The integration of AI technology and AI-based tools (e.g., machine learning (ML), natural language processing (NLP)) has accelerated the ‘computational turn’ of the field of DH, expanding its capacity to handle more complex datasets, uncover hidden patterns, and automate processes that enhanced both the scale and depth of the scholarly inquiry (Berry, 2022; Gefen et al., 2021). In text analysis, for instance, AI-powered tools, such as NLP, have facilitated the examination of large text corpora, enabling the practice of ‘distant reading’ (Moretti, 2013). Furthermore, a growing focus on data and algorithms—particularly on how to conceptualize and work with data, and how to understand as well as address the opaque nature of algorithmic processes and their underlying data transformations—has brought important questions to the forefront of DH research (Bode, 2019; D’Ignazio & Klein, 2020; Drucker, 2011; Lavin, 2021; Schmidt, 2016). In addition to enabling methodological transformation, AI-driven automation has reshaped how labor is conceptualized and executed in DH, introducing new opportunities for interdisciplinary collaborations and research innovation (Smithies, 2017). Building on AI’s longstanding presence in DH, the recent rise of generative AI technology presents both new possibilities and challenges for the field. Researchers have begun exploring how this technology can be integrated into DH research and pedagogical practices in ways that are effective, critical, and ethical (Chun & Elkins, 2023; Dedema & Ma, 2024; Guo, 2024; Liu et al., 2023).

Despite the recent progress in exploring AI applications within DH and the acknowledgement of the importance of data and related ethical implications in DH research (Gonzalez & Rodrigues, 2022), a comprehensive, empirical investigation of how AI has been reflected in DH scholarship remains limited. This is especially true when considering the historical integration of AI technologies into DH publications and research. With the recent surge in generative AI technologies, a historical perspective on AI’s use in DH is increasingly valuable for identifying patterns, gaps, and emerging challenges within the field. Our paper responds to this gap by examining AI use in DH scholarship, tracing its evolution from the 1980s to the present.

Bibliometric analysis offers valuable methods for illustrating broad trends and usage patterns in DH scholarship through empirical data. Previous studies have highlighted the strengths of bibliometric approaches—such as citation analysis, network analysis, and topic modeling—in examining DH as an interdisciplinary field and research community. For instance, Luhmann and Burghardt (2021) analysed articles from DH journals alongside those from 15 other academic disciplines, addressing the conceptual identity of DH—one of the field’s enduring debates—and underscored how DH connects with neighbouring fields like computational linguistics and information science. Similarly, Ma and Li (2021) introduced the ‘battleground’ metaphor, emphasizing the tensions and collaborations between humanities and STEM research practices and conventions. Earlier works demonstrated that bibliometric analysis also sheds light on the early history and emergence of DH as a field (Leydesdorff & Akdag Salah, 2010; Sula & Hill, 2019), as well as its gradual evolution into a ‘Big Tent’ discipline (Weingart & Eichmann-Kalwara, 2017). Furthermore, citation and network analysis using DH journal article data has provided insights into global DH research collaboration patterns and community structures (Gao et al., 2017, 2018; Li et al., 2024; Nyhan & Duke-Williams, 2014; Tang et al., 2017).

In this paper, we leverage the unique methodological advantages of bibliometric analysis and apply it to examine a publication corpus of three influential DH journals, highlighting the broad patterns of AI adoption in DH scholarship.

We collected metadata from all articles published in three influential, ‘exclusively DH’ journals: DSH, DHQ, and JCA (Spinaci et al., 2020). Published since 1986, DSH allows us to consider nearly 40 years of DH history. Metadata for DSH articles were parsed from HTML files collected by one of the authors using their university library’s subscription and access to the journal; the DHQ metadata were extracted from XML files available in the journal’s GitHub repository (https://github.com/Digital-Humanities-Quarterly/dhq-journal/tree/main/articles). For JCA, the metadata were obtained by web scraping the HTML source code of the journal’s online open-access articles.

