A (Feminist) Toolkit for Data Ethics

With the growing accessibility of machine learning tools, applications and approaches, the Digital and Computational Humanities face an urgent need to focus on ethical frameworks, research integrity and quality assurance in research outcomes. Cf. D’Ignazio / Klein 2020. This working paper builds on discussions and co-organized events of the Digital Humanities im deutschsprachigen Raum (DHd) working group ›AG Empowerment‹. However, the paper itself was written exclusively by the two credited authors, both historians by training, situated at German universities and research institutions, who were not introduced to feminism through formal training. ›AG Empowerment‹ is made up of predominantly female and non-binary early career scholars in Digital Humanities from the German-speaking areas with different disciplinary backgrounds in the Humanities. With our work, we hope to make a positive contribution to discourses aiming to make the Digital Humanities more inclusive, yet we as a group are not truly diverse, hence the working group name Empowerment. We approach our work through a lens of critical whiteness and believe, following the Data Feminism manifesto, that better outcomes for all can only be achieved through co-liberation, a process that requires participation by all and cannot be done for anyone. Accordingly, our goal is to help empower ourselves and others to participate in this process. While traditional source criticism is well established in the historical disciplines, the nature of digital projects necessitates a digital source criticism that evaluates not only historical sources but also the processes involved in creating and curating data. Data Feminism offers a compelling framework for addressing ethical challenges, but it remains underutilized within Digital Humanities (DH). has outlined seven principles to guide ethical data work in the Digital and Computational Humanities, grounded in intersectional feminist power analysis. It builds on works such as Noble 2018 and Eubanks 2018 and leverages intersectional feminist concepts like the matrix of oppression (Collins 2008) and the situatedness of knowledge (Haraway 1988). It also incorporates ideas like racial innocence (Bernstein 2011) and white fragility (DiAngelo 2019) to define the privilege hazard (D’Ignazio / Klein 2020, pp. 28–29), which refers to the inability of those in positions of power and privilege to foresee the potential harm that systems they design may cause due to their lack of lived experience. On Data Feminism for AI, see Klein  / D’Ignazio 2024. This raises the questions: Why is there a hesitation to adopt a Data Feminist approach in the Digital Humanities, and how can the principles of Data Feminism be applied within a Digital Humanities framework?

We argue that the limited adoption of Data Feminism in Digital Humanities can be attributed to a combination of misconceptions and practical challenges. One possible cause may be the assumption that Data Feminism principles only concern issues about gender or women in particular. For this reason, we have titled this article ›Beyond Data Feminism‹. This is not because we believe Data Feminism is somehow lacking or should be left behind, but to emphasize that despite its name, Data Feminism is not (just) about feminism or women. The reach of the proposed concepts and principles should extend beyond the disciplinary borders of what is typically associated with feminism. D’Ignazio and Klein make this important distinction when they explain that while women are not a minority, they are still not the dominant group within patriarchal structures and experience systemic oppression compounded by intersecting identities such as race and class. The authors refer to this as being minoritized (Smith 2016) within their specific context of power dynamics (D’Ignazio / Klein 2020, p. 26). Wernimont / Losh 2018, pp. ix–x, for example, have criticized that critical Digital Humanities and feminist Digital Humanities are continually suffering marginalization as ›fringe interests‹ at the periphery of the field, even though women are not a minority in Digital Humanities (Eichmann-Kalwara et al. 2018). Data Feminism addresses the interplay of dominant and minoritized groups and how these dynamics shape how data is collected, archived, analyzed, and used. But due to women being a prominent example of a minoritized group, it is understandable why many may have erroneously confused this with being about women. This, however, is not the case: Data Feminism, as articulated by Catherine D’Ignazio and Lauren F. Klein, focuses on power structures and how marginalization shapes data. In providing guidelines which operationalize feminist thought as actionable principles, the authors of make clear that a project can be Data Feminist in content, form, or process. Cf. D’Ignazio / Klein 2020, p. 18. Yet misconceptions about what Data Feminism means and who it is for may offer an explanation why Digital Humanities practitioners only sparingly engage with its principles, likely underestimating their universal applicability. Data Feminism applies to all data-related projects because every project is inherently political, whether its agenda is explicit or implicit. The manifesto provides tools for addressing power imbalances in data work by drawing on feminist theory’s longstanding critique of such structures. The principles build on established strategies that can be applied to (any) data work. As such, Data Feminism is a toolkit for dealing with ethical issues in data projects and thus, highly relevant for all of Digital Humanities. But despite its potential, Data Feminism has yet to achieve widespread recognition as a standard ethical framework for Digital Humanities. This likely limits its broader adoption, creating a barrier to entry that hinders the significant policy changes that could arise from widespread acceptance of Data Feminist principles.

