2.1. Archives

Our journey begins by addressing archives from the perspective of archival studies. Archives are generally understood as institutional bodies that collect, preserve and provide access to collections of records of historical interest (Prescott and Wiggins Reference Prescott and Wiggins2023). They undergird academic, governmental and community practices of collective memory construction. As Millar (Reference Millar2017) summarizes, “[a]rchives are tools that people use to look beyond the present moment and understand the wider context of a family, community or society.” Technical problems in archives can be separated into three processing stages for our purposes. To simplify somewhat, these stages involve (i) acquisition and appraisal; (ii) processing of collections, specifically the description of records and the production of metadata; and (iii) provision of access to a collection through interfaces (Moss and Gollins Reference Moss and Gollins2017). Appraisal entails evaluating whether a given document or set of documents should be included in an archive. Description, meanwhile, involves metadata creation and assessing materials for a number of attributes. Archival interfaces entail systems – useful for both archivists and archive users – for interacting with a given collection.

Modern archival practice originated in the state archives that emerged out of the creation and rationalization of various state bureaucracies during the 19th century (Cook Reference Cook2013). The two most influential principles from this period, which continue to shape archival practice, are original order (the maintenance of organizational schemes developed while records were in active use) and respect des fonds (the maintenance of records grouped according to their originator). These principles produce an enduring normative assumption in archival work: reorganization of records should be minimized (Nesmith Reference Nesmith2002; Yeo Reference Yeo2012). At the same time, archives are also theorized as sites at which collective memory is constructed and renegotiated (Schwartz and Cook Reference Schwartz and Cook2002). Contemporary archives contend with a fundamental tension between the deeply interpretive work that characterizes historical meaning-making and the seemingly infinite stream of documents produced by modern industrialized life. Understanding the implications of applying NLP to archives requires an account of this conflict, driven by the underlying transparency-, context-, resource- and maintenance-related concerns mentioned above.

Archival work in the 21st century is characterized by additional theoretical and practical desiderata that archivists may seek to balance with these principles. As described by Cook (Reference Cook2013), the move away from an early focus on preserving all evidence of organizational activity was precipitated by the corresponding bureaucracies’ growth and the novel requirement in the early 20th century that archives choose which documents to preserve. The fundamental challenge archives are faced with now is the many, many documents that the 21st-century bureaucracy or online platform produces (Thibodeau Reference Thibodeau2009). Newer frameworks engage directly with this problem of scale. MPLP, or “More Product, Less Process” (Greene and Meissner Reference Greene and Meissner2005), is an approach to managing archival backlogs that prioritizes making materials available in great volume, in contrast with other goals like meticulous description. The applications discussed in this paper are motivated by the immense challenges of handling very large collections with limited resources. We note that some of these challenges are due to problems beyond the scope of the review, such as labor shortages due to poor pay (Bredbenner et al. Reference Stephanie, Fulmer, Oliveira-abbey and Rinn2024), funding contingent on the political environment (Harper et al. Reference Harper, Burr and Blanton2024), reasonable distrust in computational methods (Jaillant and Rees Reference Jaillant and Rees2023), and the difficulty of digitization (Archives 2016; Tudico Reference Tudico2022).Footnote ¹

Based on our review criteria, most papers we discuss focus on government archives, likely because they have the resources needed to explore new ways of processing documents. National archives in particular are subject to both the voluminous digital exhaust of bureaucracy and culturally subjective notions of national identity and historical importance (Martínez-Cardama and Pacios Reference Martínez-Cardama and Pacios2022). While such institutions are not necessarily representative of typical archives (which run the gamut from small private collections to massive corporate holdings), the techniques we discuss nevertheless have wide relevance – especially as the technology involved becomes more mature.

