












































Always Room for Mindfulness


179

AI Meets Astrology
Creating a Digital Edition of Anton Brelochs’s 
1529 Practica with Transkribus and ChatGPT
Nathan R. Ericson, Library Director, Wisconsin Lutheran Seminary

ABS TR AC T: This paper explores the integration of AI tools like 
Transkribus and ChatGPT in creating a digital edition of Anton Brelochs’s 
1529 Practica Teutsch. The research uses the rare intact copy at the 
Wisconsin Lutheran Seminary Library to assess the efficacy of AI in 
transcription, translation, and tagging processes. Key questions include 
AI’s role in identifying and correcting transcription errors, speeding 
up translations to enhance accessibility, and automating TEI XML 
tagging. The study highlights both the strengths and limitations of 
these AI tools, concluding that while they offer valuable assistance, 
human oversight remains crucial for accuracy and scholarly reliability. 
Additionally, the paper examines AI’s potential in interpreting histori-
cal illustrations, such as the astrological woodcut in Brelochs’s work, 
demonstrating its capabilities in suggesting plausible identifications 
of depicted figures.

INTRODUCTION

The impetus for this paper was the discovery of a unique volume in 
the rare books collection of the Wisconsin Lutheran Seminary (WLS) 
library: the only known intact, extant copy of the Practica Teutsch for 
1529 by Anton Brelochs, who served as the city doctor of Schwäbisch 
Hall from 1517 to 1559 (Institut Deutsche Presseforschung 2024). 
(Another fragmentary copy is in the collection of the University of 
Kiel and was recently restored with the help of images from the WLS 
library copy.) At the WLS library we are planning projects that would 
create digital editions of the systematic theology works of Abraham 
Calov and Johannes Andreas Quenstedt. The discovery of the Practica 
by Brelochs presented the opportunity for a pilot project to create 
a digital edition of a much shorter work. Our hope was that in the 
process we would be able to explore the available tools, methods, 
and platforms to better inform our digital edition project plans.



180    ATL A 2024 PROCEEDINGS

Digital editions bring together facsimiles (images), transcriptions, 
and often translations of one or more manifestations of a work—usu-
ally of non-digital materials such as manuscripts, letters, and printed 
books—in an online interface that is suitable for scholarship. (For 
an example, see the Taylor Editions platform from the University 
of Oxford at https://editions.mml.ox.ac.uk/.) Digital editions are com-
monly expressed in TEI XML format, which is a text file that abides by 
the XML specification and makes use of the Text Encoding Initiative 
(TEI) standard for the representation of texts in digital form. The 
goal of the TEI standard is to produce files for digital editions that 
are both human readable as text files (and thereby suitable for 
long-term preservation) and machine readable through markup 
language (and therefore suitable for encoding meaning in a way 
that can be processed by computers). One strength of digital edition 
platforms is that multiple versions of facsimiles, transcriptions, and 
translations can be provided by creators and individually selected 
for comparison by users.

The recent advent of AI-based tools such as ChatGPT led us to ask 
how AI-based tools might assist in the creation of a digital edition, 
whether that would be through lightening the workload, improving 
the process, or advancing what is possible in creating a digital edi-
tion. Specifically, this paper explores how AI-based tools can help 
with the transcription, translation, and tagging that is necessary in 
the creation of a digital edition: Can AI identify and correct errors 
introduced in the transcription process? Can AI speed up the process 
of making digital editions more accessible through translation? Can 
AI automate the process of adding TEI XML tags to a transcription 
and accurately linking those tags to named entity authorities? As 
a bonus, in the case of the 1529 Practica, this paper also explores 
the usefulness of ChatGPT in interpreting the astrological woodcut 
featured on the pamphlet’s opening page.