And, while the principles in Data Feminism are clearly articulated and memorable, they are broad guidelines that require adaptation to specific contexts and the operationalization of Data Feminism for a practical (Digital Humanities-)use. The book serves more as a manifesto than a practical guide, although its core goal is to present actionable principles for incorporating intersectional feminist approaches into data work (Alvarado 2022). The same can be said, maybe even more so, for the more recent article (Klein / D’Ignazio 2024) in which the authors of applied the principles originally developed for data science to the field of AI. The principles are very valuable but due to their nature as principles, they are not guidelines suitable for immediate practical implementation. Many demands in concern policymakers more than individual Digital Humanities practitioners seeking to make a positive contribution to a more equitable AI landscape. As such, despite the stated aim that the principles of Data Feminism should operationalize feminist strategies for engaging with power structures in ways that are useful for those working with data, they largely remain general guiding principles. While they can serve as guides to one’s modes of thinking and approaches to data, they are not yet fully operationalized in a technical sense that would translate directly into concrete, actionable steps for a typical Digital Humanities project without further ›translation work‹ (Pichler / Reiter 2022). On operationalization in Digital Humanities, see Pichler / Reiter 2022. On data work, see Alvarado 2022. Many practitioners lack the time or resources to engage deeply with feminist theory, creating a barrier to its adoption. Even those familiar with feminist principles may need to read and re-read the book to fully engage with its content, a time commitment that many Digital Humanities professionals cannot afford within their projects. While it would, of course, be best if everybody could engage with the principles deeply, providing accessible entry points that can be seamlessly integrated into existing Digital Humanities work is more likely to have a significant impact and encourage a gradual shift towards better practices.

In this article, we aim to ›translate Data Feminism‹ for more traditionally minded (Digital) Humanists. We provide both theoretical context and concrete implementation examples to ease the burden on Digital Humanities professionals who might otherwise need to navigate this material independently. Thereby, we hope to encourage more practitioners to explore and apply these principles in their daily work, potentially motivating those who might otherwise be hesitant to engage with feminist methodologies. Ultimately, we hope to promote stronger ethical standards within the field, fostering a cultural shift in how data projects are approached. Regulatory bodies, such as grant agencies, could also benefit from these clearer guidelines, enabling them to set higher expectations for diversity, inclusivity and accountability in Digital Humanities work. This is especially crucial regarding the steady increase of machine learning and datafication, while ethics remain somewhat understudied and undertheorized in Digital Humanities. Cf. Berry 2022.

This article aims to bridge the implementation gap between Data Feminism’s conceptual foundation and its practical application in Digital Humanities. By contextualising its principles and providing concrete examples, we seek to lower the barriers to entry and demonstrate its relevance for everyday Digital Humanities practice. Our discussion begins with an exploration of digital hermeneutics and source criticism for data work in (Digital) Humanities and emphasizes the inherently political nature of Digital Humanities work. We then provide an overview of Data Feminism, based on the original book, but tailored to the Digital Humanities context, offering strategies for its practical implementation. The article concludes with a case study from our own research on women’s roles in historical letter networks, illustrating how Data Feminist principles can inform project design and implementation. In doing so, this article not only clarifies potential misconceptions surrounding Data Feminism but also offers practical pathways for its adoption in Digital Humanities. By making these principles more accessible and actionable, we hope to promote stronger ethical standards in the field, addressing the ethical challenges posed by the increasing integration of machine learning technologies with problematic development histories and dubious ethical standards. This work represents a step towards institutionalizing the Data Feminism principles as a universal ethical framework and toolkit for Digital Humanities, so that its transformative potential can be fully realized.

1. Digital Hermeneutics and Source Criticism for Data Work in (Digital) Humanities

Issues with bias in machine learning algorithms increasingly highlight how Humanities skills, such as historical hermeneutics and source criticism, will be central to Digital Humanities work of the future. In fact, far from rendering the more traditional Humanities – or the sub-field of Digital Humanities that Camille Roth refers to as the ›Digitized Humanities‹ Cf. Roth 2019. – obsolete, the AI revolution makes it more crucial than ever for the Digital Humanities to engage in critical data work. On data work, see Alvarado 2022. On critical code and data studies, see Iliadis / Russo 2016; Marino / Douglass 2023; Berry 2023a; Berry 2023b; Prescott 2023; Smits / Wevers 2021b. In such data work, Data Feminism can serve as a methodology that complements what is known in the Humanities as source criticism. On (digital) hermeneutics, see Edmond / Lehmann 2021; Fickers et al. 2022.

In historical disciplines, source criticism is an accepted technique to identify bias in data and ask how historical sources are ›situated‹. Cf. Haraway 1988. However, in Digital Humanities it must be applied on multiple levels and often for an entire dataset or collection rather than a single document. We must also acknowledge that we cannot simply recreate a dataset from scratch; instead, we are forced to work with what has been transmitted through the historical record. This means grappling with datasets that are inherently rife with problems in their very makeup. Any reparative measures we implement can only do so much to mitigate their harmful impact and contextualize the potentially problematic worldviews that this data perpetuates or documents. Accordingly, employing data feminist approaches to data work involves considering the data’s history in two ways.

