Until now, I’ve shied away from writing anything about using AI tools for genealogy. I’ve shied away from using it too. I’m not using at all in my professional work. There’s lots of reasons for this including the two biggies:
- Intellectual property theft to train large language models
- Environmental concerns about data centres which need vast quantities of clean water to run & use a lot of energy
It’s also important to clarify what we mean by “AI”, which is really just a marketing term. There are different levels of technology that on some level involve computers. For example, Optical Character Recognition, or OCR, is what we use to read newspaper and handwriting. This is not the same as:
- Large language models like ChatGPT and Google Gemini which generate text based on language patterns and probability
- Synthetic image generators like Dall-E or Nano Banana
It feels like you can’t look at a piece of software these days without some form of AI being flung at you. Copilot would like to help me. Every genealogy site has some chatbot. Even my recent purchase of a washing machine saw different models trying to tell me I needed an app for it which was AI-assisted.
Practically every poster you see now has been designed by an image generator, and boy can you tell because the art looks awful. My mother was an artist, so I deeply appreciate the time and skill it takes to create something with a pen or a paintbrush. It’s even creeping into books. A recent discussion in a genealogy book club Facebook group uncovered a series that is largely written by AI. A book I’m reading this week had a line about someone having a “Celtic Tiger tattoo”, which immediately made me wonder if they had used an LLM. The Celtic Tiger is a name for a particular period of Ireland’s economic growth. It’s never been a large cat.
At the recent back-to-back conferences in Glasgow, we were treated to several lectures on the subject of using artificial intelligence in our research. All but one were overwhelmingly in favour. But it was the negative one that resonated most with me, delivered by industry professional, Laura House.
LLMs are very positive. If you’ve interacted with one, you’ll understand what I mean. It’s like your supportive best friend, always on your side, a cheerleader. “That’s a great idea, Claire. Now, let’s try adding in this…” This positivity can lead to mistakes. A summary of a document might miss the one unique point in it because LLMs look for the most probable patterns. If you don’t read it yourself, you could miss the most important part. If you’re not already well-versed in the subject, how can you be sure that its answer is correct?
Laura referenced a university lecturer called Professor Emily M. Bender in her talk. Here’s Prof. Bender on a YouTube show discussing it in more detail. It’s 30 mins long but well worth the time investment.
Everyone will need to decide for themselves what’s the level of usage they’re comfortable with. For some, it’s “none at all”. Others might be happy to use it for some things but not others. I don’t know what the right answer is, but I do want to read more, think further and consider the long-term implications before deciding what my level is. What’s your thinking? Let me know your thoughts (preferably not written with the help of AI!!)
I tried out ChatGPT when it first became prominent, just for brainstorming titles for talks. But I never found it that useful. That talk by Laura was brilliant and we need more people deconstructing this AI hype and spelling out clearly what it is and isn’t. As you said it seems to be everywhere. Everytime I open up PowerPoint I have to close down the Copilot icon because it’s deliberately positioned to be in the way to force people to click on it. And I see so many other genealogists getting lured in by AI because of the promise of efficiency.
I’ve disabled it in Microsoft stuff. The version in Adobe Acrobat is very “in your face” and it’s making me think about switching to another pdf reader.
I have mostly used AI (NotebookLM) to create short videos from my family stories and blog posts. I can see using it to sort things out. I recently had 1`0 Muster Rolls for an US expedition into Mexico in 1916. They all had my grandfather’s name Jeremiah Sullivan. NotebookLM determined they were 2 different men, as the dates and locations overlapped. That was helpful, but I still reviewed it myself. I can also see where it may be helpful in a “Where do I look next….” scenario. But I think I will still be doing the work myself.
In that scenario, it sounds like NotebookLM just sped up the process. It’s funny, I had a very similar question the other day but was trying to work out if 2 military records were the same or different men. I just took a piece of paper drew a line down the middle and wrote things out! By the end of the page, I’d come down on the side of them being one person.
I used ChatGPT recently to try to help me transcribe a difficult will. The experience was frustrating, and also, quite frankly hilarious, before I gave up! I know prompting is key, but despite me being very clear what I wanted to see when the words still remained unclear, what I got back had chunks of text ignored as if it wasn’t there and streams of made up legalese padding out other lines. When I re-prompted to tell the model to stop doing that, it then descended into hilarity, by chastising itself repeatedly, saying what it should have done in the first place and then chattering on about what it would do better next time? I had to laugh though as the conversation ended with the response, I kid you not – “Yes, you told me not to do that, but I did it anyway”! I went back to transcribing the will myself and made a much better job if it once I got my eye in.
Having said that, I have used AI to help me construct location guides, specifically for potential Roman Catholic burials grounds in Armagh, which was more successful. I’ve also had some quick wins recently with finding named locations on maps by just Googling “where was location X in 1894” for example. This has proven very useful for finding properties or churches that have since been demolished and beats wrestling with sat nav maps in Family Tree Maker. I’ve had no success with trying to improve photographs though, with one attempt repeatedly turning my uncle into a woman?!
We need to keep challenging the “answers” that AI feeds us when it mashes together everything it scrapes brazenly from the work of others and remember it can never bring new original thinking to bear on a problem. That’s what fellow genies are for!
Great example. I think Transkribus is better for handwriting transcription, and it doesn’t provide its own commentary or cheek!
Call me old fashioned, even a Luddite but I’m just not even tempted to use it all. For anything.
Lots of people in your camp, Rosaleen!
AI can make genealogy research more efficient by helping researchers identify patterns across large collections of historical records. The real value comes from using those tools to narrow the search while still verifying important findings against original sources.