Unveiling the Past: A Genohistorical Journey into AI Handwriting Transcription

Unveiling the Past: A Genohistorical Journey into AI Handwriting Transcription

Everything changed this week. I’m not exaggerating. Finally, I gave artificial intelligence (AI) a serious look, and it is already transforming my genohistorical research, my business, and my life.

Introduction

The last time I gathered with my extended family, my nephew enthusiastically told us about AI. The technologically curious of our family gathered around as he described what could be done. He even had AI compose a new song about us in seconds on his smartphone. He told me it could create a digitized painting of whatever I wanted to see.

I have wasted many hours trying to find the perfect picture to go with blog posts over time, so that had my attention. My first experiment with AI was in using it to create an image that said what I wanted it to say, and you saw that with the last blog post. I was using a free tool called Dream Studio.

Already this technology had made things better. Finally, I can have the image I want without searching for a lesser alternative. But I can’t say that I was serious about seeing what else it could do for me.

This week, I listened to the playback of a program I missed from the Association of Professional Genealogists’ Writers SIG. In it, Stephen Little described the future of AI and genealogy. I paid real attention for the first time to what it’s going to mean for us, especially in terms of interpreting handwriting. Think of the history and records that will become available in a searchable form.

Stephen encouraged us to get the superior technology of a paid AI service, rather than settling for the free apps. The free apps simply don’t begin to show you what is really possible. I purchased a subscription to ChatGPT Plus—Stephen’s recommendation—and, today, I got serious about testing it.

AI Taught Me

My nephew told me that the ChatGPT Plus tool will be my best teacher, and that I can learn by asking it questions. There are books, and I am sure I will explore some of them. But this technology reinvents itself so fast, a book on it will border on obsolete before it gets published. So I just opened the AI dashboard and dove in.

Curious about how far it has come in interpreting old handwriting, I typed a message to it asking, “Do you interpret handwriting?” It told me it did not yet, but would be.

I realized that I was using chat GPT 3.5, which is the default. I changed it to ChatGPT 4, which is more advanced and available with some limits in ChatGPT Plus. I asked it the same question, “Do you interpret handwriting?” This time I got a positive answer, encouraging me to upload an image. I wasn’t listening carefully and uploaded a PDF of a full set of estate documents. It was huge, and ChatGPT balked before it could get it fully downloaded.

So I tried just a page of a PDF. It uploaded beautifully but told me it found no words.I finally understood that the software needed an image, not a PDF.

A two-page densely handwritten PDF has been snipped and pasted here as a PNG image.

I snipped an image of a two-page spread as a PNG file and uploaded that. In a few seconds, it came back with this text:

[Left Page]

[At the top, it’s unclear, seems like a title or heading]

… this day came John [Surname unclear] administrator of [Title/Name unclear] who made oath that said [Title] is indebted and swear… [The rest of the text is too faded and cramped together to be legible without guessing]

… time presumed to celebrate [possibly “their”?] marriage but a short time heretofore [unclear] … the clerk [to issue?] [unclear] … they were married in … by … [too faded] … of the peace in the presence of [Names unclear]

[The middle and bottom sections of the left page are too faded and compressed, making it difficult to distinguish individual letters and words.]

[Right Page]

… this day came William [Surname unclear] administrator of [Name unclear] of [unclear] County to [unclear] and [The text is too faded and compressed here]

… [Title] to said decedents estate which is [unclear] and appraised … [much of the remainder of this section is illegible]

… [The last few lines on the page are too compressed and faint to read accurately]

Not too terribly impressive, I realize, but it was definitely indicating that we’re on our way to something.

I Taught AI

I wondered if it would do better with a smaller, clearer piece of text, so I snipped the top portion the left page.  To sharpen the faded text, I pulled the PNG file into free GIMP software and used the Filter–Enhance–Sharpen feature to make the words sharp and clear.

Stephen Little had emphasized several times that we should simply tell AI what we want it to do, addressing it like another human. Let it learn from you, he said. So I decided to add a few instructions before I uploaded the refined image. With the new PNG file, I typed these words: 

Please convert this handwriting to text, wherever you can recognize words. If you don’t recognize a word, put the words “[illegible]” in place of it. Put a carriage return at the end of each line.