First, the history of how the sources came into existence must be investigated, using historical-critical methods such as source criticism or hermeneutics, alongside questions about the sources’ provenance and archival history. These considerations familiar to Humanists and the hermeneutic method that undergirds our very discipline are perfectly congruent with the principles of Data Feminism: For instance, examining and challenging power (Chapters 1 and 2) and, particularly relevant for data in the Humanities, considering context (Chapter 6) (cf. D’Ignazio  / Klein 2020). Second, the history of how those sources became data must be examined. Questions arise, such as who was responsible for the digitization, what were the needs behind the data model, and whether historical-critical methods were considered during digitization and reflected in the metadata. Third, it is important to consider who created the metadata, whose interests the project serves, and what their goals were. Factors such as their societal position within a hierarchy of power and privilege, as well as the academic discipline relevant to the digitization, may, too, have influenced how it was digitized and which aspects either received special attention or were neglected. These considerations influence the affordances Cf. Norman 2013. of the resulting data, i.e. what can, cannot or should not be done with it. This may, for instance, concern questions like wether the data created is following a research-driven or curation-driven digitization paradigm (cf. Flanders / Jannidis 2015). Research-driven modeling is often highly detailed but may not translate well to larger corpora, while curation-driven modeling tends to involve only non-specialized, surface-level metadata collection. However, curation-driven digitization is designed to accommodate a wide range of research scenarios, even if the data is not very detailed or particularly accurate. Such datasets, while potentially containing incorrect, untraceable or unverified details, are easier to compare on a larger scale and are often created as part of broader digitization efforts. Nevertheless, curation-driven modeling is still not ›neutral‹ modeling because the decisions and priorities of the projects and people involved are still reflected in the data model. Examples of concrete steps to be taken to mimize their impact in a Data Feminist source criticism will be given later.

Only those in relatively powerful positions are typically able to create data on a larger scale, archive it, and decide how it is treated. Thus, as the authors of point out, data generally emerges from structures of power and reflects societal hierarchies inherently worth making explicit. Data demands even greater scrutiny when it originates from particularly problematic sources, such as those known to contain racist slurs. However, even sources that may not appear problematic, such as metadata which are commonly perceived as neutral, can be rife with issues that warrant investigation. Such issues with one’s sources may already be known or should be carefully examined during source criticism to uncover biases that may be present in unknown datasets.

This involves not only analyzing the data itself but also consulting theoretical literature that addresses power structures from the relevant historical period. As the authors of and Noble 2018 have concluded, working with data does not require less theory or signal the end of theory – it requires more.Referring to the 2008 article (Anderson 2008), the authors argue that correlation without context is not enough because it recirculates racism and sexism and perpetuates inequality (D’Ignazio / Klein 2020, pp. 171–172), citing Safiya Umoja Noble’s (Noble 2018) that showed how Google search reinforces bias beyond its correlation with sexism, racism and colonialism in our society. Contrary to Chris Anderson’s deliberately provocative claim, the authors of argue that we need more theory, not less. Without theory, survey designers and data analysts must rely on intuition, supported by ›common sense‹ ideas about the things they are measuring and modeling. This reliance on ›common sense‹ leads directly down the path to bias (D’Ignazio / Klein 2020, p. 162). Without a strong theoretical foundation, those annotating data inevitably have to fall back on ›common sense‹, where stereotypes and biases reside, which can then become inscribed into the data, reproducing or even reinforcing those biases. If theoretical principles are not properly operationalized in the data, the results may fail to reflect the object of study accurately, and resulting algorithmic models may not actually engage with the information they claim to analyze in the way they are expected to. This issue is further compounded by the historical perception of tasks such as data entry, labeling, and metadata creation as menial work, unworthy of credit or attribution. Digital Humanities scholars have repeatedly demonstrated that this is, in fact, significant Humanities work, involving hermeneutic processes that significantly shape the final output. For example, see the recent work by Alvarado 2022. However, due to this past perception, it is often impossible to trace who performed the work, preventing an investigation into the potential biases they may have introduced. Scholars reusing such data today cannot know past metadata creator’s motivations – whether they were simply completing a task, had specific agendas, or held political views that, while considered acceptable or even standard opinions at the time, are now recognized as harmful or problematic. However, it must be noted that even political views deemed unproblematic can influence a dataset and are worth acknowledging and making traceable. Throughout our work, we frequently encountered the evolving nature of terms and conventions in the context of feminism and gender-inclusive language. It is essential to use these terms with critical awareness and to provide justifications for their use, both in the digital processing of historical sources and in accompanying texts. Even versioning metadata is necessary so that the temporality of these attributions or choices of words or labels remain transparent for future users of the data. Additionally, scholars must reflect on and make transparent their own societal positions of privilege and power that inevitably influence and inform their scientific practices. Our commitment to inclusive language reflects an intentional distancing from reactionary currents in feminism. However, the limited diversity among the members of our working group and among the German Digital Humanities at large means that we can only offer a perspective rooted in critical whiteness when discussing intersectional forms of discrimination. See also Ravulo et al. 2023. For instance, imagine the major consequences on research possibilities in the seemingly minor case of manuscript metadata where scribes are always identified as men, despite there often being no evidence to support this assumption apart from received stereotypes about how the past works, possibly further shaped by the metadata creator’s personal worldview. We are grateful to Hannah Busch for this insight into her research in Digital Forensics for catalog metadata that are leveraged in machine learning applications, see Busch 2019. On inclusive metadata practices, see Mandell 2019; Mähr / Schnegg 2024. This highlights the importance of crediting all data work in modern projects to ensure transparency and traceability, which are essential to understanding the biases built into algorithms. While transparency alone does not guarantee ethical outcomes, it at least allows for explainability. On explainability in Digital Humanities, see Berry 2023a; Berry 2023b; El-Hajj et al. 2023; Ries et al. 2023. This is particularly important regarding the demands for using AI to do this seemingly ›less important‹ data work in order to save resources in Digital Humanities projects. When researchers do not know who created the data one reuses or where it originated, achieving transparency in AI ethics becomes nearly impossible.