2.2. Natural language processing

NLP broadly refers to the use of computation to understand – segment, label, reorganize, aggregate and interpret – natural language. The field can be understood as an outgrowth of applied machine learning and computational linguistics (Eisenstein Reference Eisenstein2019).Footnote ² Modern NLP is itself grounded in large-scale data curation efforts, from the Penn Treebank (Marcus Reference Marcus1993), a dataset of Wall Street Journal articles with part-of-speech labels and other syntactic information numbering in the millions of words, to Common Crawl (Common Crawl – CC-MAIN-2025-08 2025), a collection of billions of scraped web pages that is the foundation for many LLM pretraining projects (Baack Reference Baack2024). With respect to archives, three practices emerge as particularly relevant: supervised text classification, unsupervised representation learning and information retrieval. Supervised text classification entails training a model to predict labels (or classes) from text, typically based on human-labeled training data, and has natural applications to archival tasks like sensitivity review (Baron et al. Reference Baron, Sayed and Oard2022), which entail sorting records into predefined categories. Unsupervised approaches, such as topic modeling (Blei et al. Reference Blei, Ng and Jordan2003), produce interpretable vector representations of documents or collections, often based on internal patterns like the co-occurrence of words. Although information retrieval as a field extends beyond NLP, many of its representational techniques originate in language technologies. In particular, search is an umbrella category with functions that range from finding exact terms in documents to using contextual representations of text to retrieve semantically similar documents (Jurafsky and Martin Reference Daniel and Martin2023).

While these three families of approaches (as well as other NLP tasks) have historically been pursued somewhat separately, contemporary NLP involves work that is premised on more task-agnostic models – a reality made possible first by more sophisticated word embedding models like Word2Vec (Mikolov et al. Reference Mikolov, Chen, Corrado and Dean2013) and then by the transformer architecture (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, Polosukhin, Guyon, Luxburg and Bengio2017) and its use in LLMs (Raffel et al. Reference Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li and Liu2020). This type of model was popularized in the late 2010s in the form of BERT (Devlin et al. Reference Devlin, Chang, Lee, Toutanova, Burstein, Doran and Solorio2019) and GPT-2 (Radford et al. Reference Radford, Jeffrey, Child, Luan, Amodei and Sutskever2019) and is usually trained to predict the next token given some context using massive text datasets.Footnote ³ While the task of predicting the next token is not necessarily of interest in itself (except insofar as it allows for generating text), the resulting models have great value as encoders. That is, they transform text into representations that can be used for many potential applications. In addition, many models exhibit the ability to perform a wide range of tasks out of the box without additional training data (so-called zero-shot performance), although rigorous evaluation of many of these capabilities remains difficult given their context-dependent and interpretive nature. Archival processing tasks in particular may be difficult to represent through a traditional NLP evaluation method like benchmarking (Raji et al. Reference Raji, Denton, Bender, Hanna, Paullada, Vanschoren and Yeung2021).

Contemporary transformer-based NLP systems are technically capable of incorporating much more additional context, but in practice still only consider limited knowledge of the user (Soni et al. Reference Soni, Schwartz, Sedoc, Balasubramanian, Duh, Gomez and Bethard2024). They tend to be data- and energy-hungry (Hedderich et al. Reference Hedderich, Lange, Adel, Strötgen and Klakow2021; Strubell et al. Reference Strubell, Ganesh, McCallum, Korhonen, Traum and Màrquez2019), opaque (Fanconi et al. Reference Fanconi, Vandenhirtz, Husmann, Vogt, Bouamor, Pino and Bali2023) and difficult to repair or update (De Cao et al. Reference De Cao, Aziz, Titov, Moens, Huang, Specia and Yih2021). These qualities are at odds with the nature of archival infrastructure – context-driven, sometimes lacking in resources, ideally relatively transparent and composed of many interlocking schemas and standards.Footnote ⁴ This mismatch means that NLP’s great promise for managing large archives meets some evident obstacles, although these tensions can be negotiated through both technical strategies within NLP and creative use of NLP by archives.

2.3. Review methodology

To scope this review, we focused on published academic work, located both within the field of archives and in the NLP literature. An initial set of papers from NLP was found by searching the Association for Computational Linguistics Anthology (aclanthology.org) for the word “archive.” Additional related papers were drawn from Association for Computing Machinery venues such as the Journal of Computing and Cultural Heritage. We also searched The American Archivist, Archivaria and Archival Science for “NLP” and “natural language processing.” Most of the application papers found through this process are case studies. Because the application of generative AI to archives is still nascent, we did not carry out a full systematic review; we instead highlight recent work for the purpose of discussing emerging trends.