Here’s the image I uploaded:

The upper left page of the previous image has been enhanced to have much more contrast between the words and the white background.

It cheerfully responded with these words: 

Sure, I’ll transcribe the visible text from the image with a carriage return at the end of each line, inserting “[illegible]” where the words are not clear:

And it did what it promised, sending me back a much improved interpretation of the handwritten text. 

The State of Alabama Probate Court Special Term

Bibb County 3 December This 29th A.D. 1857

This day came Jonas [illegible] inventor of the last will

and testament of Jonathan Bailey deceased and filed

his ape and schedules for allowance and final settlement of

said estate which is [illegible] examined audited and

filed for allowance and final settlement on the

third Monday in February next and ordered by

the court that notice be given by publication one

a week for three weeks in [illegible] for the Independent

Monitor a news paper printed in the City of [illegible]

notifying all persons interested to appear at the Office

of the judge of probate in Centreville on said day and contest

said ape if they think proper

Of most of the words it marked “illegible,” I had trouble interpreting, too. And it missed a few things, but still left me with something I could clean up faster than I could have transcribed it on my own. It also interpreted some words perfectly that I did not. And it knew the old-fashioned word “carriage return,” making sure each line of the handwriting contained the same words as each line of text—making it much easier to match one against the other.

Now that I know what to send ChatGPT Plus, I can cut many hours off the process of transcribing documents in the future. I suspect I can even teach it to recognize some of the text it missed.

And then there’s the bonus.

For fun, I asked if it would write a blog post for me about today’s adventure, and it did! It wasn’t my style by any means, and I don’t intend to let a machine do my writing for me anyway. But it wrote usable prose—pretty remarkable—and I liked the title it chose, which I kept with minor tweaks.

Then this sweet little piece of AI technology said:

For the visual side of this story, I can create an illustrative image to accompany your post. Let’s envision a scene with antique documents, a magnifying glass, and perhaps a hint of digital elements to represent the AI component. How does that sound?

I, of course, said, “Please do!” It volunteered!

And it cranked out the image you see featured with this story. 

Then, being on a technological roll, I went to Google Docs and used its voice typing feature to dictate this blog post, so I had very little typing to do.

Wrapping Up

I recognize that there are many things we need to work through when it comes to AI. Certainly there are issues of ethics. and none of us wants to become dumb as we let a smart machine do our work.

But what I experienced today tells me that we can get a lot of tasks out of the way that a machine can do better than we can. And we can pour our energy into doing what it takes a real soul to do. We can love recreating our ancestors’ lives. We can bring their long-buried experiences back into consciousness.

Call to Action

I encourage you to become AI aware. There are some people in my life, 20-plus years older, who refused in their 50s and 60s to believe that the internet was going to change the world and refused to learn computers. Then came smartphones—also a NO. They are now handicapped because they waited too long to learn. By the time they realize the world demands email and smartphones for basic life, their brains were not cooperating anymore. It was too hard to learn.

I tell you with all the passion of my being, AI is about to change everything. Some things good. Some things bad. You will want to be able to tell the good from the bad and milk this technology for all it can give you. It gets better every day. One day soon, I will be able to upload a large PDF of old handwritten documents and, in seconds, have a beautifully transcribed text document. It’s improving that fast. Our old age will demand its use, and we need to be ready.

Share...
Share

10 thoughts on “Unveiling the Past: A Genohistorical Journey into AI Handwriting Transcription”

  1. Linda Morgan Clark

    Welcome to this new AI adventure!

    I’ve followed a similar trajectory into AI, and I wake up most mornings now with ideas for what I want to use it on next!

    I’m currently using Perplexity (which one expert calls “extractive AI” to develop research guides on specific geographic areas where I research. I was blown away yesterday when it gave me a comprehensive resource list of places to search for the records of Cornish miners who were indentured to work in the silver mines in Mexico in the 19th century (my gr-grandfather’s story). Google would have NEVER turned up what Perplexity gave me, simply because it would not have “understood” what I was looking for even if it was not on the internet.