2. Digital Humanities Work is More Political Than it May Seem at First Glance

Digital Humanities practitioners may assume that research ethics are not relevant to their work, as their subjects may be long dead. This naive approach to ethics is illustrated by Rehbein’s reflection that he rarely considered moral or legal issues in his research, believing that studying long-dead historical figures or past events had little impact on the present (Rehbein 2016, p. 631). They may equally assume that because their work is not actively political or activist, they do not require activist methodologies like feminism. However, this is a misunderstanding. In fact, every project is inherently political: Simply by selecting certain topics, materials, or sources over others, one makes a political choice, prioritizing time and resources for a particular project while others, potentially more deserving or urgent, are overlooked. This is evident in many Digital Humanities projects, where, despite the discipline’s promise of innovation, much of the work focuses on the same canonical topics and sources. This is particularly true for projects that do not involve digitizing new sources, as they often rely on pre-digitized material selected through a politics of digitization that is neither fair nor equally representative of all historical materials. Cf. Zaagsma 2023. This disparity is especially apparent in the digitization gap between the Global North and the Global South that clearly privileges data deemed relevant by Westerners who hold the power and resources to digitize and manage their digital cultural heritage. Cf. Fiormonte / del Rio Riande 2022.

As Mark Hall discusses in his 2019 contribution , the Digital Humanities are at risk of perpetuating old canons when they keep analyzing data produced following former research agendas: The first data to get digitized in high quality was frequently chosen because of the availability of high-quality scholarly editions. Cf. Hall 2019. On the discourse around dark sides of Digital Humanities, see also Smithies 2022. Former scholarly editing priorities focused on what was considered part of the exclusive canon of the supposedly best literature, including only what was deemed (Western and white) ›high culture‹ like, for instance, William Shakespeare or Johann Wolfgang von Goethe. Cf. Hall 2019. When Digital Humanities apply new methods to old sources – commonly presented as a strength, maybe even the unique sales proposal of the field – they may, in fact, reinforce old canons that many in traditional humanities have already moved beyond to look at a more diverse range of sources. In doing so, Digital Humanities research may further amplify research priorities of the past which are already overrepresented in available data.

But, of course, the problem extends beyond the issue of research priorities: The traditions, sources and collections Digital Humanities build upon are inherently shaped by hegemonic norms and structures rooted in patriarchal, colonial, racist, capitalist, and other systems of power. Digital Humanities research engaging with the material, visual, and textual cultural heritage collected by dominant groups risks perpetuating these hegemonic patterns through their data-driven research if power dynamics are not critically examined and contextualized. This issue is not limited to the design of new epistemological frameworks aimed at preventing the digital reproduction of discriminatory structures; it also includes a critical interrogation of the processes by which data and sources are accumulated. Often, collection practices are based on colonialist, Eurocentric, and cis-male viewpoints, raising important questions about how the Digital Humanities will address archival gaps that hinder the exploration of narratives beyond those centered on cis-male, white experiences. This historically rooted imbalance is evident in the fact that data concerning marginalized groups is frequently absent from archives, and when preserved, they are often filtered through the perspectives of hegemonic groups. As a result, archives simultaneously produce and perpetuate both visibilities and invisibilities, as well as expressabilities and inexpressabilities. The Digital Humanities im deutschsprachigen Raum (DHd) working group ›AG Empowerment‹ has already touched upon these issues throughout a series of panels and workshops held at DHd2023 (Trier / Belval) and DH2023 (Graz) that are documented in blogposts and the original conference abstracts: Borek et al. 2023a; Borek et al. 2023b.