3.1. Appraisal

The first stages of a document’s life in an archive consist of acquisition and appraisal of whether the document should be included in the archive. Some archives take a positive approach of seeking out materials, while others instead define a stream by which they come to own documents (Schmidt Reference Schmidt2011). In many contexts, however, archives’ central goal of preservation is complicated by the endlessly proliferating streams of text which characterize the modern world. Resource constraints and massive backlogs can additionally limit archives’ processing capacities, thus making automated appraisal desirable. Automated processes have particular utility for managing acquisition of born-digital documents, which are usually much more numerous than a typical analog collection (Jaillant Reference Jaillant2022).

The acquisition process in a given archive is influenced by its goals (Cook Reference Cook2011). An archive that collects all documents produced by an organization need not make appraisal decisions based on content; modern appraisal originates in the 20th century, when the volume of documents produced by state and corporate bureaucracies massively increased (Buckland Reference Buckland2017a; McKemmish et al. Reference McKemmish, Piggott, Reed and Upward2005). Heuristics have been used to constrain the amount of documents ingested into state archives. For example, one type of policy for appraising personal communication entails selection based on the seniority of the sender or recipient, such as the U.S. National Archives’ CAPSTONE procedure (Jaillant and Rees Reference Jaillant and Rees2023). This standard necessarily results in an archive stratified according to the power of candidate documents’ authors; it also produces series that chiefly correspond to author identity. Rules of this kind can also have legal ramifications, e.g., if messages with evidentiary content are not recognized as such and then not acquisitioned into an archival collection (Johnston et al. Reference Johnston, Wallace and Punzalan2022; Shilling Reference Shilling2006). NLP offers the potential to more flexibly make similar kinds of automated decisions, but nevertheless involves additional challenges.

In an inventory of prior work on computer-assisted appraisal, C. A. Lee (Reference Lee2018) describes a dominant paradigm of (1) using software to filter digital documents via heuristics in metadata, such as time of creation, and (2) manually choosing messages based on content. This is grounded in concerns about transparency and context sensitivity; complete automation of appraisal is not preferred because it is opaque, and it lacks the contextual interpretative capabilities of a human archivist. Prefiguring other papers in this review, Lee’s article focuses on digital records generated by large organizations, like governments. Digitization is costly. Thus, digitally automated appraisal is only reasonable for documents that are either born-digital or made readable through digitization and optical character recognition (OCR) prior to an archive attempting to collect them.

NLP can be used to automate complex and subjective judgments about the content of documents, so long as there is available data representing analogous judgments. A major content-based desideratum for many archives is some notion of historical importance (Cox and Samuels Reference Cox and Samuels1988). The possibility of using automated text classification for communication archives in particular, given an operationalization of this notion, has attracted attention from multiple domains (Schneider et al. Reference Schneider, Adams, DeBauche, Echols, McKean, Moran and Waugh2019; Vellino and Alberts Reference Vellino and Alberts2016). For example, Risi et al. (Reference Risi, Sharma, Shah, Connelly and Watts2019) attempt to assess how well future historical relevance can be predicted contemporaneously using about two million cables sent within the U.S. Department of State from 1973 to 1979. A very small subset of these documents was later included in a collection constructed by State Department historians, Foreign Relations of the United States (FRUS), that reflects pivotal moments in U.S. foreign policy. Because it is more carefully prepared for consumption by researchers than the rest of the cables, meaning its content takes on greater authority in the context of computational studies that rely on clean data; as Jo (Reference Eun Seo2020) shows, studying FRUS at scale with computational techniques evinces major shifts in the ideologies of American diplomats. The theoretical problems Risi et al. enumerate range from the relationship between an event’s historical relevance and events that have not happened yet to a lack of clarity about the relationship between discrete events and documents. Using document representations from a topic model as input to a classifier, they achieve an F1 score of 0.86 on a balanced test set of FRUS and non-FRUS cables – an impressive score, but likely not usable in a real archival system.Footnote ⁵ They find that historical importance is not easily predicted by contemporaneous assessments, casting doubt on the feasibility of using that information in an automated fashion.Footnote ⁶

Venkata et al. (Reference Venkata, Young, Bell and Green2022) are similarly concerned with government documents, albeit at the British National Archives. In 2018, the volume of digital materials belonging to the British government was estimated at “more than 16 billion emails and 3 billion documents” (Ranade Reference Ranade2021). Venkata et al. conduct a survey of open-source or commercially available AI-based methods for appraising this massive number of documents. They are interested in training classifiers using prior appraisal decisions; these appraisal decisions are themselves grounded in the notion of “enduring value” to the public (although the operationalization of this notion varies). In some settings, documents are classified as retained or not; in others, they are assigned a specific “retention schedule” score corresponding to a degree of default postprocessing. Based on their experiments, the authors observe that no system is satisfactory on its own – i.e., using them would require a human in the loop.

