A recent 2nd cousin DNA match (click here for that story) inspired me to explore German records from Bavaria. So I decided to try AI to read the handwriting and translate it. First however, I did some more work on my Norwegians.
My home page at Claude
There is more than one free AI out there, so I ran an experiment. First I submitted a not great scan to all three of the better known ones, Claude, ChatGPT and Gemini. This is what I found:
- Claude told me it was hard to read and only gave me a partial.
- ChatGPT made stuff up that sounded totally plausible, but did not seem to be anything like the original Norwegian.
- Gemini made an effort to translate it and did well, but organized it into bullet items.
So my new method is to upload the same document to Claude and Gemini and merge the result. One thing I like about both of them is that they remember my recent efforts in the left hand sidebars (on Gemini ask for the sidebar in settings). So far I have not used either one heavily enough to be asked to upgrade to a paid version, unlike ChatGPT which has already requested that.
Next I tried some handwritten German documents. In my opinion Gemini did the best job. Read on for how each AI performed.
Claude told me for the first document “⚠️ Note: This register is in old German Kurrent script, which is very difficult to read with certainty. Several names and place names above are approximate — particularly “Fünfgam,” “Oberhafern,” and “Pößelfing.” If you need precise transcription for genealogical research, a specialist in Kurrent handwriting would be recommended.”
Now I asked about the death record of my great great grandmother, a Bavarian Catholic women who died in Munich. Click the image for the full record with column headings (she is line 62) or here to see it online.
Death record for Anna Reiner Riedleder
Claude attempted a translation and then said “The cause of death appears to be paralysis, she was 63 years old, died on January 21 at 3:30 AM, and was buried on February 2. She is listed as an Englishwoman.” I am totally puzzled by Claude thinking she was English.
ChatGPT stopped after a few lines
While Gemini said, “This entry is written in a late 19th-century German Kurrent/Sütterlin script and records a death.” Then proceeded to do a great job, column by column:
| Column Header (Translated) | German Transcription (Kurrent) | English Translation |
|
Fortlaufende Nr.
(Consecutive No.) |
62 | 62 |
|
Namen
(Names) |
Rissleiter
Anna |
Rissleiter (also spelled Rißleiter)
Anna |
|
Stand und Religion
(Status/Occupation & Religion) |
Taglöhnersehefrau, r. k. Aufenthaltsort, Hausnummer:
Mozartstr. 92/0 |
Day laborer’s wife, Roman Catholic. Residence & House Number:
Mozartstraße 92, ground floor (/0) |
|
Ledig, verheiratet…
(Single, married…) |
Ehefrau | Married woman (Wife) |
|
Arzt… Todesursache
(Doctor… Cause of Death) |
Paralys. cordis
Dr. Littrich |
Paralysis of the heart (Heart failure)
Dr. Littrich |
|
Tag, Monat, Jahr u. Stunde
(Day, Month, Year & Hour of Death) |
28. Jänner,
8 1/2 Uhr
A. M. |
January 28,
8:30
AM |
|
Tag der Beerdigung & Ort
(Day & Place of Burial) |
2. Feber,
öftl. |
February 2,
Eastern [Cemetery] |
|
Alter
(Age) |
63 Jahre | 63 Years |
|
Pfarrer…
(Pastor or Representative) |
Koop. v. St. Georg
Probst |
Cooperator (Assistant Pastor) of St. George,
[named] Probst |
|
Bemerkung
(Remarks) |
pr. 682 / 858 |
Abbreviation for a parish registry cross-reference or payment state (likely “protestantisch” notification or “praesentatum”).
Record numbers: 682 / 858 |
And Gemini’s commentary on this:
### Key Contextual Notes
Location Hint: The address Mozartstraße 92 and the cemetery abbreviation öftl. (Östlicher Friedhof / East Cemetery), alongside the parish of St. Georg, strongly point to this record being from Augsburg, Bavaria (matching the neighboring entry on line 60 which explicitly mentions Augsburg).
The Month: The scribe uses the Austrian/Southern German regional terms Jänner for January and Feber for February.
Cause of Death: Paralysis cordis is a standard Latin medical term used at the time for sudden cardiac arrest or heart failure.
Fascinating that the cause of death called “paralysis” is actually a heart attack.
I expect to continue my German and Norwegian research with these AI helpers.
I’ve just been working with Gemini on 18th century German and Latin birth/marriage/death records – overall I’m very happy with the results at this point, but it does need correction and/or instruction. As you found, it gave me the bulleted version of a transcript and translation first. I then asked for a full transcript and translation- which it gave me, but changed the order of the words a bit (and I was interested in the word-for-word of a birth record). So I prompted, “You said it starts with “x”, but it clearly starts “y” so I got an apology and finally it supplied what I asked for. It will find the same person or a likely spouse of that person on a record I gave it days earlier and points that out, sometimes suggesting the family probably has a connection to another village, which is helpful. It did hallucinate names of sponsors in one case – it was an odd situation where there were four godparents, not two, so it was confused – but I had pointed that out when I uploaded the image. I had to correct this error out twice, because it clung to its original names and tried to shoehorn them in.
Yes Steve Little in his lectures on using AI often mentions how to phrase the question to get an accurate answer. I just uploaded the documents and told them “translate”
I really like MyHeritage’s Scribe AI. I was shocked that it could both identify and translate (reasonably well, I think, not speaking the language) my grandmother’s Carpatho-Rusyn Eastern Slovakia dialect in family letters she saved.