This suggests that Digital Humanists need to be more self-critical and aware of our practices than we may initially think, questioning the self-image of rebellious innovators breaking away from the received confines and problematic traditional structures of academia. This sense of ›vocational awe‹ – a term originally coined by Fobazi Ettarh and applied to Digital Humanities by Melissa Terras – can obscure the field’s internal challenges. Cf. Ettarh 2018; Borek et al. 2023c. These include domination by privileged white Westerners, precarious employment conditions, and gender imbalances. ›Vocational awe‹ describes a profession’s uncritical belief in its own inherent goodness, which can blind its practitioners to the field’s flaws – a phenomenon that D’Ignazio and Klein describe as a ›privilege hazard‹ in the context of Data Feminism. Cf. D’Ignazio / Klein 2020, pp. 28–29. While not all of Digital Humanities practitioners may be in positions of power, many Digital Humanities practitioners are in positions of privilege. This comes with the responsibility to use it well.

3. Data Feminism for the Digital Humanities

But how can Data Feminist approaches be operationalized within Digital Humanities? What research areas could serve as practical entry points for those wishing to apply Data Feminism methods? Addressing the challenges posed by Data Feminism requires the Digital Humanities community to build structures that critically confront biases and imbalances. But we are not empty-handed in dealing with these challenges: Lessons can be drawn from various projects, which offer valuable insights for opening up spaces for dialogue within the field. The ›Full Stack‹ Feminism project, for example, aims to create a toolkit that integrates feminist methods and design principles to promote more socially conscious technologies and infrastructures in Digital Humanities: This involves applying feminist design principles, decentering traditional voices, and incorporating intersectional feminist methodologies into Digital Humanities projects. Cf. Liu 2020; Webb et al. 2023; Webb 2023; Webb / Fox 2022. In addition, there is an extensive body of research that can be drawn on for applying Data Feminist principles to the Digital and Computational Humanities. Beyond the foundational work on feminism and intersectionality in Digital Humanities Cf. Risam 2015; Wernimont 2015; Losh / Wernimont 2018; Bordalejo / Risam 2019; Block 2020; Wiens et al. 2020; Smyth et al. 2020; Earhart 2022; Gao et al. 2022., numerous studies directly engage with Data Feminism. Cf. D’Ignazio / Klein 2020; Lang et al. 2023; Rezai 2022; Juen 2021; Keck 2021; Borek et al. 2023a; Borek et al. 2023b. However, much of the existing Digital Humanities work focuses on recovering the contributions of historically overlooked women. Cf. Buurma / Heffernan 2018; Aleksander 2014; Hall 2019; Dang 2020; Dickel et al., n.d.; Wreyford / Cobb 2017; Bui et al. 2021; Keck 2021. This suggests that Digital Humanities has not yet fully embraced the broader applicability of Data Feminism to diverse areas of data work beyond these specific, though important cases. Data Feminism is organized into principles that can be applied at various stages of the research process. The following overview is organized according to the chapters of the original book, , which are themselves structured around these principles. Cf. D’Ignazio / Klein 2020. Somewhat confusingly, the chapters each have a title and subtitle, which can interchangeably be used as references.

4. A Data Feminist Approach to Historical Letter Network Data

To apply the information discussed previously, we present an example of Data Feminism in practice from our own research: Applying Data Feminism as a framework to women’s networks in the correspondence of Early Romanticism. This PhD project is situated in Digital Humanities at Philipps-Universität Marburg and has the working title Epistolares Schweigen oder hidden figures? Kreatives Kapital, Rollenzuweisungen und Funktion weiblicher Korrespondenz im Briefnetzwerk der Frühromantiker*innen. For more information about the PhD project, see here. Furthermore, the PhD project is part of the DFG project ›Korrespondenzen der Frühromantik. Edition, Annotation, Netzwerkforschung‹, see: https://briefe-der-romantik.de/. The enrichment of data with LOD was part of the work in the project group based in the Academy of Science and Literature Mainz together with Aline Deicke and Clara Seibold. The study focuses on a quantitative analysis of women’s roles and functions within the letter network and, among other aims, explores whether women’s often overlooked creative potential can be formalized through historical network analysis. The letters as data are collected in context of the ongoing project ›Correspondence of Early Romanticism‹, which compiles and analyzes letters exchanged between Romantic authors from 1790 to 1802 by reusing existing editions, supplementing them with original manuscripts, and adding inferred letters to fill gaps. In the following, we discuss lessons learned from this work.