    Then to sweeten the task, all I had to do was use the Zotero connector on my browser to create an item for the results of my prompts. Those results contain REAL links, recommended reading, places to visit, etc. It summarizes its findings this way: “This resource list combines academic, cultural, and local perspectives to provide a well-rounded understanding of the 19th-century indentures of Welsh miners in Mexico. Each source will contribute uniquely to your understanding of the economic, social, and cultural dimensions of this historical migration.”

    With what you’ve described of your experience I’m now going to que up ChatGPT4 again and see what kind of an image it will create for me on this topic!

    Can you spell E.X.C.I.T.E.D. ?

  2. If I can’t spell it, Linda, guess what can–and give me its ancient etymology, no doubt? Tanks so much for this wonderful report and for the Zotero plug. I do expect that Zotero and AI will be working together eventually, for maximum benefit. There are a couple of plugins I haven’t explored yet.

    I hear that ChatGPT 5 is expected this summer!

  3. Great post. I really appreciate your forays into finding new tools for genealogists. I have a ton of letters and journals from my family that I thought I would have to read aloud to get them transcribed. That project always ends up at the end of the list because it is so hard. Light at the end of the tunnel!

  4. Family Search now has an experimental section that uses AI to transcribe documents that were previously unindexed, primarily deeds and wills. I’ve been able to find a lot of documents pertaining to my biggest brick wall ancestor and the transcriptions, while not perfect, were about 90-95% accurate and very easy to fix.
    I have been avoiding learning about using AI even though in the past I have been an early adopter of most technology. Your post has convinced me I should at least get more educated on it.

    1. So glad to hear it, Karen, and thanks for mentioning the FamilySearch miracle. I have also had some amazing finds. I can’t imagine what will show up as more and more document sets are transcribed!

  5. I would be curious to see if Transkibus could do better with the areas marked illegible – some of which I can read.

    I think we have to also focus on the second part of the AI transcription process, which is checking over every detail of the transcription to make sure its correct. Its easy on such a clear image like this one, but I can imagine it being much more difficult on harder to read documents, where it is more likely that the AI can hallucinate an answer and people are less likely to be able to “check” its work.

    Even here, there are some words that I think are incorrect – ape? And we need to be cognizant that with fast comes errors.

    (and by the way…you can’t copyright an AI generated image in the US…so the copyright notice on there is invalid)

    1. Many thanks, Concetta. I had been given different and wrong information about copyrighting AI images. You are correct–they cannot be copyrighted. I suspect that the AI translations will always get some of the handwriting wrong. It will be our duty to go behind the technology and check it. I think the greatest value is going to come in helping us to find documents and mentions in documents that we would have missed otherwise. I am not familiar with Transkibus, but it sounds promising. ChatGPT 5 is due out likely this summer, and I hope to see improvements in the handwriting interpretation and other things. OH, and this morning, I put another set of documents through ChatGPT 4 to see if it could interpret the batch. It couldn’t interpret any of them. They were all tightly written and badly preserved and scanned. As you say–another challenge on the way to WOW. Thanks again!

  6. You are so on the money with our need to be open to learning new things. I completely understand people wanting to be cautious and having concerns about AI but one of the things I wrestle with as a retired family historian is the amount of “work” there is yet to be done and not enough volunteers. I think my generation is pulled in so many directions – care for grandchildren, care for the older generation and volunteer work. And I find those that volunteer, volunteer for more than just one organisation. We have to get smarter about our “work”. Yes to AI helping me with recording minutes of meetings, yes to AI helping me write marketing material. Thank you so much for sharing your ChatGPT instructions. You are so clever to have come up with those. I really appreciate generous souls such as yourself who share your learnings with the rest of us.

    1. Thank you, Alex. I’m loving ChatGPT-4o, since it launched Monday, too. My husband left me in the car for a minute last night, and I started playing with it. I asked it to give me a table that showed the percent of the Calhoun County, Alabama, population in each age category in 1830, 1840, 1850, and 1860. A few seconds later, there it was. I then asked it to tell me how many churches there were of each denomination in Calhoun County every decade. It gave me that. I said, “Please name the Baptist churches in 1830 there.” It got four of the five. I know I could find all this information myself–and that I need to check its work–but it is such a great start on research!

Leave a Comment

Your email address will not be published. Required fields are marked *

Shopping Cart
Share
Share

I stand with the innocents of all nations and pray peace to all.

X
Scroll to Top