Although NLP models typically classify each document individually, Rolan et al. (Reference Rolan, Humphries, Jeffrey, Samaras, Antsoupova and Stuart2019) question whether the historical relevance of a document can be predicted in isolation. An item-level approach to appraisal misses the contextual importance of documents that gain their meaning from others. Work in government archives also speaks to the additional labor required to use a model within a preexisting standard. As we discuss in Section, descriptive frameworks can be both difficult to operationalize and specific to a given archive. While NLP seems potentially well suited to assisting with the automation of appraisal decisions, the literature indicates mixed results. The core problem that Shabou et al. (Reference Shabou, Tièche, Knafou and Gaudinat2020) identify is in producing a “defensible” method which aligns with archival norms. Having experimented with automation in a Swiss government archive, the authors conclude that automating appraisal requires “prioritizing the implementation of metrics with high automation potential” – an approach which may prejudice archives toward the use of simpler and less seemingly subjective variables in deciding which documents are worthy. Shabou et al. ultimately produce a proof-of-concept tool for manual appraisal which incorporates automated scores related to a given document (e.g., topic) without giving decision-making wholly over to the machine.

Notably, many archives’ initial processing includes some form of privacy review (Yoxall Reference Yoxall1984). This entails the review of records for information that may be inappropriate for public access; it is frequently discussed in the context of the Freedom of Information Act, which mandates the release of government information that is not sensitive when requested (Windon and Youngblood Reference Windon, Youngblood, Lacity and Coon2024). Technical paradigms discussed above, and supervised classification in particular, have been used to partially or fully automate this process (Baron et al. Reference Baron, Sayed and Oard2022; Giannella et al. Reference Giannella, Branting, Van Guilder and Baron2024; McDonald Reference McDonald2019; McDonald et al. Reference McDonald, Macdonald, Ounis and Gollins2014). Even journalistic archives face related problems with recognizing the copyright status of poorly documented records (Taurino et al. Reference Giulia, Sweeney, Facklam and Smith2025). In general, detection of sensitive information is an exemplary problem in that it is culturally contingent, dependent on context (Nissenbaum Reference Nissenbaum2011) and sometimes in tension with other archival desiderata.

3.2. Description

In addition to classification decisions, NLP also lends itself to the production of metadata. In many archives, the primary interface through which materials are presented to users is the finding aid. This is a document that describes the contents of a particular collection, series or even individual record (Chung and Irwin Reference Chung and Irwin2017). It is used both as a proxy for the relevant material and as a way of providing context, usually in the form of narrative text (Wiedeman Reference Wiedeman2019). Usability, satisfaction of formal requirements (such as those of common formats, like the Encoded Archival Description XML standard), and the creation of (contextually) helpful unstructured descriptions are all important desiderata (Berner and Haller Reference Berner and Haller1984; Cox Reference Cox2008). Similarly, metadata plays a role in the representational schemes discussed below – it can be a mechanism for users to find records – and in description (Buckland Reference Buckland2017b).

In existing work using NLP, we find more discussion of the creation of new types of metadata than of integration with preexisting schema. One straightforward way to extract descriptive information from archival documents is named entity recognition (NER), the labeling of sequences in a text as referring to a certain person, place or thing with a proper name (Daniel and Martin Reference Daniel and Martin2023). As described by Ehrmann et al. (Reference Ehrmann, Hamdi, Pontes, Romanello and Doucet2023), a survey of NER in archives, the ability to search a collection for named entities is desirable for many users whose historical work may be focused on a set of individuals or institutions. As Schneider and Chan (Reference Schneider and Chan2016) put it, NER is a natural tool for addressing two major goals of archivists: “screening for confidential, restricted, or legally protected information … and preparing the files for use by researchers, including by efficiently optimizing discovery and access to the material’s intellectual content.” Although NER models are intended to generalize, they are likely to perform best on entities included in their training data (Agarwal et al. Reference Agarwal, Yang, Wallace and Nenkova2021). As such, performance in a novel archival setting might be worse without additional effort put into labeling additional documents and retraining an NER model.