In this study, we analysed abstracts rather than full-text articles because abstracts provide higher information density. We expect that articles focused on AI will include relevant terms in their abstracts. Therefore, we excluded articles without abstracts or publication years, as well as those with metadata information indicating that the articles are not in English. This resulted in a corpus of 1,587 abstracts from DSH (69.3% of the total 2,290 articles published between 1986 and 2023), 720 abstracts from DHQ (95.2% of the total 756 articles published between 2007 and 2024), and 181 abstracts from JCA (99.5% of the total 182 articles published between 2016 and 2024). After data collection, four data analysis experiments were conducted to examine the discourse surrounding AI in these abstracts: temporal analysis of AI mentions, collocation analysis, word vector analysis, and topic modeling.

For the temporal analysis of AI mentions, we measured the presence of AI terms in each abstract and tracked their variation over time. We defined ‘AI terms’ as ‘AI’ and ‘artificial intelligence’ only, rather than creating an extended thesaurus, to ensure that all terms discussed in this study specifically refer to AI. Abstracts were identified as containing AI mentions if they included either the word ‘ai’ or the 2-gram ‘artificial intelligence’ (where ‘artificial’ is immediately followed by ‘intelligence’ within the same sentence). We then calculated the proportion of articles with AI mentions in their abstracts published each year.

Collocation analysis was conducted to identify the meaningful terms that co-occur with AI terms in each abstract. We first generated a word count table using the CountVectorizer from the Python module Scikit-learn (Pedregosa et al., 2011). To reduce morphological variants of words to a common stem, we applied the SnowballStemmer for English from the NLTK library (https://www.nltk.org/api/nltk.stem.SnowballStemmer.html). Next, we calculated the number of abstracts in which each stemmed word co-occurs with AI terms. Following the ideas of Algee- Hewitt et al. (2020) and Gilkison and Kurzynski (2024), we employed Fisher’s exact test (Fisher, 1922) to identify statistically significant collocates of AI terms, excluding common words like ‘the’ and ‘a’. Additionally, referencing Gilkison and Kurzynski (2024), we calculated the ratio of the actual number of abstracts where each stemmed word co-occurs with AI terms to the expected number of such co-occurrences (r-exp), which is defined as:

In addition to words that co-occur with AI terms, we also identified words with similar semantic meanings with AI terms through word vector analysis, which quantifies the semantic similarity between terms based on their contextual usage. Using the Python library Gensim (Řehůřek & Sojka, 2010), we pre-processed the corpus with simple_process and converted each stemmed word into a vector with the Word2Vec model (Mikolov et al., 2013). For better performance with our relatively small dataset, we adjusted the ‘min_count’ parameter to 1 to ensure no word was excluded, and switched the ‘sg’ parameter from 0 to 1 to use the skip-gram algorithm instead of the CBOW algorithm. We then calculated the cosine similarity between each stemmed word’s vector and the vector of the word ‘ai’, ranking their semantic similarity based on these cosine similarity scores.

Topic modeling is a technique used to ‘discover hidden thematic structure in large collections of texts’, and is widely used in DH (Blei, 2012). For this study, we applied topic modeling to examine the themes in abstracts and their relationship with AI terms. We used the MALLET (MAchine Learning for LanguagE Toolkit) for this purpose, which offers efficient sampling-based implementations of latent Dirichlet allocation (LDA), Pachinko allocation, and hierarchical LDA in its topic modeling toolkits (https://github.com/mimno/Mallet) and is commonly used by digital humanists (e.g., Graham et al., 2012). We imported our corpus into MALLET and generated 30 topics. When importing data into MALLET, we specified the ‘--token-regex’ as ‘[\p{L}] +,’ since the default regular expression removes words with 1 or 2 characters, including “ai.” We also combined the default English stop word list with the French stop word list from https://github.com/stopwords-iso/stopwords-fr (excluding the word ‘ai’). The French stop word list was necessary because some abstracts contain French words without metadata indicating their language, making it impossible to filter them out during data collection. In determining the number of topics, we tested various topic numbers in increments of 10 (from 10 to 50) and observed that the log-likelihood initially rose, then declined, reaching its peak at 30. We then compared the average weight of each topic in abstracts with AI terms to those without. Finally, the three co-authors independently labelled each topic based on the 20 keywords for each topic, and then synthesized their results to determine the final label for each topic.

Out of a total of 2,488 abstracts, 28 (1.1%) contain AI terms. The temporal variation in the proportion of articles with AI mentions in their abstracts published each year is illustrated in Figure 1.

