Disclaimer: This article was written on November 18, 2025, using the paid version 5.1 of ChatGPT. ChatGPT is constantly evolving and changing, which means that its skills and capabilities may change in the future.
ChatGPT is a great new tool, and you can use it for many things – asking questions on the meaning of German words, genealogy research help, and so much more. But one thing it cannot do well yet? Read the old German handwriting.
The tricky thing is, when you ask it to give you a translation, ChatGPT makes you THINK it can read the handwriting. It produces a result with so much confidence – and great English language skills, and perhaps even the right names and dates – that of course you would think the English translation it provides is accurate.
But when an expert German translator actually reads the handwriting, and then compares that handwriting to ChatGPT’s translation, 9.99 times out of ten, there are MANY mistakes. I have discussed this with many expert translators in the field, and at this point, we are all on the same page. ChatGPT is not ready to translate the old handwriting. But the general genealogy public doesn’t seem to know this (and for good reason, ChatGPT is cocky!), so I wanted to write an article as a PSA and help you make sure your family tree is accurate.
Let’s take a look at some examples of various genealogical documents put into ChatGPT, along with the prompts “Transcribe this document into German”, followed by “Translate to English”.
Example 1: 1864 Baptismal Church Record of Gustav Adolf Müller, Godparent Column
ChatGPT Prompt:
Let’s Analyze What It Said…
Ok! So let’s say you didn’t read the old script. The child’s name is Gustav Adolph Müller, so that would make sense that he would be named after a godparent, Gustav Adolph Baersch. The child’s last name is Müller, so having a Selma Müller as a related godparent would also make sense. And then the pastor’s name as the third godparent. Seems like an accurate translation, right?
Unfortunately not…there are a number of mistakes in this VERY short transcription.
Let’s look at ChatGPT’s German transcription and English translation, compared to my human transcription and translation on the right. The bolded words show the discrepancies.
ChatGPT German Transcription:
Gust. Adolph
Baersch, wohl
dann Georg Baersch
Selma Marie Müller,
Pfarr. N. Rinds
Actual German Transcription:
Gust. Adolph
Baersch, vertreten
durch Georg Baersch
Selma Marie Müller, Schwest. des Kindes
ChatGPT English Translation:
Gustav Adolph
Baersch, presumably
then Georg Baersch
Selma Marie Müller,
Pastor N. Rinds
Actual English Translation:
Gustav Adolph
Baersch, represented
by Georg Baersch
Selma Marie Müller,
sister of the child
As you can see, if you took ChatGPT’s translation of this column at face value, you would be misinformed and/or missing out on information. First and foremost, ChatGPT said Gustav Adolph was “presumably then” Georg Baersch – likely leading you to believe he went through a name change, rather than being represented at the baptism by a completely different person. This could lead you down a rabbit hole that wouldn’t even exist, wondering why Gustav used to be named Georg!
Secondly, if you actually translated the handwriting yourself, you would learn a great fact for your family tree – that Selma was the sister of your ancestor! If you had just believed ChatGPT’s translation, you would miss out on learning this fact, as well as have an incorrect name on your record – a third godparent – Pastor N. Rinds – a third godparent that doesn’t even exist as a person.
And that’s just one column on a twelve-column church record – imagine the other mistakes that would crop up if you had ChatGPT transcribe and translate the entire entry. Yikes!
Example 2: 1833 Marriage Record, Paragraph Form
ChatGPT Prompt:
Let’s Analyze What It Said…
Again, let’s say you don’t read the old German handwriting. You had thought this was a marriage record, but you guess you were wrong…ChatGPT says it’s a death record, so it must be.
And it has a lot of great notes for you, such as that “Aeffenheine” is likely a dialect word and “apbellirten” is likely “apoplektischen” (stroke-related). It also gives you some great genealogical context, adding that smallpox was extremely common in the area during this period, insinuating that it makes sense that your ancestor died from the disease. ChatGPT also explains the oddly-formulated 21 and 5/7 weeks into the 28-week illness, explaining that this was a normal formula for pastors to use during this time – so all seems to make sense, and you record the death in your family tree. Would you have been right in doing so?
Let’s take a look at the ChatGPT output and my own translation side by side.
ChatGPT German Transcription:
Gestorben:
Den 15ten September, Nachmittags 3 Uhr, ist das zu
Groß Siemen wohnhaft gewesene Arbeitsmanns-Ehepaar
Johann Vick und dessen Ehefrau Marie, geb. Müller,
einen 21, 5/7 der 28 Wochen währenden Nachmittags
mit Pocken behafteten Knaben Aeffenheine und daselbst
daselbst daselbst daselbst begraben worden: mit derselben Krankheit
des Knaben haben die Kränklichen Geschwister: mit derselben
Krankheit Zwillings-Geschwister, Hinderknecht Hinrich Gekeler
und der Häusler Friedrich Dettmann, daselbst,
und das auf erkannter apbellirten (?) Krankheit daselbst.
