Large language models such as OpenAI’s ChatGPT have rapidly become prominent tools in academic contexts, including digital scholarly editing. Although not originally designed for editorial practice, they are increasingly integrated into experimental editorial workflows. This review assesses its potential to support key editorial tasks and surveys recent literature on their use in digital scholarly editing. Drawing on published case studies and our own experiments, we analyze the benefits and limitations of employing ChatGPT at different stages of creating and analyzing digital scholarly editions, including transcription, TEI encoding, and named-entity recognition. At the same time, the review addresses significant concerns. As a commercial and largely closed system, ChatGPT operates as a ‘black box’ that allows only indirect control through prompting, raising challenges for transparency and reproducibility in academic contexts. Issues of model bias and data governance must be carefully considered when integrating such tools into research workflows. ChatGPT should therefore be adopted cautiously and critically. Nevertheless, its capacity to lower technical barriers and to consolidate tasks that previously required multiple tools and advanced programming skills makes it, or similar artificial intelligence tools, a potentially valuable assistant in digital editing workflows.
In the rapidly evolving landscape of the Digital Humanities, large
language models (LLMs) Martina Scholger and Roman Bleier are both affiliated with the Department of
Digital Humanities at the University of Graz, where they work in the field of
digital scholarly editing—Scholger as a senior scientist and Bleier as a
postdoctoral researcher. We thank Christopher Pollin of DH Craft for reading an
early draft of this article and for providing valuable feedback and helpful
suggestions, and Thea Schaaf for proofreading.
Introduced in November 2022 (OpenAI 2024), the chatbot ChatGPT is a generative artificial intelligence
(AI) system that communicates through natural language, processing user inputs—known
as prompts—to produce textual responses. ChatGPT is developed by OpenAI, a company
founded in 2015 in San Francisco by a group of machine learning experts, research
engineers, and scientists.
The chatbot can be used as an assisting tool for scholars in the process
of editing, annotating, and curating (historical) texts. The authors evaluate ChatGPT
primarily as practitioner-users, i.e., as editors working with the web-based
graphical user interface, rather than as developers engaging with the underlying
models or APIs. All experiments were performed using the ChatGPT web interface,
specifically the GPT-4o model, accessed between July 2024 and July 2025, under both
free and Plus tier access. The source material used in the experiments comes from
digital edition projects in which the authors are involved, namely
The technical architecture of ChatGPT relies on Generative Pre-trained Transformer (GPT) models, which are fine-tuned for natural conversation using a combination of supervised learning and reinforcement learning from human feedback. These LLMs belong to the broader class of foundation models, which are designed to support multiple downstream applications rather than being optimized for a particular, narrowly defined task. Therefore, ChatGPT can in principle be applied to a wide range of tasks at different stages of a digital scholarly editing workflow.
The exact training data of the underlying models is not publicly
disclosed, and the models themselves are not openly available. According to
OpenAI, training data consists of “information that is publicly available on the
internet, information that we license from third parties, and information that our
users or our human trainers provide”.
Another factor affecting reproducibility concerns the availability of different model versions within the ChatGPT platform. As of June 2025, free-tier users have access by default to GPT-4o, OpenAI’s flagship multimodal model, though access is subject to usage limits. Once these limits are reached, usage automatically defaults to a smaller, cost-efficient fallback. Users of subscription tiers such as Plus, Pro, or Team receive higher usage limits and access to a wider range of models through the model selector, including models optimized for complex reasoning tasks, high-performance multimodal interaction, and lightweight, cost-efficient processing. The presence of multiple model versions within a single platform therefore introduces an additional layer of unpredictability for scholarly workflows.
ChatGPT operates through a conversational interface in which users interact with the system using natural language prompts. While text is the primary input modality, interactions can also begin with images or voice. A prompt is the instruction or query provided to the model, ranging from simple questions to complex scenarios, directing the LLM in its response generation. In addition to user prompts, the system relies on internal system prompts that shape the model behavior, but remain invisible to users. Prompt engineering refers to the practice of designing prompts to guide model outputs effectively. Considerations of prompt length and structure are also relevant, as model inputs are processed as tokens and subject to context window limits. There are various prompting techniques, such as zero-shot, few-shot, and chain-of-thought prompting, each designed to approach the model differently based on the complexity of the task (Saravia 2022). In editorial contexts, careful prompting is essential for achieving consistent results, particularly when tasks involve formalized output formats or domain-specific conventions. Mastering prompt engineering has become its own discipline and is particularly important in digital humanities contexts, where tasks frequently involve specialized terminology, historical language material, and encoding standards that require precise instructions (Saravia 2022, Pollin 2025).