Women and individuals beyond the cis-male category are particularly affected by the limitations of historical archival practices. In many collections of cultural heritage, gender is not documented at all or when it is, not as a self-identification but is instead externally imposed. But this gender data gap can distort both the interpretability and validity of historical research findings. On the gender data gap, see Criado-Perez 2020; Jahnke et al. 2024; Neuber et al. 2024. The term gender data gap was popularized through Criado-Perez 2020. However, we distance ourselves from her binary conception of gender. In the context of the Romantic period, dominant historical narratives, which are still prevalent in current scholarship, often reflect patriarchal assumptions. Women are frequently represented as muses, assistants, or passive observers, rather than as autonomous individuals with their own agency or as creators. For instance, the (Kyvernitou / Bikakis 2017) ontology supports making such gendered representations explicit. Our project takes a critical stance on these inherited perspectives, framing such female perspectives within the concepts of care work and mental load. Acknowledging these dimensions is essential for a more comprehensive understanding of women’s experiences and contextualizing their roles in historical networks of correspondence.

In order to do so, our work explores how digital methodologies can be used to make visible the structural absences and biases in historical datasets. Transparency in data creation and annotation, as well as an active engagement with the epistemological implications of missing data, are essential to this approach. Gender-sensitive representation, though only one aspect of appropriate data modeling, requires careful attention. Retrofitting historical classification systems to accommodate current understandings of gender identity is not straightforward and involves ethical and methodological challenges. Widely used standards such as the PICA acquisition schema, TEI-XML, or gender modeling in Wikidata typically do not accommodate fluid gender identities, for a long time did not distinguish between sex and gender, and rarely account for temporality or uncertainty. In most cases, these models enforce a binary framework that contradicts both current theories of gender and, crucially, the complexities in the lived reality of historical individuals. Cf. Illmer 2022. Gender is shaped by the cultural and institutional norms of a given time period and therefore inherently possesses a temporal dimension. However, digital encoding systems, if they allow for differentiated gender encoding at all, rarely account for this temporality. They often lack mechanisms to record the temporal context of the attribution, as well as the context in which the metadata was created, including underlying assumptions and the definitions of terms at that time. Since terms used to describe marginalized groups frequently shift in meaning, sometimes becoming slurs or being reclaimed, it is essential to include contextual information for both the historical period being described and the time when the metadata is added, along with the cultural situatedness of the metadata creator. Moreover, gender attributions are subject to change, which is why it has been proposed to model gender as an event to account for its temporality. Cf. Andrews et al. 2024. A further complicating factor is the difficulty of reconstructing gender identity from historical records, particularly where the documentation itself is sparse or biased. For example, in the case of Anastasius Lagrantinus Rosenstengel (1687–1721), there is scholarly debate as to whether this individual was a lesbian woman living as a man or a trans man, which is not easily resolved with available records. Cf. Steidele 2021; Almstedt 2024. In most historical contexts, individuals were assigned a binary gender in line with cultural expectations, which was reflected in their gendered first names, and only few historical sources document forms of non-binary identification. On the history of non-binary gender: DeVun 2021. On queer voices in Digital Humanities, see Webb 2022. Our work does not claim to resolve these complexities, but it seeks to identify patterns of gendered roles within the constraints of and social frameworks represented in our source data.

However, such sophisticated forms of gender encoding remain rare and often elusive in the Digital Humanities. This also applies to the underlying data of our case study: The scholarly editions our project builds upon seldom annotate gender as a research category. Cf. Scott 1986. To make gender analyzable as part of a dataset audit examining the gender data gap, the dataset was semi-automatically enriched with gender information.A dedicated study by the authors of this working paper introducing dataset audits for mitigating data gaps in historical data is forthcoming (Lang / Suárez Cronauer 2026, forthcoming). While we acknowledge that automatically assigning gender based on first names is highly problematic, and that modeling gender involves complexities the Digital Humanities have only begun to address, our case study demonstrates that even a simple dataset audit can expose data gaps within authority control records and, by extension, other reconciled data sources. This was annotated as ›assigned_gender‹ to acknowledge the provisional and externally attributed nature of these classifications. This facilitates future reinterpretations or corrections and provides transparency for subsequent researchers. In this context, it must be noted that the act of classifying, as discussed in Data Feminism’s Principle 4 (›Rethink Binaries‹), is inherently contentious and thus, must be critically assessed in each instance to determine whether it causes more harm than benefit. However, stopping to classify is not an option either: Our data exists within pre-defined systems that we reuse more than we redefine them and it has already been classified in various ways. The digital ecosystem as a whole is built on classification and categorization. Therefore, it may be more constructive to leverage existing classifications and make them our objects of study, recognising that they do not represent reality accurately or holistically. Rather, they reflect a dominant view of reality that privileges a narrow range of perspectives over a plurality of voices. As the well-known saying goes, All models are wrong, but some are useful (George Box). Studying our models can help us learn about the world through the discrepancies between these representations and the realities they are meant to encode. For a great example of using digital methods to scrutinize skews in the system of Digital Humanities, see also Eichmann-Kalwara et al. 2018. On learning about reality through a process of iterative modeling, see McCarty 2003, p. 1232: Research in humanities computing begins then, in the breakdown, when tools become models. It proceeds in an iterative cycle of constructing, testing, analyzing, and reconstructing these models in order to discover how the thing imitated maybetter be known.