Er hieß auch laut Zeugniß der
sämmtlichen hiesigen evang.-luth. Gemeinde-
Glieder und der selben bekannten glaubwürdigen Personen
Zeugen des Kindes waren: Hauswirth Jochen
Behrens daselbst, Hauswirth Georg Stüven
zu Groß Siemen, Hauswirth Jochen Hinrich Gekeler
zu Klein Siemen.
Aufgenommen den 18ten Xber 1833.
Brand.
Actual German Transcription:
den sechszehnten Dezember tausend achthundert dreÿ und dreÿsig wurden morgens neun Uhr nach, unter Nummer 21,551 den 28ten 9br erhaltener Staatserlaubtniss, geschehenem gesetzlichen Aufgebothen und ertheilter Dispenz über das kirchliche Hinderniss, mit wechselseitiger Einwilligung, öffentlich getraut Xaver Ehrhart des verstorbenen Schullehrers von Erlach Xaver Ehrhart und der noch lebenden Appollonia Schmid ehelicher Sohn; und mit ihm die auch ledige Maria Anna Weber, des hiessigen Burgers und Schneidermeisters Wendelin Weber und der Helena Brunner eheliche Tochter. Zeugen der Trauung waren: Wendelin Weber von hier Anton Lang Schullehrer von Erlach.
Renchen, am 15ten Xber 1833
ChatGPT English Translation:
Death:
On the 15th of September at 3 o’clock in the afternoon, the child of the laborer’s married couple
Johann Vick and his wife Marie, née Müller,
living in Groß Siemen, died. He was
21 and 5/7 weeks into his 28-week illness,
and was afflicted with smallpox. He was buried
there (in Groß Siemen). The weakened siblings suffered from
the same disease; likewise the twins,
the farmhand Hinrich Gekeler and the cottager Friedrich Dettmann,
both living there, were affected by the same recognized illness.
According to the testimony of all members of the local
Evangelical-Lutheran congregation, and of other known
trustworthy persons, the child was named Johann Heinrich Friedrich.
The baptismal sponsors were:
farmer Jochen Behrens of the same place,
farmer Georg Stüven of Groß Siemen,
and farmer Jochen Hinrich Gekeler of Klein Siemen.
Recorded on the 18th of December (Xber) 1833.
Brand.
Actual English Translation:
On the sixteenth of December one thousand eight hundred thirty-three, at nine o’clock in the morning, after obtaining the state permission Nr. 21,551 on November 28, the occurrence of the legal marriage banns, and the dispensation granted regarding the ecclesiastical impediment to the marriage, with consent from both sides, the following were publicly married: Xaver Ehrhart, legitimate son of the late schoolteacher Xaver Ehrhart, from Erlach, and Appollonia Schmid, still living; and Maria Anna Weber, also unmarried, legitimate daughter of the local burgher and master tailor Wendelin Weber and Helena Brunner. Witnesses to the marriage were: Wendelin Weber from here and Anton Lang, schoolteacher from Erlach.
Renchen, December 15,[1] 1833
Schleer
[1] This says 15, while above says 16. Likely a mistake by the scribe.
What’s Going on Here?
What?!?! ChatGPT just totally made the transcription and translation up! Almost NOTHING is the same – I can’t even bold the discrepancies, as there are too many. I had to actually double-check myself that I uploaded the correct record into the tool – I honestly thought maybe I did put a different death record into ChatGPT by mistake, as I have never seen it hallucinate so completely. But no, I had uploaded the 1833 marriage record of Xaver Ehrhart and Maria Anna Weber in there, and this invented death record is what it gave me.
Did you catch the word Appollonia, the groom’s mother’s name? That’s likely where it got the apbellirten – apoplektischen (stroke-related) word mentioned above. A little different meaning, don’t you think?
So while the church record column above had some minor mistakes, this record was just completely wrong. And if you believed it to be true, you would have had a child dying of smallpox on your family tree – where it should have been the wedding of your ancestors Xaver and Maria Anna. Not to mention a different set of parents – Johann Vick and Marie Müller – and the fact that ChatGPT seemed to invent some random twins along the way! Crazy!
Example 3: 1900s Photograph
Let’s finish up with a shorter one – not a long paragraph-form record. This is the back of a photograph that we have had passed down in my own family.
As before, I put the document into ChatGPT and told it to transcribe the text into German and then translate it into English. The bolded words below show the discrepancies.
ChatGPT German Transcription:
Here is the transcription of the handwriting exactly as written:
alt
der Jüngste.