All inputs and outputs processed by ChatGPT are internally segmented
into tokens, which constitute the basic unit for both text processing and usage
accounting. Tokens are the building blocks of text that OpenAI models process. They
can be as short as a single character or as long as a full word, depending
on the language and context. Spaces, punctuation, and partial words all
contribute to token counts. This is how the API internally segments your
text before generating a response.
How exactly tokens are calculated is often not immediately intuitive
for users. OpenAI suggests that, as a rule of thumb, for English texts one token
corresponds to approximately four characters, or about three-quarters of a word.
Consequently, 100 tokens amount to roughly 75 English words. OpenAI provides an
online tokenizer that allows users to test how specific inputs are segmented into
tokens.
Beyond prompt-level interaction, ChatGPT also allows users to
configure model behavior through customizable applications known as Custom GPTs.
For creating them, no programming knowledge is necessary. The configuration screen
allows you to enter a picture, a name of the GPT, a short description, detailed
instructions, a conversation starter to facilitate the start of a conversation,
and the upload of an extra knowledge base. When such files are provided, the
system can retrieve relevant passages from them during response generation,
thereby implementing a form of retrieval-augmented generation (RAG). It is
recommended to provide the additional knowledge in a minimalistic format (e.g.,
JSON or Markdown) to save tokens instead of ‘chatty’ formats like XML, where each
tag consumes tokens. While this built-in retrieval mechanism is suitable for
small-scale text documents, larger corpora typically require interfacing with
external databases.
ChatGPT can be individually configured using ‘custom instructions’, a personalization feature distinct from Custom GPTs, available in the user settings. These allow users to provide details about their interests and working context and to specify how the model should respond, for example, by indicating a preferred tone or level of detail. The instructions persist across conversations so that the user does not have to repeat these preferences in each prompt.
ChatGPT can handle and process text in various data formats like plain text, XML, HTML, JSON, CSV, or Markdown, as well as multimodal inputs such as image, video, and audio as input. Likewise, it can generate output in most textual data formats; in addition, newer multimodal models support synthesized audio output, while the generation of standalone video files is not natively supported in the standard ChatGPT interface. Advanced Data Analysis (ADA), an integrated capability of GPT-4-class models, further expands its functionalities to the automatic generation of Python code for performing exploratory data analysis tasks. It can process structured data formats, including Excel, CSV, and JSON, and produce outputs as Python code, statistical summaries, visualizations, charts, interactive tables, and more, thereby supporting eexploratory data analysis within a single environment
For developers, OpenAI provides convenient and well-documented access to its API through multiple interfaces, including the programming language Python, the cURL command-line tool for making HTTP requests to APIs, and Node.js, a widely used JavaScript runtime environment to build applications, provided one has an API key. The API supports integration into a wide range of software systems, applications and services. However, the focus of this article is the chatbot interface, and the API and its development capabilities will therefore not be discussed in detail.
ChatGPT exists within a broader and rapidly evolving ecosystem of generative language models. Proprietary systems include Claude by Anthropic, Gemini by Google, as well as models integrated into platforms like Microsoft Copilot. At the same time, open-weight and open-source models are advancing rapidly and in some benchmarks are even outperforming proprietary systems in certain tasks (Chen et al. 2024). Prominent examples include the Llama family developed by Meta, Mistral AI’s Mistral and Mixtral models, and Alibaba’s Qwen series.
When assessing a language model’s openness, it is important to
differentiate between open-source models and open-weight models. Open source
refers to sharing the full source code, the underlying algorithms, training
methods, and data, while open-weight models only share the pretrained model
parameters, which allows use but limits transparency and control. Open-weight
models are easier to use but less transparent, while open-source promotes
transparency and allows modification, but also requires more developer effort
(Ramlochan 2023). For
digital scholarly editing, these distinctions are relevant for sustainability,
inspectability, and long-term reuse. The current leading open-source models can be
consulted at the Chatbot Arena.
In conceptualizing how generative AI might support scholarly editing, the Institute for Documentology and Scholarly Editing (IDE) has proposed a processing pipeline (Pollin et al. 2025), tracing the path from source material through edition as data to edition as publication via intermediate editorial products such as digital images, transcriptions, and the user interface. It further differentiates between structural tasks (e.g., planning, documentation, tool integration, quality control) and operational tasks (e.g., text recognition, entity recognition, normalization, enrichment). Mapping ChatGPT's capabilities onto this pipeline helps locate the stages at which it may be effectively integrated into editorial workflows, and where their limitations become apparent. We will attempt a brief evaluation of some tasks in this section.