THE PRACTICA TEUTSCH FOR 1529 BY ANTON BRELOCHS

The WLS library’s rare book collection contains about 4000 items 
across 2200 titles. Of these titles, two are from the 1400s, 400 are 
from the 1500s, 600 are from the 1600s, and 1200 are from the 1700s 
or later. Beginning in 2001, we sought to at least begin to answer 
the question of which of our rare books were truly rare and which 

https://editions.mml.ox.ac.uk/


Listen and Learn Sessions    181

were old but perhaps not so rare. A study of the 30 oldest books in 
the collection discovered several that had not yet been digitized, sev-
eral for which there were less than ten known copies, and one—the 
1529 Practica Teutsch by Anton Brelochs—for which the only other 
catalog record found, at the University of Kiel, noted that the bot-
tom half of all pages was missing. A subsequent email conversation 
indicated that the catalog description was a bit generous regarding 
the current condition of the item. The WLS library copy is intact, 
though with slightly wrinkled pages and other wear; photographs 
of the item can be viewed in the Internet Archive at https://archive.
org/details/brelochs-practica-1529/. 

The production of German, Italian, and Dutch practicas, annual 
booklets of astrological calendar-based prognostication not dis-
similar to the modern Old Farmer’s Almanac, began soon after the 
invention of the printing press and flourished throughout the six-
teenth century. Their literary design is that of the overhearing of a 
“privileged conversation” (Green 2012, 82) between an astrologer 
and the rulers of a territory, where the buyer gains a glimpse of this 
otherwise-obscured knowledge. Most practicas follow a prototypi-
cal form, including elements such as the name practica, a title page 
containing a woodcut with representations of featured planets for 
the year, an identification of the author, a dedicatory epistle with 
biblical justification for astrological prognostication, and finally sev-
eral chapters containing predictions for various spheres of earthly 
life for the coming year (Green 2012, 111–15).

The WLS library copy consists of 16 pages in two separately 
stitched quartos, four leaves each, held together by a tape binding. 
Pages are 15 cm wide by 20 cm tall with horizontal chain lines. 
The work and its instances at Kiel and Mequon (where Wisconsin 
Lutheran Seminary is located) are now listed in VD16 (Bibliography 
of Books Printed in the German Speaking Countries of the Sixteenth 
Century); in addition to the Internet Archive, a digitization of our 
copy is housed in the WLS institutional archive and digital library.

AI AND TRANSCRIPTION

Can AI-based tools help with the transcription of written or printed 
documents? One type of tool is already in regular use for transcription. 

https://archive.org/details/brelochs-practica-1529/
https://archive.org/details/brelochs-practica-1529/


182    AT L A 2024 PROCEEDINGS

Handwritten Text Recognition applications such as Transkribus 
(https://www.transkribus.org/) and OCR4all (https://www.ocr4all.
org/) use AI technology to determine likely bounding boxes and 
lines of text within both printed and handwritten documents and 
allow users to train models for the recognition of characters and 
words. Transkribus is hosted online and operates on a subscription 
model; OCR4all is a locally installable Docker image. (Potential users 
should note that as of this writing, OCR4all has not been ported to 
Apple silicon.) 

I used Transkribus to perform initial transcription of the Practica 
using the Transkribus Print M1 model that recognizes, among other 
typefaces, the Schwabacher typeface used in the Practica. I sampled 
both color and grayscale images of the pages. Although the color 
images were easier for my human eye to read when checking the 
transcription, it seemed that Transkribus produced equal results 
with either type of image. The sixteen pages of the Practica (the 
first and last of which contained substantially less text than the 
others) took a total of 10 minutes, 10 seconds for Transkribus to 
process. I then proceeded to correct Transkribus’s transcription in 
the interface that Transkribus provides for this purpose, comparing 
the transcription with the original document and its images, but 
only after saving the raw transcription to a text file to experiment 
on it using another AI-based tool, ChatGPT. The edited transcription 
is currently viewable on Transkribus at https://app.transkribus.org/
sites/brelochs1529/doc/2694510.

My first attempt to see how well ChatGPT could identify and fix 
transcription errors involved creating my own custom GPT with 
ChatGPT 4o, to which I provided instructions that it should take any 
text it received, identify and correct errors in the raw transcription, 
and highlight the changes in the response it produced. The interactive 
custom GPT design process produced the following set of instructions:

This GPT assists users in detecting possible errors in OCR transcriptions 
of historical German texts, specifically from a 1529 German pamphlet or 
practica teutsch. The GPT identifies and corrects transcription errors, high-
lighting the changes made in bold type while ignoring any text enclosed 
within angle brackets. It does not make orthographical corrections unless 
they are necessary to resolve a transcription error.