—
— Marta (Kusine)
—
(?)
Kronopolit? den 2 h.
Actual German Transcription:
Ally
der Jüngste
—
Marta (Tochter
—
Traugott der 2te
ChatGPT English Translation:
old
the youngest.
—
— Marta (cousin)
—
(?)
Kronopolit? on the 2nd h.
Actual English Translation:
Ally
The youngest
—
Marta (daughter
—
Traugott, the second (likely 2nd son)
Let’s Analyze What It Said…
Again, if you didn’t read the script and read these notes from ChatGPT underneath – “The first word alt is clear”…”Marta (Kusine) is clear”, you would think that those words were definitely correct. After all, ChatGPT says that they are clear to read, so why would you doubt the tool?
That would lead you to put Marta on your family tree as a cousin, rather than a daughter, and completely miss out on Ally’s name. Not great!
However, I was impressed that ChatGPT did admit that the final line was difficult to read – that is the first time I have seen it do that. It guesses that the final line is a place name, which is incorrect, but at least it admitted that it wasn’t sure. Making progress, ChatGPT!
ChatGPT is also wrong about the “den 2 h.”. If you read this note, perhaps you would think this was the time that the photograph was taken, when in fact it is the name of the second son, Traugott – again leaving you wondering what one of the children’s names is. But at least it didn’t try to trick you with its confidence this time – it said “could be” and not “clearly”.
That being said, ChatGPT only gave you one child’s name out of three children – and an incorrect relationship with it. Again, not great.
Your Objections
Having given my presentation on “Can AI Read the Old German Handwriting?” several times now, I have heard a handful of objections when I try to tell people that ChatGPT cannot read the old German handwriting (yet). I wanted to address them below.
1. “I always tell ChatGPT to transcribe into German first, and then translate to English. This makes the translation accurate.”
As you can see above, this is what I did for all of these examples. ChatGPT does not – in all the cases I have seen – transcribe the handwriting into German with 100% accuracy, thus leaving you with an incorrect English translation when it uses its transcription as a basis. It sounds like correct English, but it is translating from incorrect German – leaving you with inaccurate results for your family tree.
2. “I tell ChatGPT ‘Pretend you are an expert in Kurrentschrift, one of the best transcribers in the field. With that in mind, transcribe the text into German and then translate into English.'”
When someone suggested this in the chat on my webinar, I was intrigued – would that work? I tried it out on a vital record I had been working with, and the transcription and translation were still completely incorrect – oddly, even more incorrect than before, in this instance. But what interested me the most was the notes that ChatGPT provided me with after the translation. It stated:
Notes from a Kurrentschrift expert’s reading:
- “im Gericht gefaßten” (“recorded in court”) appears clearly — confirming the recognition was made judicially.
- “über d. 39” means under record number 39.
- The handwriting’s long s (“ſ”) and Kurrent-style “e” loops can make “gefaßten” and “beurkundeten” look tricky, but the phrasing follows standard legal formulae from early 20th-century civil records.
Well, when I read that, I actually went back and looked at the document. Did it say “im Gericht gefaßten” somewhere? It must, if it said that phrase appears clearly. But no – that phrase was NOWHERE to be found on the document. ChatGPT just made it up and made us think it knew what it was talking about, as a “Kurrentschrift expert”. The other two points it made from an “expert’s” point of view are also completely made up and do not reflect what is written on the document – even though it gives us reasons to explain why it is right.
These “notes from a Kurrentschrift expert’s reading” made me doubt myself, and I am someone who works with Kurrentschrift every day. If you don’t know the handwriting, I wouldn’t blame you for believing these points!
3. “I don’t need everything to be right – I just want the names and the dates.”
That would be great – if you could be certain that the names and dates ChatGPT gives you were accurate! Unfortunately, as you can see in the examples above, it completely makes this biographical information up sometimes as well! And besides, as genealogists, most of us want all the juicy details a record offers, right?
Conclusion
At this point in time, it is still advisable to use a human to translate the old German handwriting. And of course, that could change soon – I’m sure these tools will just keep getting better and better. If you are still set on using an AI/machine translation tool, Trankribus and Google AI Studio do seem to be more accurate than ChatGPT – but they still make a number of mistakes! I would only use them if you can read the handwriting yourself and you’re able to identify where on the document those inaccuracies are located. You can use AI as a starting point, but let yourself, the human, be the final editor.
As mentioned at the beginning of the article, however, ChatGPT can still be a huge help in genealogy. I use it all the time to ask definition questions, of both historical words and abbreviations (and I then verify its answers by googling what it tells me and doing a bit more research on the meaning it provided). It’s also a pretty good translator of both Fraktur (the font in German newspapers) and of typed German text – much better than either Google Translate or DeepL, in my opinion.