The initial enrichment process involved matching individuals with Gemeinsame Normdatei (GND) authority record identifiers and importing gender information using OpenRefine. For individuals lacking a GND-ID or gender data within the GND, the Python library ›gender-guesser‹ was used to infer gender from first names. The enrichment of the database with external information was done within the project’s Digital Humanities group, including the work of Aline Deicke and Clara Seibold. Because gender assignment is inherently problematic, this step was followed by manual verification. Importantly, automated classification tools, and indeed most digital tools, tend to perform better on data from the centre or mainstream of the data they were trained or built on. This means that minority groups, often only represented as sparse outliers in data, are less likely to be classified accurately, placing scholars in a double bind: On the one hand, such tools can ease the burden of filling in large amounts of missing data; on the other, the groups most affected by these gaps are also the most likely to be categorized incorrectly. There are several possible approaches to this problem. One is to do the work manually, as some activists choose to do, considering the extra time invested is a meaningful way to honour those lost voices as part of a curatorial ethics of care (Caswell / Cifor 2016). For more information see, for example, recent work on counting feminicide D’Ignazio 2024. Another is to use the tool but follow it by manual verification or even tool criticism. We chose the latter approach of using tools to speed up processes and make this important work feasible within project time constraints, while following up with careful manual verification. Building on the concept of strategic essentialism (cf. Eide 2016), Suárez Val 2023 has conceptualized this as strategic datafication. The method is neither definitive nor neutral, but it serves as a starting point to interrogate what can be known within the limits of existing authority records and software. As Dominique Schirmer notes, there is an inherent tension between the ›requirements of reproduction‹ and ›dangers of reproduction‹ when gender must be inferred to enable analysis.Cf. Schirmer 2023. Without such annotation, gender cannot be operationalized as a research variable. Yet the act of assigning it introduces new uncertainties. To acknowledge the limitations and challenges of gender assignments made by third parties, automated or not, and to indicate their provisional nature, we recorded this as ›assigned gender‹. This facilitates future reinterpretations or corrections and provides transparency for subsequent researchers. While our database remains a work in progress and is not yet complete, these preliminary figures offer insights into the persistent challenges associated with historical research data, particularly in terms of gender representation.

The enrichment of data with gender information and the using of gender as an analytical category enables more complex, diverse perspectives on datasets. With ›assigned_gender‹ added and embedded as information in data about persons in historical sources, we are able to address and filter marginalized groups like women in corpora in which they otherwise would be hidden from quantitative historical research. Thereby, we are able to ask new research questions and open the field of quantitative historical research for debates on gender aspects. In the presented use case, for example, we find a significant disparity in male and female gender data, which raises critical questions about the historical sources underpinning the database. The imbalance reflects both the archival loss of women’s writings and the biases of earlier collection practices: Research, such as Wernli 2022, indicates that many letters authored by women in the early 19th century were lost, destroyed, or only preserved through editorial frameworks that reduced women to gender roles defined by patriarchal norms. Whether as housewives admiring their spouses, assistants supporting their husbands, or muses serving as passive inspiration for male authors, these gender roles reflect persistent patterns of positioning women in relation to male figures, rather than as independent agents. This historical imbalance not only perpetuates gender disparities in datasets built upon these sources, it also shapes how datasets are constructed for data availability reasons. These figures exemplify the gender data gap within historical authority records and its implications for digital historical research.

While our method of assigning gender based on first names is limited, its utility lies in enabling an initial audit of the dataset. This implements a key principle of Data Feminism: The value of making absences visible. Even a preliminary analysis can expose the effects of data gaps, prompting further investigation and opening up space for new research questions to be asked. For instance, which roles are systematically attributed to male versus female correspondents? How does gender distribution shape access to particular topics or social functions within the letter network?

Embedding ›assigned_gender‹ into our dataset allows us to extract and examine the participation of historically marginalized groups, in this case women, in a corpus where they might otherwise remain invisible. This in turn facilitates the integration of gender history and feminist critique into quantitative historical research. While this approach does not yet account for non-binary identities or fluid gender expressions and may be considered crude in some respects, it lays a foundation for future work that explicitly addresses gender distributions in our corpus, a possibility the original state of the data did not afford. This serves as a good example of how projects can make important contributions by providing structured data on issues such as gender representation. Although the measure is relatively simple, it still requires reflective depth to be implemented responsibly within the limitations of digital systems that inherently classify. It also requires proper documentation, such as through the use of datasheets, to help future users understand how the dataset was created and what kinds of research questions it can or cannot support.

