are known as an Excel maven. You know more about Excel than most people ever get to. I know a big theme we've talked about for years now has been these tools are going to be appearing everywhere, percolating down to all of our software. And this was a big one for a couple of different reasons. So tell us more about Claude for Excel and why it's important for genealogists and the broader knowledge economy. Yeah, so speaking of things that are still not perfect, but that doesn't mean that they're not incredibly useful. Welcome to Claude for Excel. I've been watching Claude for Excel as well as the other tools with respect to spreadsheets, which is co-pilot. Co-pilot also has with Microsoft has some built-in features with Excel. And then also Gemini is integrated into Google Sheets. So all of them have been working on trying to put the button, put the AI button inside of our spreadsheets now for coming up on two years. And up until about six months ago, they were all uniformly terrible. Every time I tried it, I spent more time arm wrestling with the feature than I did actually benefiting from it. And I, you know, that's one of the reasons why we've never talked about is because it just wasn't good enough. But the allure of using them is just so high because it's so hard to get information as hard as it is to get text in and out of your chatbot. Spreadsheets are even harder. Spreadsheets are a really, really complicated document format. There's not just a table inside of spreadsheet. You could have tables and charts. You could have three different tabs and four different tables and nine different charts. So when you, when you want to get that information into your chatbot to get help or guidance, it's been pretty complicated and kind of clunky. And so just a lot of people don't even try. But about six months ago, the these spreadsheet integration plugins started getting a lot better. And right now, I think the best one that's out there is Cloud for Excel. They released a version in January that finally got usable. So what this means is there's actually a piece of software that you actually have to install on your computer. And so you go to the cloud website, you search for the cloud plugin for Excel, and you download a piece of software and then you install it into Excel. And what this gives you, once you actually install it as a plugin is access to cloud directly in a side panel within Excel. So you can actually have a chat with your chatbot about what's going on in Excel without having a copy and paste stuff back and forth to the chatbot. You can just say, take a look in that left table and tell me and describe it as what you see. Summary's the table or even more importantly, you can say, please add a new column to that table or take that table of information that's about my research report and create a chart from it that shows all of the events that are captured over time periods. So it doesn't just read it. It can actually update the table as well. It takes a long time to learn how to use a lot of those features and functions. And so just to be able to say, I know the feature that I want, just go and do it. That's just going to be a huge benefit for so many people who haven't put in the thousand hours learning how to use Excel. A couple remarkable things about this tool and and I've been I'm pretty good at Excel, but I'm not an expert. But knowing knowing that there's things that I would like to be able to do in Excel, but I just don't own that book or that manual or I have that my fingertips, the instructions on how to do it. Having that available to me in Excel is a big deal. And folks, especially folks who've been using Clawed for a while that recognize that there's something just a little bit better about Clawed during this season of the AI revolution. It's pretty neat. And I'm glad to see some of these Titans of industry working together so that they realize that if Clawed's the best model, the best tool for the job that Microsoft's going to let them inside the tent a little bit. Well, you know, today's just a rapid fire episode. So we're not going to go too deep, but but the one thing that I will say, like all AI tools, they are best wielded by people who are already familiar with how to do the thing that they're asking for help for themselves. And and and Clawed for Excel is no different. I find it regularly doing ham fisted and just outright wrong things when I ask it without direct prompting. So like every single time and something's like, why the heck would you do that? That's just a bad way to build a spreadsheet. So that being said, it's really good if you already know how to do the thing, you just can't remember how to do it or you don't want to be bothered to do it because it's a really a finicky thing like, you know, creating a chart from a table and formatting all the axes correctly. Like that's the kind of thing that most people have to look up every single time that they do it myself included. I know how to do it. I know that it can be done so that I know how to ask for it, but it's kind of a it's kind of a finicky thing to do. This is perfect for that. So that's the first thing is it's it's you got if you know what to ask for like all AI, it'll get it done. If you don't exactly know how to ask for it and you don't have to verify that it did it correctly, you could get yourself in a bad situation. So that's one point. The other thing is that I find it is very inefficient like it's token heavy. It sends a lot of information back and forth to cloud and it's very slow. I find it's great for what I want to do something and excel well I'm working on something else like I'm building a presentation or I'm creating a report or whatever. So I have cloud running on Excel. running in one window and I've got my, you know, word running in the other window. And so I'll give guidance to Claude about how to do it and what I want it to do and let it run for four or five minutes and let it do its thing while I'm working in the other window. So that if you're comfortable with that, it's going to run for a little while. By the way, it runs way faster than I do to do the thing, but it still takes a while to do it. It's not, it's by no means as an instantaneous or efficient. So you got to be careful you don't chew up your daily allocation of Claude credits while you're doing this kind of stuff. But that being said, I'm finding it incredibly useful and I'm using it more and more as it gets better and better. So maybe one day we'll do a special episode just on the best add-ons plug-ins and unit taskers. I've got