We have already mentioned the multifaceted applications of ChatGPT as a tool to support
editorial work, ranging from text summarization and translation to automated XML/TEI annotation and other
natural language processing (NLP) tasks. It can also be used as a ‘buddy’ to provide
constructive feedback on existing data models or XML/TEI code, helping to identify inconsistencies or
suggest alternative solutions. Furthermore, ChatGPT can assist in the development of small programs or scripts for the web presentation or
data analysis of an edition, utilizing
From a utility perspective, ChatGPT primarily supports exploratory and preparatory stages of editorial work rather than final, authoritative decisions, as its non-deterministic nature and lack of stable versioning limit scholarly reproducibility. While this challenges academic rigor, the non-deterministic behavior can also be used as an advantage. Editors can generate multiple versions of the same output (e.g., alternative TEI encodings) and critically identify the optimal encoding approach. In this way, non-reproducibility can function as a form of digital hermeneutics, potentially transforming what appears to be a limitation into a tool for deeper interpretive engagement with editorial decisions.
GPT models are generally considered not to be trained on specific
domains or tasks, although user-guided fine-tuning can be applied to that end.
They are general-purpose models that can assist in different tasks related to the
digital editing pipeline, from data creation to final publication. One area of
experiments includes custom GPTs, i.e., user-configured applications built on top
of a pre-trained GPT model. Examples include
Beyond individual tools, broader community efforts further illustrate these developments. A workshop at the DHd 2024 conference in Passau, organized by the Institute for Documentology and Scholarly Editing (IDE), provided a broad overview of ongoing experiments with generative language models in digital scholarly editing (Pollin et al. 2024). The ten contributions covered a wide range of editorial tasks, including transcription, OCR post-correction, markup generation, normalization, named entity recognition, annotation, translation, and text summarization. As demonstrated in Pollin's workshop series, generative AI can be applied across diverse DH use-cases: from transforming unstructured data into TEI to generating and modeling RDF data to data analysis and visualization (Pollin 2025). Alongside these exploratory applications, research has emphasized the need for robust pipelines and systematic evaluation in LLM-assisted scholarly workflows. Recent contributions assess, for example, the quality and reliability of LLM-assisted TEI encoding (Strutz and Vogeler 2025, Strutz forthcoming), the automated handling and resolution of editorial interventions (Strutz and Scholger 2025), and the automated extraction and normalization of metadata from historical letter transcriptions (Strutz and Scholger 2026). Galka and Vogeler compare the integration of different LLMs within a workflow for adding named entity recognition annotations (2026). These studies highlight both the potential of LLMs for complex, context-sensitive tasks and the methodological challenges that remain.
In order to evaluate ChatGPT’s practical value for concrete editorial tasks, we selected two common use cases: (1) converting plain text into TEI-compliant XML, and (2) identifying and classifying named entities. All experiments were performed using GPT-4o in the browser interface. For more elaborate examples, we refer to slides including chat protocols in Pollin 2025.
In the first example, a letter from the digital edition Euer Gnaden,
Unterzeichneter1 bittet um die gnädige Erlaubnis,
künftigen Sonntag die Redoute2 besuchen zu dürfen3, weil er
1.) im
letzten und sechsten Jahre seiner Studien und bereits ein Jahr länger an der
Akademie ist als es seine Vorfahren waren, da ihnen diese Erlaubnis erteilt
wurde.
2.) auf das ihm jederzeit von S[einer] H[ochwohlgeboren] dem
Herrn Direktor erteilte Zeugnis seines guten Verhaltens und sittlichen
Betragen stolz ist, und sich daher schmeichelt, dass die Besorgnis einiger
daraus entstehenden üblen Folgen bei ihm nicht statt haben dürfte.
3.)
die Vorschrift einer für diesen Fall geforderten anständigen und sicheren
Begleitung nicht besser erfüllt haben könnte, indem er in der Gesellschaft
der Frau Schwester des Herrn Direktors sich befinden wird.
Den 20ten
Hornung 1794
Joseph Hammer
1.) PvTh D.2.1.14.
2.)
Eine Tanzveranstaltung oder auch ein eleganter Maskenball. Um welche
derartige Veranstaltung es sich hier handelt, ist nicht feststellbar.
3.) Die Erledigung durch Jenisch s. das nachfolgende Stück,
findet sich auf demselben Blatt wie die vorangegangene Bitte HPs mit dem
Vermerk „Von der Kanzlei den 20. Hornung 1794“.