Role and Goal: The primary goal is to ensure accuracy in OCR transcriptions 
by identifying and correcting errors in the text. The role involves carefully 

https://www.transkribus.org/
https://www.ocr4all.org/
https://www.ocr4all.org/
https://app.transkribus.org/sites/brelochs1529/doc/2694510
https://app.transkribus.org/sites/brelochs1529/doc/2694510


Listen and Learn Sessions    183

reviewing the text, making corrections where needed, and highlighting 
these changes for easy identification.

Constraints: Avoid making unnecessary orthographical corrections. Ignore 
any text within angle brackets as these are locators for human readers. 
Focus only on identifying and correcting transcription errors that affect 
the text’s meaning or coherence.

Guidelines: Always highlight corrections in bold type. Review the text for 
inconsistencies, nonsensical parts, or obvious errors resulting from OCR 
transcription. Use historical knowledge of the text to make informed cor-
rections where possible. Include corrections with the main body of the 
response as opposed to afterward.

Clarification: If the text is ambiguous or unclear, make an educated guess 
to correct the transcription errors. If unsure, err on the side of providing 
a coherent reading (rather than maintaining the original transcription 
unless it is clearly incorrect).

Personalization: Maintain a professional and helpful tone, ensuring clar-
ity and precision in responses. Be patient and thorough when reviewing 
the text for errors. Automatically process pasted text according to these 
instructions, even if no explicit prompt is given.

 The results using the custom GPT were not good. It wasn’t all 
bad; the GPT made several good corrections in the section that 
I examined. However, it also made unrealistic and unnecessary 
corrections, missed obvious corrections, and occasionally allowed 
nonsensical words to pass through unfiltered. Overall, it proposed 
very few corrections compared to the number that I identified. Ben 
and Sara Brumfield observed similar results when exploring the 
use of AI-assisted transcription for their crowdsourcing platform, 
FromThePage. They concluded that it was better to ask ChatGPT to 
highlight obvious errors (creatively using the “hmm” or “thinking 
face” emoji to highlight these locations for crowdsourcing partici-
pants) than to ask ChatGPT to supply corrections (Bastida 2024).

For my second attempt, I used a different approach. Recalling 
the late-2023 trend of ChatGPT sometimes refusing to follow explicit 
instructions (often referred to as the AI being “lazy”), I attempted a 
bit of misdirection: asking ChatGPT to perform a secondary task that 
would, in the process, accomplish the primary task I had in mind, 
that of transcription correction. Instead of using a custom GPT with 
complex instructions, I simply directed ChatGPT 4o, “The following 
is 1529 German. Please convert to modern orthography,” and then 
provided the raw text under consideration.



18 4    ATL A 2024 PROCEEDINGS

The results now were much better. ChatGPT’s fixes generally 
agreed with my fixes, and what ChatGPT produced generally made 
sense where the original lacked sense—and did so accurately. There 
were exceptions, of course, in which ChatGPT mistook a word and 
altered the overall meaning. Context may determine whether this is 
a critical problem. On the one hand, it may not detract from a casual 
reader’s ability to get an overview of the content of the work. On the 
other hand, it will not serve as an accurate or dependable basis for 
scholarly research. In addition, an obvious drawback to this approach 
is that once ChatGPT provided modern orthography for the 1529 
German, the transcription was no longer a diplomatic transcrip-
tion that accurately represented what was on the page. Therefore, I 
conclude that the best method for transcription improvement is not 
the use of ChatGPT, but rather the training of the recognition model 
used in Transkribus (or other HTR applications) so that it is more 
attuned to the nuances of the document being transcribed. This will 
be our library’s approach going forward as we seek to transcribe 
longer works, namely training a custom model on the first n pages 
of the work so that the remaining pages may be transcribed with 
greater accuracy.