Production of useful metadata for multimodal and non-digitized archives is another popular application of NLP. Non-digitized archives with preexisting finding aids still contain text on which one can run NER. This is the case in Wright et al. (Reference Wright, Evans and Green2023), where a search engine was built for an archaeological archive using NER; their finding aids already documented location, time period, etc. More radically, non-digitized text archives can be made accessible with computer vision models designed to recover a specific part of a document’s textual content, such as the U.S. Census race label recognizer described in Jansen and Marciano (Reference Jansen and Marciano2025) for research on Japanese internment or the card classifier used in B. C. G. Lee (Reference Lee2018a) to retrieve death certificates in a large Holocaust archive (both of which do not require OCR). Archives with audio data also tend to have textual finding aids which are possible inputs to NLP-based interfaces themselves; however, full description of the materials may require transcription of that audio.

In contrast to the content-oriented focus above, Davet et al. (Reference Davet, Hamidzadeh and Franks2023) explored the inclusion of paradata – here, the byproducts of automated archival pipelines – in metadata. Emphasizing the different questions that different users might have, they recommend including information about the data used to train and validate any models used, as well as their version and performance on representative data. The quirks of a system can have a significant impact on derived metadata; indeed, Liu et al. (Reference Liu, Heitman, Soh and Whiteley2025) creatively use the failings of an NLP tool – Google’s language identification application programming interface (API), which struggles with indigenous languages – to find records of indigenous culture in colonial archives. Given that the content of training data may have unpredictable impacts on a model’s application to a different domain (Elsahar and Gallé Reference Elsahar, Gallé, Inui, Jiang, Ng and Wan2019), providing details about that data also constitutes an attempt to provide context. Notably, this concern about a lack of transparency in archival description also applies to completely analog archives; as Douglas (Reference Douglas2016) argues, a collection is influenced by archiving practices and “interested parties” (e.g., the subject of a personal collection) before it is even ingested into an archive.

3.3. Interfaces

The question of how one should order and present archival collections is one of the central focuses of archival theory – as the paradigmatic principle of original order ensures. The local context of materials impacts their meaning, whether on a web page or in a physical space. For example, a diary may be grouped under materials related to a certain marginalized community, rather than as part of a personal collection. This means the diary will necessarily be interpreted differently – likely as an exemplar of the marginalized group’s history, rather than as a document of an individual experience (Caswell et al. Reference Caswell, Migoni, Geraci and Cifor2017; Gabiola et al. Reference Gabiola, Brilmyer, Caswell and Zavala2022). The historiographic responsibility of archivists first detailed a century ago and its more recent evolution into an admission that archivists will always be implicitly supporting one interpretive frame (Cook Reference Cook2013) has significant implications for computational processes that present archival content to users (Trace Reference Trace2021). What seems to be the appropriate context for a document is always itself changing in response to shifts in the normative belief systems of the surrounding community (Caswell et al. Reference Caswell, Migoni, Geraci and Cifor2017). In response to the historiographical imperative and increasing recognition of the community impact of archives, many archives have made efforts to make their collections more accessible (Caswell et al. Reference Caswell, Cifor and Ramirez2016). A core element of this effort has involved implementing search engines and other exploratory tools that support traversal of collections’ contents. This subject already has a great deal of overlap with the applications discussed below because the family of technologies supporting many archives’ search engines – information retrieval systems – partially constitute a subfield of NLP.