one more story to share with you Mark. This is one that could be a little bit of interest to genealogists. Wikipedia made an announcement just about 10 days ago on March 20th concerning artificial intelligence and articles on Wikipedia. And the headline was a bit more splashy and click baby than what ended up I think being a very reasonable statement. The story is this about 10 days ago on March 20th, the English Wikipedia editors voted 44 to 2. That's 44 to 2 in a full landslide. Yeah, we'll call that a landslide. During their formal process for what they call a request for comment to ban AI generated content. And so the policy they voted on is straightforward. The use of large language models to generate or rewrite article content is prohibited. Their reasoning is worth quoting. They said that LLLms large language models can go beyond what you ask of them and change the meaning of a text that it is not supported by the sources cited. That Wikipedia, the single largest reference work ever created, the largest encyclopedia ever created, saying that AI generated text is not reliable enough for their standards. Now there's a couple things we're pointing out about this. One, I would have agree with their statement that large language models can go beyond them and change the meaning of a text that it's beyond what the sources cited say. That is true for beginners. And you cannot expect every Wikipedia editor and every Wikipedia author to be a skilled AI user. It's certainly possible to do create articles that would be worthy to be included in Wikipedia, but a beginner is not going to be able to do that. And so especially if you had a reluctant or just a new user of these tools, they are not going to create an article that's of the quality needed. Yeah, well, it's worth noting like for anybody who's ever posted or edited a Wikipedia article, there's not 44 people. There was 44 people that you talked about. Those are just like the people who hold the editorial title. There are thousands of people who submit and modify tens of thousands of people who submit and modify English language articles. So this is the big pool, right? So I can see how they'd be very fearful about how it would completely upend their editorial workflow that they use today. So I'm kind of, I'm actually I'm pleasantly surprised. And I would say fully supportive of their stance on this one. I mean, they're essentially, they run the worlds in cyclopedia. And that's essentially what Wikipedia is for better for worse. It is the most used and largest in cyclopedia in human history. And so race for today. Well, the fact that they wouldn't want to introduce AI into that mix and the potential for AI slop into that mix is fully understandable. And I personally, I think it's a great idea. This is a makes sense, a good prohibition, but it's not a complete prohibition. They did carve out what I thought were two very good exceptions. The first is this editors can use AI to suggest basic copy edits to their own writing. Now this is one of my first props. I wrote a prompt like this that if you did not want to have AI creating your work, but you want to grammarly on steroids, it does not rewrite your stuff, but it checks it for grammar spelling mechanics. All those things a good editor would do. And it does not edit your text. It creates you a list of flagged concerns. And that's exactly what Wikipedia said. They said, yes, you can use AI when you're you've ratchet it down to grammarly on steroids. And the other exception I thought was pretty reasonable. They said you could use it as a first pass translation, but I'd love this even to use this exception. You can use it to translate from one language to another, but only if you're fluent in both languages, the source language and the target language. And that's one of the things that we've learned that is critical because if I use AI to translate something from French to English, but I don't read French, then I'm not qualified to verify the accuracy of that translation. Wikipedia wisely said, you can only do this, therefore, if you speak both the source language and the target language, I thought that was kind of brilliant. Yeah, you can see that they didn't come to these conclusions shooting from the hip. The folks that actually run Wikipedia and by run, I mean volunteer their time to actually help create and manage the world's largest encyclopedia because most of them are volunteers. These folks have been thinking about this for a long time. So the conclusions that they came to, I think they're very, very well thought out and considered. So I'm really glad to see I was a little bit worried about Wikipedia in the age of AI because AI Slop is going to infect a lot of places. And if I think this is going to help preserve the quality of the Wikipedia, you know, encyclopedia with these ideas, it's not going to make it worse. I think it's actually going to help preserve it and keep it like with, as we've talked about before, like the introduction of AI Slop is going to forever change the nature of the internet. I'm really glad that they've decided to hold back on this with Wikipedia. This is not a reactionary, luddite rejection of AI. This is actually a well thought out and nuanced policy debt for in the spring of 2026. Seems to be, I think it's a good move. Every single organization is going to be faced with some form of this question. How do we incorporate AI into the thing that we do here? And some folks, some organizations are, you know, they're leaning in and some folks are holding back. But the most important thing is you have to make the decision because it's going to be a big one, right? Like whatever side of the fence that people land on on this one, you know, down the torpedoes versus, you know, hold back and avoid at all costs. Whatever decision is that every organization makes is going to affect other organizations going to be. And, you know, for the next turn of the wheel, the next, you know, two to five years. What a great episode, Mark. I love that we got to hear the interview with David Vance. We have one more recorded interview for our next episode. Do we want to tease that or just surprise folks? Which means we'll tease it. It's another person named David. Industry insiders. We'll see how many industry insiders can discern who we might be talking to. Well, I'll tell you this. It's very much worth waiting for and looking forward to. Thank you, thank you, thank you, thank you, Mark. Really appreciated all of this. Thanks, Steve. And thanks everybody out there in AI Land. We will talk to you soon.