Links to the transcriptions produced may be found on the project 
website at https://nericson.github.io/brelochs/. 

AI AND TRANSLATION

Given that the 1529 Practica is written in German and mein Deutsch 
ist nicht so gut, I next experimented with asking ChatGPT 4o to 
translate both the raw transcription and my edited transcription 
into English. The difference between the two results was difficult 
to measure because of ChatGPT’s “temperature” setting, which 
provides some randomness in the responses given when the same 
prompt is given multiple times. (Future experiments could make use 
of the ChatGPT API, where the temperature may be set “lower” by 
the user.) Accounting as best I could for “artistic license” in transla-
tion, the translations of the raw and edited transcriptions appeared 
overall to be similar, with several improvements in the translation 
of the edited transcription and an occasional better reading in the 
translation of the raw transcription.

https://nericson.github.io/brelochs/


Listen and Learn Sessions    185

My conclusion from this experiment is that AI-based tools may 
be very useful for improving accessibility to the content of works 
written in languages not familiar to the user, especially when an 
overview of the content is more important than specific detail. 

As a side point, in my experimentation I also observed that it’s 
important to work against ChatGPT’s tendency to paraphrase. If 
I asked ChatGPT to translate “very idiomatically,” it would often 
substantially shorten the text under consideration, perhaps still 
conveying the overall meaning, but also eliminating much of the 
detail. I found a much more workable translation by asking ChatGPT 
to translate “rather idiomatically”—or even “rather literally” if I 
wanted a more word-for-word representation of the original language.

Links to the translations produced may be found on the project 
website at https://nericson.github.io/brelochs/. 

AI AND TAGGING

I briefly explored using ChatGPT 4o to create the tags necessary for 
producing a TEI XML document, using instructions such as “Please 
encode this text in TEI XML, providing tagging for entities such as 
dates, people, places, and heavenly bodies. Use terms from LCSH and 
LCNAF.” While ChatGPT provided the shell of a TEI XML document 
and identified many named entities within the text, it didn’t always 
connect that named entity correctly with authority files. For example, 
it connected the term Astronomie with LCSH identifier sh85009152, 
Athletic shoes, rather than with identifier sh85009003, Astronomy. 
(Numerically close, but not quite.) It connected the person Ptholomei 
Pheludiani not with Wikidata identifier Q34943, Ptolemy, but with 
identifier Q303, Elvis Presley. Since many others are pursuing the 
use of AI in named entity recognition, I decided not to pursue my 
investigation of AI-based tagging further at this time.

AI AND IMAGE ANALYSIS

Finally, I tested whether ChatGPT would be able to identify the figures 
pictured in the woodcut on the first page of the Practica (see figure 
1). Using descriptions in Green (2012, esp. 112–13 and 125) of the 
woodcut images usually found in Practica, I am confident that the 

https://nericson.github.io/brelochs/


186    ATL A 2024 PROCEEDINGS

three figures are gods represented by planets due to their situation 
above the earthly sphere and due to each figure being equal in size 
to the others. Furthermore, I am confident that the figures are those 
of the gods Mars, Venus, and Saturn, not only due to Mars’s sword, 
Venus’s feminine form, and Saturn’s scythe, beard, and wool-clad 
feet, but also due to their prominence (Saturn) or specific mention 
(Venus, Mars) in the Practica’s text.

Figure 1: First page of the 1529 Practica Teutsch by Anton Brelochs

Interestingly, in November 2023, ChatGPT 4 correctly identified 
only the central figure of Venus, and with some uncertainty. The 
prompt I used at that time was rather exploratory: “There’s [a] pic-
ture in this document. The document is a scan of a practica teutsch 
from 1529 by Anton Brelochs. I have no idea what the picture is 



Listen and Learn Sessions    187

about. What can you tell me about who is in the picture or what 
they represent?” ChatGPT 4 responded:

The left figure, a man, is dressed in what looks like armor or attire befit-
ting a soldier or knight, suggesting he may represent martial prowess or 
chivalry. He holds what appears to be a sundial, which could symbolize 
the importance of time or astronomy.