You can also use ChatGPT if you are wanting to write an e-mail request to an archive or church in Germany. I always tell people to write in both languages – German and English – when submitting your request, as that gives the European recipient the option to reply to you in their native language or use their English if they like. But we don’t want to just assume they speak English. So you can type your English e-mail request into ChatGPT and then give it the prompt, “Translate this text into German”. You can then copy and paste the German it provides in your e-mail to the archives. I have been very impressed with the translated results and its German grammar skills!
And of course, there are many ways to use ChatGPT for research, which I won’t go into here – that would be another article! But feel free to share any research ideas in the comments that you think might help others.
I hope you enjoyed the article and it helps you!
13 Responses
ChatGPT wasn’t too bad with transcribing and than translating my newspaper articles typed in Fraktur from OldNews. I still needed to verify and edit everything!
Yes, as I said in the conclusion, it does pretty well with Fraktur! I was impressed too! Glad you are verifying though!
Great analysis! Thanks a lot!
I learned to read Kurrent first. Handwriting, abbreviations, and obscure words (both German and Latin) sometimes trip me up. The way I use AI is typically giving it the text I have read with my eyes, as inaccurate as it might be since I am not a fluent German speaker, and asking it what those funky words might be given the context. Often the AI will suggest words that I will then be able to see – either yes or nope, there’s no way that’s a “t”! It offers suggestions, but my eyes are the deciding factor. Transkribus can often give me a jump-start to unfamiliar handwriting, sometimes it’s easier to correct it than to start from scratch. The key is learning and seeing it with your own eyes. The Reddit community is a huge help when I need another set of eyes – I’ve learned so much there!
I agree wholeheartedly that it can’t properly translate whole records, however, it has helped me in several instances when I can make out all but one word in a sentence [at least I can’t find a German word spelled that way] . I type the German sentence into ChatGPT along with the year and type of record and ask it to suggest what the unknown word might be. In many cases it comes back with suggestions on what the word is and has helped me to correctly transcribe the unidentified word.
Hi Carole! Thanks for the idea. The difference here is that you are able to read the old handwriting – so you’re able to check if the transcription ChatGPT suggests is correct or not. Many people can’t read the handwriting and are taking what it says at face-value. If you can read the handwriting and are able to see if it’s right or wrong, I can definitely see using it to help you with an unknown word.
I uploaded a handwritten poem. The author was a school child whose German family settled in my area in the late 1880s. I had to go back and explain certain local terms and landmarks. Some of the translation attempts were downright comical! AI definitely misread readily decipherable words and their significance.
So, there was almost as much work correcting the translation as just doing it myself.
I do appreciate how quickly AI can produce documents for printout. This is a bonus for me, as I have impaired hand dexterity.
I reviewed an older print out from AI.
It said that Pferd= ford, an old term for horse. I cannot verify that translation anywhere!
You can’t verify the translation because this word does not exist. An older German term for Pferd would be “Ross”. Words spelled similar to “ford” are Furt, meaning ford and Förde, meaning bay or fjord. No horses found here!
As with so much of our difficulties in this area, a lot of the problem comes down to the handwriting of the scribe. In the examples you used (which are unfortunately quite typical of so many), the writing appears to be somewhere between abysmal and bad. Even with my limited experience reading Kurrentschrift, I would have an extremely difficult time making out some of the letters. Although, bad as the handwriting is in your first example, I can easily see it doesn’t say “wohl dann”. This is also where context can come into play to help, and if ChatGPT is making things up, instead of trying to use context, that’s a big problem. Think I’ll just wait until those who are actively playing with ChatGPT in this area (more power to them, and also a big thank you to them) surprise us by saying “Eureka! They’ve finally succeeded.”
Do you have any experience with Grok? I’m wondering if it can do any better than Chat gpt. Experience has shown me that any AI can be extremely cocky when it is outright making things up (on subjects I’m good with) so I greatly appreciate your warning with old German writing as I still struggle with that.
I have never studied German so I am at the whim of the various translation tools. Right now I am working on a government document. I enjoy working out the translation using the tools you have suggested Katherine: Google Translate, DeepL, Wordmine, and Thode’s book. Then I send my translation to ChatGPT to translate my translation into English. Does anyone else do it this way and what do you think of ChatGPT translation?
I just spent an enjoyable two hours arguing with ChatGPT over its transcription of a 1736 Acadian church book entry of a marriage. A document I can almost read completely without help. I actually had to show it an alphabet chart from a BYU Script Tutorial website to prove my point.
Is Grok any better? I know it was the only thing that worked when I was locating certain communities. And I thought it was helping me figure out names. Just haven’t asked it to decipher many notes because I have no idea if they are correct or not.