Overall, the leading DH journals show a sharp increase in AI mentions in their abstracts after 2019, with a particularly notable surge beginning in 2023. This trend aligns with the recent development of AI technology, especially generative AI, which has drawn significantly increased attention from DH scholars, potentially leading to more extensive discussions in these journals. Additionally, it is noteworthy that four articles mentioning AI terms were published before this recent rise. With the earliest dating back to 1990, they applied AI to tasks including text interpretation (De Vuyst, 1990; Terras, 2005), text encoding (Terras & Robertson, 2004), and optical character recognition (OCR) (Heckmann et al., 2014). This finding supports Berry’s (2022) observation of digital humanists’ extended history of using automation (pp. 446–447), which predates the recent surge of AI in public discourse, and further highlights that the term ‘AI’ was occasionally explicitly used in these early practices.

There are 16 stemmed words identified as statistically significant collocates of AI terms (p < 0.001 in Fisher’s exact test, excluding ‘ai’, ‘artifici’, and ‘intellig’). These collocates are presented in Table 1, ranked in ascending order of p-values (i.e., more statistically significant collocates appear higher in the list).

The 16 statistically significant collocates of AI terms span both technical and humanities-oriented concepts. Technical terms like ‘system’, ‘machin’, ‘openai’, and ‘gpt-2’ highlight the technical focus in discussions of AI within DH scholarship. Simultaneously, humanities-oriented terms such as ‘historian’ indicate the application of AI in specific humanities disciplines and underscore the prominence of humanistic perspectives in DH work. Additionally, the presence of terms such as ‘ethic’ further highlights the field’s emphasis on ethical issues. This observation is also echoed in digital humanities scholarship (e.g., Gonzalez & Rodrigues, 2022), especially in the context of teaching with emerging AI technologies (e.g., Chun & Elkins, 2023; Berry, 2022).

Among the 12,418 distinct stemmed words (excluding ‘ai’ itself), it is unsurprising that many of the terms with the highest semantic similarity to ‘ai’ are technical. For example, ‘intellig’ (5th), ‘portal’ (6th), and ‘microcomput’ (11th) were among the terms ranked the highest in semantic similarity to ‘ai.’ This potentially demonstrates an alignment between DH and information science in perceiving these technical elements as integral to the study of AI.

As shown in Figure 2, the 28 abstracts containing AI terms have substantially higher average weights (at least twice as high) in topics 10 (Ancient Texts), 17 (Digital Cultural Heritage), 23 (Narratology), 24 (Machine Learning), and 25 (Digital Media) compared to the 2,460 abstracts without AI terms.

Topics associated with AI in the topic modeling process aligned with themes previously revealed in the analysis (i.e., technical and humanities-oriented). Technical aspects are strongly demonstrated in Topic 24, which is comprised of keywords related to ML methodology (e.g., ‘machine’, ‘learning’, ‘accuracy’, ‘classification’, ‘features’). Articles with abstracts where Topic 24 is dominant (weight > 0.4) reflect a particular interest in technical aspects, including the application of AI and ML to tasks like OCR (Heckmann et al., 2014), text classification (Zheng & Jin, 2019), and information extraction (Sánchez-Salido et al., 2023).

AI is also aligned with humanities-oriented themes. Topics 10, 17, 23, and 25 each represent an intersection between DH and a specific humanities field. Topic 10 highlights keywords related to ancient texts and language systems (e.g., ‘Greek’, ‘ancient’, ‘language’, ‘system’, ‘linguistic’); Keywords of Topic 17 centres on cultural heritage (e.g., ‘cultural’, ‘heritage’, ‘historical’, ‘archive’, ‘collection’); those of Topic 23 focuses on narratology (e.g., ‘narrative’,’ character’, ‘story’, ‘novel’, ‘plot’); and those of Topic 25 concerns media studies (e.g., ‘media’, ‘video’, ‘game’, ‘art’, ‘virtual’). This suggests that certain humanities disciplines—linguistics, heritage science, media studies, and a subfield of literature, narratology—may be at the forefront of incorporating AI advancements, reflecting their active engagement in AI-driven research.