Prompt instructions to convert a letter from Joseph von
Hammer-Purgstall to Bernhard von Jenisch into TEI.
XML/TEI encoding of a letter from Hammer-Purgstall to
Jenisch generated by ChatGPT (GPT-4o).
In the second example, the same TEI annotation task was applied to
reports of envoys present at the Imperial Diet of the Holy Roman Empire that
was held in Regensburg in the year 1576 (Leeb et al. 2023). The envoys, Dr. Wolfgang Eulenbeck
and Dr. Andreas Paulus, reported regularly to Prince-elector (Kurfürst) August
von Sachsen, who was not present at the Imperial Diet. The form and structure
of these reports are very similar to the above letter from the
The following quote shows the plain text of the letter from May 9
1576, to August von Sachsen: Neunden Maii, Regenspurgk rhäte. Präs. Annaburg, 12. Maii anno 76.
Zu s. kfl. Gn. henden.
Durchlauchtigster hochgeborner churfurst,
ewern kfl. Gn. seint unsere underthenigste gehorsame dienste jeder zeit
mit trewem vleiß zuvorn. Gnedigster herre.
Auf euer kfl. Gn.
gnedigsten bevelch und uns zugeschickter instruction haben wir uns aufn
wegk nach Regensburgk erhoben, und seint gestern umb acht schlage vor
mittage alhier ankommen. Und wiewol wir under wegs verstendigt, als solte
die röm. ksl. Mt. auf den vierzehenden tag ditz monats Maii alhie
angelangen, so werden wir doch durch des Reichs marschalch alhier
berichtet, das sich irer ksl. Mt. aufbruch zu Wyen bis in mitten ermelts
monats verweilett, und ire Mt. wenig vor ende desselben alhier ankommen
werden, inmassen euer kfl. Gn. aus inligender copey der höchstgedachten
ksl. Mt. schreiben an den bischoff [1v] alhier, item hern Conraden zu
Pappenheim und doctor Viecheussern gnedigst haben zuersehen.
So
thuen euern kfl. Gn. wir hirbey underthenigst ein vorzeichnus ubersenden,
was vor gesandten (derer doch wenig sein) bis auf itzo alhier
angelangt.
Und, euer kfl. Gn. mit underthenigstem gehorsamb trewe
dienste zuleisten, seint wir schuldig und geflissen. Datum Regensburgk,
den neunden Maii anno 76.
Ewer kfl. Gn.
underthenigste
gehorsame
Wolffgang Eylenbeck, Dr.
Andreas Paull
propria[?]
The letter to August von Sachsen encoded in TEI using
ChatGPT (GPT-4o). The entire second paragraph was left out in order to
fit the beginning and end on the picture.
Close observation of the result shows that some text is missing.
For some reason, the part of the opener indicating when the letter was
presented to August von Sachsen: ‘Präs. Annaburg, 12. Maii anno 76’ is missing.
Furthermore, identification and expansion of abbreviations is a challenge.
While abbreviations, such as ‘s.’, ‘Mt.’ and ‘kfl.’ are correctly resolved,
others were not even identified as abbreviations, e.g., ‘Gn.’, or words were
incorrectly identified as abbreviations, e.g., ‘ewern’, ‘bevelch’,
‘instruction’. This can easily be resolved by providing a list of the
abbreviations used in editions of the Imperial Diets.@when
attribute. However, ChatGPT does not automatically take the Gregorian calendar
reform into consideration. It has to be explicitly mentioned to add
@calendar.
First result from ChatGPT trying to identify persons and
places in the letter of 1576. Wrongly identified places are
highlighted.
Geoname references of the places generated by
ChatGPT.
ChatGPT can be accessed via the OpenAI website, and an API is
available for programmatic use. The tool requires minimal setup, and the web
interface is minimalistic and straightforward. It works well on all modern web
browsers, and a ChatGPT app is available for mobile devices. Users require
internet access and an OpenAI account. The interface is designed primarily for
conversational interaction—asking questions and receiving responses—and keeps
this core functionality readily accessible. Critical remarks have been raised,
however, regarding its simplicity and limited customization options (e.g., Sebastian 2023). Additional
services beyond the free tier, including expanded API usage and processing
capabilities, are only available under a pay plan based on OpenAI’s token
system.
The various language models used within ChatGPT, and their
differences, are documented and compared on the OpenAI website.
ChatGPT is described as a multilingual chatbot that can
communicate in over 50 languages.
