21 Comments
User's avatar
Steve Minniear's avatar

Thanks for this post. I've used AI once to transcribe a 1939 handwritten letter, just to see how it worked. And it worked pretty well. Well enough that I am considering going through the process with all thirty something letters. It will provide me with a benchmark to compare and contrast my efforts at transcription over ten years ago. AI should be seen as a helpful tool, but not something to rely upon solely.

Expand full comment
Kirsi Dahl's avatar

100% agree. I will say that I’ve seen SIGNIFICANT improvements from month to month in how it reads and transcribes. It’s learning so fast. But, yes, always verify!

Expand full comment
Anne’s Family History's avatar

I tried transcibing with ChatGPT and found it sometimes hallucinated even though asked to transcribe word for word. I have also used Transkribus and GoogleDocs to transcribe handwriting - neither of these tools hallucinated.

Expand full comment
Paul Chiddicks's avatar

Thanks for sharing with us your valuable insights about what you learnt by experimenting with different AI platforms to help translate historical documents. Very revealing in your summary about their own self promotion.

Expand full comment
Jane Chapman's avatar

Thank you @Kirsi Dahl It is really useful to see your comparison of AI tools for the purpose of transcribing. I have recently been using Transcribe History - https://transcribehistory.com/ to transcribe some handwritten documents from Archives New Zealand. It was working really well, almost perfectly, until some time during last week when it started throwing up error messages and transcribing about 1 in 5 documents. So I am looking for an alternative. I tried Transkribus with the same documents and it was well off.

Expand full comment
Kirsi Dahl's avatar

Thanks Jane. I hope this post gave you some good options for alternatives! Every one that I tested I did on the free versions, which is a bonus!

Expand full comment
Jude Rhodes's avatar

Thanks for an excellent summary, really helpful. AI certainly has it's place in research, I use it as a starting point to then use traditional methods.

Have you tried transkribus? www.transkribus.org

Expand full comment
Kirsi Dahl's avatar

Thanks Jude. I have not tried Transkribus, but it seems many have and all have had good experiences. I will be checking it out. :)

Expand full comment
Emma - Journeys into Genealogy's avatar

Thanks for sharing this Kirsi, it makes for very interesting reading. I was unaware that there were 5 different AI platforms. I've only tried ChatGPT so far so will look into the others.

Expand full comment
David Shaw's avatar

The outperformance of Co-Pilot I think is related to it's generally better capacity for "seeing" images. I detest Co-Pilot for an entire range of defects but grudgingly admit it does well handling certain types of images. As far as the self promotion and "I'm the best," I am not surprised since A.I. is "trained" by its creator.

Expand full comment
Kirsi Dahl's avatar

I am surprised by how much I was surprised. 😆

Expand full comment
Nancy G. Carver's avatar

Very informative. So they all think highly of themselves. I see people using ai for transcription a lot over in the genealogy and ai forum on Facebook. They also use a tool called Transkribus which, it appears, one must train in order to have best results. I have found most attempts to be tedious, because it is necessary to double check very carefully, and unlike ocr (which will usually print weird characters when it can’t decipher) the ai models often just blithely make things up, which is harder to spot. It does do fairly well with my dad’s journal (good penmanship) and other very legible things, but so would older tools.

Expand full comment
Kirsi Dahl's avatar

Yes, I have had complete fabrications from AI in the past. I have seen improvements over the last year but to your point, always checking the work is a key part of any process involving AI.

Expand full comment
David Shaw's avatar

I'm sure it will be used at some point to forge documents.

Expand full comment
Kirsi Dahl's avatar

Sadly, it seems there are always those that find a way to sully things.

Expand full comment
Jill Swenson's avatar

What an interesting experiment and the results are fascinating to compare. Until things improve with AI, I'm inclined to believe it remains an illustion to think a machine can do the transcription for me. I am curious whether you found any of these methods significantly faster than reading it and transcribing it yourself.

Expand full comment
Kirsi Dahl's avatar

Great question, Jill. If I were doing only one document, it would possibly be easier and maybe faster to do it myself. Many times I’m doing a series of related documents (such as a series of handwritten letters between sisters) and I find it to be much more efficient to leverage the capabilities of AI.

Expand full comment
Jill Swenson's avatar

Thanks for your honest and candid reply. Like so many tasks, sometimes it is easier and faster to do it yourself than to try and find a work-around. Although avoidance of tedium may have been one of the instigators of many inventions.

Expand full comment
Barbara at Projectkin's avatar

Wow, Kirsi! This is brilliant. You've done the work so <shuffles papers> 🙂‍↔️ I don't have to. I love this. As soon as I get a chance, I'm digging in on one of my war records. I'm thrilled!

Expand full comment
Kirsi Dahl's avatar

Ooh, good luck! I've been absolutely shocked by how fast AI improves. One month it won't even try to transcribe a handwritten document, and the next it's transcribing that same document nearly flawlessly. The motto "if you don't succeed, try, try again" was made for this type of thing.

Expand full comment
Barbara at Projectkin's avatar

Exactly! It directly parallels the advances in image search and (frighteningly) facial recognition as well. Great stuff. Thank you.

Expand full comment