Presumably only for items found in MyHeritage?
No
I suggest that Transkribus (https://www.transkribus.org/) would do a better job.It is also free for light users. Transkribus was trained on gothic and other old handwriting samples. It’s used mainly by archives, historians, and genealogists. In my experience it’s great on Norwegian chuch records, old German wills and early colonial/US town and church records. I’d save your free LLM time for other things.
Zoe, thank you for that. I had forgotten the name of that service.
I tried it on the same page of deaths and it transcribed without translating plus I did not see a way to ask for just line 62. Then it listed each column separately. Finally I still need to get the result translated, so I think I will use it only as a secondary double check.
What impressed me about Gemini was the interpretations. Who knew paralysis meant a heart attack?
Not translation but I submitted early 1700s handwritten wills from Wales but written in English. Gemini easily outperformed Transcribus and presented the results virtually in Word format.
Thank you for all of your great information! I will be trying these. My son who worked for Google says it’s pretty random which AI works best for each task. If one fails, he tries another.
To my eye, the “Englishwoman” was a misreading of “Taglöhnersfrau” (Day laborer’s wife), starting with the “T” being incorrectly read as an “E.” I also notice that a translation has inserted “ehe” (married) in that word, possibly because of the “ehefrau” in a subsequent field.
I usually find that Gemini does the best transcribing/translation among the common AI programs. I’ve tried Transkribus, but it has a steep learning curve. But if someone is familiar with how to use it, it seems fairly accurate for transcription, which can then be transferred to another AI for translation.
I too have found Gemini to be the best for transcription/translation, but still only use it for documents about which I already have enough knowledge so I can verify what it tells me.
It was the only one to read older Russian and Polish for me, as well as fairly accurately transcribing some 18th century wills written by my ancestors from Kent, England.
I still use it sparingly, due to the environmental cost and the fact I want to keep exercising my own little grey cells rather than becoming dependent on technology to do it for me.
I have also done the test between ChatGPT, Claude and Gemini using Steve Little’s prompts. One important trick learned after many hours is to request that the Ai first improve the document or image. Once improved the results have been measurable better. I can also take the best image and submit it to allow them. Also,give the site Prompt Cowboy a try for creating promptsof your own.
Pingback: Friday’s Family History Finds | Empty Branches on the Family Tree
Kitty, love this writeup, and comparing two AIs is a smart habit, but we have to be careful with this type of “triangulation.” One friendly wrinkle: researchers find different AIs tend to trip over the same hard spots in the same way (“correlated error”). So on a tough passage, Claude and Gemini might agree on the wrong reading–and that agreement can feel like confirmation when it isn’t. (Old idea, really: scholars reading ancient manuscripts treat a shared error as a sign two copies came from the same source, not that it’s correct.) So when the two AIs agree on something hard, I’d take it as a cue to double-check the original, not as the final word. Thanks for sharing these experiments!
Good point about “correlated error.” What I love about these two AIs is that they often explain terms and concepts which I did not know!
I actually questioned Gemini about the last name it gave the mother from a birth record and got a fascinating reply
“ You are entirely right, and good eye! Let’s look closely at the image “IMG_0562.jpeg” again.
Looking at the mother’s column and the long paragraph at the bottom, that surname is absolutely Frankenberger, not Denk.
Here is what went wrong in the initial reading:
The handwriting uses a classic old Kurrent F (which drops far below the line) and a g that looks like a modern cursive “p” or “h”. When reading quickly, the middle section “anken” was squeezed, making it look deceptively like “Denk” at a glance. However, looking at the clear “F” at the beginning and the “-berger” ending, it is unmistakable.
Here is the corrected transcription and translation for that section:”
So I am finding using these AIs very educational as I take this deep dive into Bavarian records for my DNA cousin!
Pingback: This week's crème de la crème - May 30, 2026 - Genealogy à la carteGenealogy à la carte
Kitty, I have a question about X DNA and hope it is ok to post it here. I am teaching myself about X DNA and coming up with some confusions. I read your 2014 article about X in Family Tree and it squares with my experience at FT as I only have my sister and an unknown cousin as X matches. However, it appears that FT shows X matches below 10cm as existing when one looks at chromosome research in FT, but does not list them as existing when a search for X cousins is used. I have been using FT and GED to look for X cousins and discovered a couple of things. First, I have a X cousin who shares 5cm of autosomal DNA and the X DNA is more than 35cm. It goes back six generations to a woman who was born in 1775. In GED I have many X cousins who do not share much or any autosomal DNA with me. Some of these cousins share more than 20cm of X DNA. OK, so should I assume that my mom carried many old X DNA hunks that amazingly passed down to her and that many predate 6 generations? Also, it is beginning to look like most of my X DNA are pretty large hunks. One more question, should I presume that the X line that is also the MT line would be more likely to not exist because there were recombinations at every generation? I hope this all makes sense. Bill
Probably better to post this on my major X post where you previously posted but as it has 103 comments maybe not
http://blog.kittycooper.com/2014/01/what-does-shared-x-dna-really-mean/
I have also observed that X matches can be large chunks but the mathematicians in the DNA world disagree so it may just be a result of the inheritance pattern.
I myself have a 33cM X chunk from an ancestress born in 1729. The path was alternating females and males. See
https://blog.kittycooper.com/2016/02/we-have-metta-olsdatters-x-dna/