AI tools, including ChatGPT, are frequently discussed in the
context of accessibility and the provision of barrier-free content. The OpenAI
web interface is simplistic and screen-reader friendly. But the actual benefit
comes from the use of the ChatGPT API in various new applications. For
instance, such systems can generate automated descriptions of images, real-time
captions, and simplify complex texts into easier language.
There is a broad spectrum of ethical concerns being discussed in the context of AI use. Applying ChatGPT in an editorial workflow should therefore be carefully considered, and not only its benefits but also its broader implications must be taken into account. For instance, the increase in power consumption of technology companies and the building of enormous data centers cause environmental issues, which call for sustainable solutions when adopting AI in daily work (Rohde et al. 2024).
The responsibility for output that an AI tool produces is particularly important in an editorial context. ChatGPT and similar tools can create image-based transcriptions, translations, and, as we have seen, also XML/TEI markup. Nevertheless, the tool may also introduce errors and interpretations of the source material not intended by the editor. Like in all AI-driven workflows, also in editorial practices, the concept of trustworthy AI, focusing on transparency, accountability, reliability, and fairness, should also be considered (AI HLEG 2019, Pope 2024). A comprehensive overview and analysis of the various discussions surrounding the ethical dimensions is provided by Stahl and Eke (2024). Our final section focuses primarily on questions of data governance, transparency through citation, and bias.
OpenAI assures its users that all data shared with ChatGPT is
encrypted, and that inputs and outputs they receive are rightfully theirs.
In academic contexts—research, academic writing, and teaching—the
question of how to cite (and acknowledge) the use of ChatGPT poses several
issues. Conversations are tightly bound to a user account, and there are no
permalinks, no DOI, and no URL that allows for persistent sharing. There is not
yet a general consensus on how to cite ChatGPT conversations.
Finally, we briefly address the question of bias. Even if we
‘talk’ to a machine, ChatGPT is not objective. The models are trained on an
enormous amount of data, most of it from a North American and European context,
and the biases introduced through that data influence its responses. Gender,
racial, religious, age, or political bias, and other negative behaviors have
been identified and criticized.
ChatGPT is not a tool that was explicitly developed with the digital humanities or the scholarly editing community in mind. Nevertheless, it has a wide range of potential applications for researchers and editors, and can support different stages of the development or analysis of digital editions. It can assist with transcription and data creation, support the development of scripts for data processing and analysis, and facilitate related tasks.
There is still relatively little research and benchmarks on the use of LLMs for editorial purposes, and the possibilities of what can be done with ChatGPT and similar AI tools are constantly changing. We are aware that the evaluation of such technology is only a snapshot that might not be relevant in a couple of months, and that all predictions for future developments are quite difficult, as development moves so quickly. As we outlined in the article, we chose to evaluate ChatGPT (GPT-4o) over other LLMs, as it is well-known and currently widely used by humanities scholars.
Several limitations significantly constrain the use of ChatGPT as a tool for scholarly editing. First, the development of the tool is closed off as it is a commercial product. For editors, this results in a black-box situation that can be vaguely controlled using prompts. If prompts are not carefully crafted, the results can be misleading or even totally wrong. Second, it remains unclear how OpenAI handles the data provided by users and to what extent the data is used to train future models. Therefore, users have been warned to be careful and not to share sensitive or unpublished material with AI. For scholarly editing, this significantly restricts the range of sources for which ChatGPT can be used responsibly. Third, while the chatbot interface is suitable for small-scale experiments and exploratory tasks, complex multi-stage editorial workflows for an entire corpus—from the TEI encoding to the recognition of named entities and their enrichment with norm data—can only be effectively implemented via the API. The associated costs cannot be covered by every project and also contradict the principles of an open-knowledge society and equal access to research. This is why it is so important that we focus more on developments in the area of open-source LLMs in the future. Finally, ethical issues caused by the bias in the models have to be kept in mind when incorporating these tools into scholarly workflows.
The use cases and literature presented in this review demonstrate ChatGPT’s high potential as an assistant for digital editing. Tasks that previously required several tools and technical expertise can now be performed with comparatively little effort using carefully designed prompting instructions. At the same time, we have also discussed that the quality of the results can diverge depending on the source material of the edition and a user’s ability to craft ‘good prompts’. In any case, a human editor must remain in the loop to control, evaluate, and, if necessary, further process the output. This is especially highlighted by our little experiment using German texts from two different historical language periods. It shows that ChatGPT works less effectively with non-standard and historical language material. As models continue to improve and open-source alternatives mature, the more pressing question for the digital editing community is no longer whether to engage with these tools, but how to use them in transparent, critically informed, and methodologically rigorous ways.