The central figure, which appears to be a partially unclothed woman, is 
surrounded by a circle filled with stars and a landscape, likely representing 
the celestial sphere or the universe. She might symbolize nature, virtue, or 
even a specific celestial body like Venus, often associated with femininity.

The right figure, another man, is holding a staff and book, wearing robes 
that suggest scholarly or clerical status, possibly representing wisdom 
or learning. He could be a scholar, a cleric, or an astrologer, given the 
context of the image.

That sounds plausible enough. However, it’s not quite the full 
picture—and, notably, ChatGPT improved its interpretation in its 
next version. In June 2024, ChatGPT 4o clearly identified all three 
figures in response to the prompt “Here is the opening woodcut from 
a 1529 practica teutsch, or astrological prognostication pamphlet. 
What is being explicitly or implicitly depicted or represented? Who 
are the figures depicted?”:

Left Figure: A soldier or warrior, depicted with armor and a sword. This 
figure represents Mars, the Roman god of war, symbolizing strength, 
conflict, and aggression.

Middle Figure: A naked woman, holding a book and with a peaceful 
expression. This figure likely represents Venus, the Roman goddess of 
love, beauty, and harmony.

Right Figure: An elderly man with a long beard, holding a scythe. This 
figure represents Saturn, the Roman god associated with time, aging, 
and harvest, often depicted as the grim reaper.

Based on this example, ChatGPT appears to hold great potential 
for helping to interpret artwork—at least providing suggestions for 
research, if not solid conclusions.

CONCLUSIONS AND FUTURE RESEARCH

Based on this exploration with the AI-based tools Transkribus and 
ChatGPT, I conclude that current AI-based tools have some usefulness 



188    ATL A 2024 PROCEEDINGS

in the preparation of digital editions and in making early printed 
works more accessible to modern readers. Transkribus quickly 
provides a base transcription that can be edited by a human editor; 
the application also allows for transcription models to be trained 
to improve future transcription. ChatGPT allows the end user to 
quickly assess the content of an early printed work, whether in the 
original language or in translation. For scholarly use, the oversight 
of a human editor remains crucial. There is potential for the use of 
AI in the creation of TEI XML documents and tags, and for its use 
in interpreting artwork of the era.

This paper has explored little about the content of the Practica 
itself. Future work might compare, for example, the defense of 
Christian use of astrology provided by Brelochs in the Practica’s 
preface with Martin Luther’s introduction to the 1527 Wittenberg 
edition of Johannes Lichtenberger’s 1488 Prognosticatio—the pre-
decessor of and prototype for the practicas that followed—in which 
Luther softly pans the usefulness of astrology for knowing anything 
specific about the will of God. Indeed, in such an investigation, AI 
might play a different type of role, namely helping to identify the 
explicit and implicit arguments made by each author and aligning 
points of agreement or disagreement between them.

REFERENCES

Bastida, Ana. 2024. “AI-Assist in FromThePage: Using HTR in People-
Centered Transcription.” FromThePage. Accessed July 14, 2024. 
https://content.fromthepage.com/ai-assist-in-fromthepage-using-
htr-in-people-centered-transcription/. 

Green, Jonathan. 2012. Printing and Prophecy: Prognostication and 
Media Change 1450–1550. Ann Arbor: University of Michigan 
Press.

Institut Deutsche Presseforschung. 2024. “Brelochs, Anton.” 
Biobibliographisches Handbuch der Kalendermacher von 1550 
bis 1750. [Biobibliographical handbook of calendar makers from 
1550 to 1750.] Accessed July 14, 2024. https://www.presseforschung.
uni-bremen.de/dokuwiki/doku.php?id=brelochs_anton.

https://content.fromthepage.com/ai-assist-in-fromthepage-using-htr-in-people-centered-transcription/
https://content.fromthepage.com/ai-assist-in-fromthepage-using-htr-in-people-centered-transcription/
https://www.presseforschung.uni-bremen.de/dokuwiki/doku.php?id=brelochs_anton
https://www.presseforschung.uni-bremen.de/dokuwiki/doku.php?id=brelochs_anton