Reindexing of records, contra original order, is sometimes ideal for specific communities’ needs. For instance, interfaces that rely on NER, not physical collocation, are one option that can be especially useful for academic users. This is likely because search of NER labels is an effective and relatively efficient way to find documents related to a given historical figure or organization (Schneider and Chan Reference Schneider and Chan2016). One example of this is Handler et al. (Reference Handler, Mahyar and O’connor2022), a platform for exploring a newspaper archive. That project uses NER to index the archive and presents results returned via three separate interfaces for different contexts in which one might view the query (a system built in response to journalist interviews). This exemplifies how archival desiderata considered “typical” may not be a priority, depending on the archive’s domain and its expected users. Similarly, both Ranade (Reference Ranade2016) and Hinrichs et al. (Reference Hinrichs, Alex, Clifford, Watson, Quigley, Klein and Coates2015) construct record-level indices for individuals and locations (respectively) because they are sensible axes along which to explore the archives in question – the British National Archives and a corpus of 19th-century commodity trading documents. In general, bespoke search systems that incorporate NLP need not be opaque or context-agnostic; in fact, NLP can be used to extract features of particular interest to a specific user audience.

The contextual nature of a user’s search needs can be addressed in other ways. Topic modeling is a means of categorizing a corpus of documents according to emergent themes; it constitutes an interface that might assist historians and others interested in topical through lines (Muresan and Harper Reference Muresan and Harper2004). In existing work on archives, however, topic modeling is frequently a first-line approach for producing record-level description (Coeckelbergs and Van Hooland Reference Coeckelbergs, Van Hooland and Mugnaini2020; Hengchen et al. Reference Hengchen, Coeckelbergs, van Hooland, Verborgh and Steiner2016) and thus only conducted once, in order to produce a content index. These approaches rely on emergent patterns in the corpora they describe, rather than descriptors chosen by archivists. As Underwood (Reference Underwood2014) puts it, fitting a topic model entails “harmonizing … initial human assumptions with the complex evidence presented by the documents themselves” in a way that makes those assumptions explicit.

Automating processes related to appraisal and the creation of metadata primarily implicates decisions by archivists. By contrast, interfaces inherently involve decisions being made by users. As Hedstrom (Reference Hedstrom2002) argues, the advent of digital archival interfaces both reminds us that archival infrastructure constrains interpretation and that the typical interfaces an archive provides to a user, such as finding aids, impose a “linear narrative structure” on the contents of a collection – sometimes obscuring discontinuities and gaps. Archivists are obligated to provide context for the records in a given collection; this context ideally includes the construction of the archive itself. Moreover, the histories of digital records are important parts of their information content. Acker (Reference Acker2017) suggests that digital records are characterized more by the traces of their mutation across different platforms than even their content; archives’ power over the evidentiary status of records depends on recognition of this quality. Automating archival processes, then, requires attention to both the modifications made on the metadata paratext associated with a record and the “exhaust” created by the processes themselves, like the paradata that Davet et al. (Reference Davet, Hamidzadeh and Franks2023) refer to.

4.1. Generative AI in archival infrastructure

Numerous organizations have begun to explore the use of generative AI, such as LLMs, to describe and index records. The exact features entailed vary with the archive. For instance, Potter and Saccucci (Reference Potter and Saccucci2024) – a project exploring LLM-based indexing at the Library of Congress (LOC) – is concerned with producing standardized Machine-Readable Cataloging (MARC) descriptions in addition to specific subject headings.Footnote ⁷ In this section, we review the LOC project as well as three generative uses of generative AI: a framework for using LLMs to interface with web archives, an article describing the use of LLMs to produce more comprehensive metadata for the Grand Ducal Archive of the Medici, and an article presenting an LLM and its uses in the Chinese state archive from which its pretraining data came. The second use exemplifies how LLMs may be used in contexts where traditional approaches to processing and interface construction may be inappropriate, while the third and fourth demonstrate the use of an LLM to automate a descriptive process in a more classical archival setting. We avoid reviewing papers that do not explicitly discuss a particular archive, such as Nguyen et al. (Reference Nguyen, Nguyen, Nguyen, Buntine, Fjeld, Tran, Tran, Thanh and Miyoshi2025), or that deal with personal collections, such as Garcia and Weilbach (Reference Garcia and Weilbach2023), although these works also suggest possible uses of LLMs in archives.