5. Conclusion: Making Data Feminism Work in Practice

Data Feminism, as conceptualized by Catherine D’Ignazio and Lauren F. Klein, represents a strand of intersectional feminism that critically examines the predominance of white, cis-male perspectives in data science. In it, the authors have attempted to operationalize existing feminist strategies to be used in data science and AI contexts by defining seven key principles to guide such work. Data feminist critique extends beyond data analysis to encompass the modeling, collection, curation, and presentation of datasets, even the development of algorithms and digital infrastructures.

Data Feminist approaches hold significant potential for advancing more ethical research within the Digital Humanities; however, they have not yet been fully integrated in the German-speaking Digital Humanities community. A notable implementation gap exists in the development of concrete guidelines and frameworks for embedding Data Feminist principles into everyday Digital Humanities project practices. Addressing the challenges posed by Data Feminism requires the Digital Humanities community to build structures and establish best practices that critically confront biases and imbalances. Fortunately, there is much work that the Digital Humanities community can draw on. We have developed this guide to provide essential information, offering an accessible entry point and outlining strategies specific to Digital Humanities for implementing ethically responsible scholarship. We hope that we have not only provided an accessible introduction to the Data Feminism principles for Digital Humanists through providing a Digital Humanities implementation example, but also offered an entry point into the available secondary literature, opening up the rich body of sources from the Digital Humanities and related fields for those interested in beginning to work with and explore these ideas further.

To make Data Feminism effective in Digital Humanities, it is essential to engage seriously with its principles rather than using them as a form of pinkwashing.This requires integrating more theory and methods from specific academic disciplines into Digital Humanities research, while avoiding a simplistic return to traditional humanism, which Bianco has termed ›retro-humanism‹ (cf. Bianco 2012). Risam 2015, picking up the term, argues for a stronger relationship between theory and praxis in the Digital Humanities by applying developments like cultural studies, feminism, postcolonial studies, critical race studies, or queer studies to them. Furthermore, the time and resources needed to enrich data and fully implement ethical approaches to data in practice should not be underestimated. Lastly, raising awareness of Data Feminism by incorporating it into teaching and academic curricula is crucial to ensuring its influence on future scholarship and practice.

As we have argued and the authors themselves have pointed out, the term Data Feminism leads many to the misconception that the framework is relevant only to feminism and feminists. We hope to have dispelled the common misconception that Data Feminism is solely about women or activist issues. Data Feminism offers feminist methodologies to address problems that affect everyone working with (historical) data. We hope to have convinced the Digital Humanities community of its broad usability and applicability, particularly given its strong alignment with established traditional historical methodologies like source criticism, to which it can be seen as an extension for the digital sphere. In doing so, we have presented Data Feminism as a universally relevant framework for conducting ethically sound data work.

In this context, one may wonder whether this truly fulfills the aspiration stated in the introduction: to illustrate how Data Feminist principles can inform project design and implementation. Are such seemingly simple steps enough to make data or projects feminist (irrespective of what people may typically imagine when they encounter the term)? Feminism means grappling with hidden power structures, so any project that engages with these structures could reasonably be considered feminist to some extent, even if it does not focus on topics typically associated with feminism, such as gender studies. Entire fields, such as critical archival studies, could likewise be understood as feminist under this broader definition. Is this ›genuinely Data Feminism‹ or simply quality assurance and good scientific practice? This is the false binary we have aimed to debunk in this article. Data Feminism is one method within the broader toolkit of quality assurance in research and should be viewed as a way to enhance good scientific practice. It is a framework that more scholars should integrate into their everyday research. Scholarship that does not account for a plurality of perspectives – even, and especially, when that plurality is lacking in the data – is not good scholarship. Research that fails to acknowledge the role of power structures in shaping datasets and historical records during analysis and interpretation must be considered significantly lacking in quality.

Can what we have provided truly be considered a how-to guide to Data Feminism in the Digital Humanities? Perhaps not, as the presented case study is only one small example of how Data Feminism can be applied, inspired by the many other examples discussed in relation to the Data Feminism principles. Our implementation example is not a definitive solution but rather a call to action: an invitation for others to engage in similar work and contribute their own approaches. It represents a first step toward making a dataset ready to even begin asking critical questions about issues such as the gender data gap.

Were we able to offer an easy how-to for fully applying the principles? Not exactly. Like the principles of research ethics and good scientific practice, applying Data Feminism requires interpretation and adaptation to one’s specific context. One researcher’s dataset may already include gender assignments but may lack other forms of representation. Data Feminism encourages us to scrutinize our data and tools holistically, making us more aware of potential pitfalls and inspiring us to take active steps to address them. In this spirit, we have aimed to bridge the implementation gap between the manifesto and the daily working practices of Digital Humanists and more traditional historians. We hope that with these guidelines in hand, others will find it easier to identify which steps, whether simple or more complex, can help make data work in the Digital Humanities more inclusive, ethical, and rigorous.