4.2. Remembering with LLMs

A growing number of exploratory papers have also examined potential uses of LLMs and AI more broadly in memory projects, especially educational ones (see Kouzelis and Spantidi Reference Kouzelis and Spantidi2024). This work tends to focus on the use of AI to memorialize events which create archival silences (Ghaddar Reference Ghaddar2016; Hawkes Reference Hawkes2012), such as genocides (Makhortykh et al. Reference Makhortykh, Zucker, Simon, Bultmann and Ulloa2023). Toth (Reference Toth2021), for instance, uses a semantic search system to support the recovery of Holocaust testimonies. In general, however, the use of LLMs in this context remains nascent. Thus, we speculate on the basis of available documentation of reports on ongoing efforts and work on these technologies within NLP.

As discussed above, LLMs can be used to summarize and synthesize text. Moreover, frameworks such as RAG allow one to cheaply integrate an external collection with an existing LLM (Lewis et al. Reference Lewis, Perez, Piktus, Petroni, Karpukhin, Goyal, Küttler, Lewis, Yih, Rocktäschel, Riedel and Kiela2020). However, LLMs are not completely faithful or unbiased authors (Bender et al. Reference Bender, Gebru, McMillan-Major and Shmitchell2021). GPT-4 has been found to generate descriptions on the basis of demographic information that are stereotypical if not overtly offensive (Cheng et al. Reference Cheng, Durmus and Jurafsky2023). Bias has also been observed in RAG systems, including in historical narratives (Fan and Gardent Reference Fan, Gardent, Muresan, Nakov and Villavicencio2022). Any proposal to use LLMs to generate new text raises questions about appropriateness, fidelity and authenticity.

4.3. Narrative generation

While many potential uses of LLMs closely resemble previous approaches with earlier techniques, the generation of coherent text opens up much more novel possibilities. As such, archival theory has not yet grappled in detail with how to think about and work with such systems. Nevertheless, inspired by recent excitement around “bringing history to life” (Eisikovits Reference Eisikovits2021), we can look to past debates about the more speculative writing of history, as expressed in the work of Saidiya Hartman and Stephen Best.

Saidiya Hartman’s exploration of the limits and possibilities of the archive (Hartman Reference Hartman2008; Reference Hartman2019) provides one means of conceptualizing generative AI in the archives. Hartman’s central contribution is “critical fabulation,” an approach to interpretivist archival research that entails writing speculative prose about the subjective experiences of marginalized people who are present in the archive but silent. Meanwhile, Stephen Best critiques the expectation that cohesive histories can be recovered from archives (Best Reference Best2018), fearing the treatment of the archive simply as “a rich repository of material from which one may recover and reanimate the lives of the enslaved.” Best argues that the recuperative stance taken by Hartman entails interpretive irresponsibility: grafting contemporary expectations and norms onto the experiences of people with very different self-conceptions. While we cannot do justice to their work here, the debate between Hartman and Best provides a useful touchstone for thinking about the emerging possibilities and dangers around the use of LLMs in archives.

In principle, generative AI could be used to try to generate the kinds of first-person testimonies that are absent from the historical record for many marginalized groups. It also constitutes a new way of producing cohesive histories out of archives; LLMs can be used to smoothly aggregate textual information and induce the appearance of continuity. In light of Hartman’s work and Best’s critique, it may be worth considering (1) what kinds of recovery and reinterpretation are made possible by the use of LLMs and (2) whether these processes are still accountable to the histories of discontinuity and disruption that Best identifies. Any use of LLM-generated text in archives, especially in the context of perspectives that have been silent, risks the imposition of contemporary perspectives and desires (including those encoded in the LLM) onto historical materials. It will also have to contend with the nature of LLMs’ representations of identity groups, which can often tend toward stereotypes (Wang et al. Reference Wang, Morgenstern and Dickerson2024), as well as the discontinuity of identities salient at one time but not another. This is not to say that LLMs are never useful in an archival context where silences abound. Rather, the use of LLMs to critically fabulate the perspectives of the under- and unrepresented itself requires critical scrutiny.

5.1. Sensitivity to context

Context can mean many different things (Burke Reference Burke2022). We recognize three main categories of archival context: the context that records in a collection give to other records in a collection, the context that archivists endow records with through description, and the context that external information known by users gives to records in a collection. We focus on the first and second aspects of context, given their more obvious relationship to archival processing.

First, there is a basic conflict between the NLP paradigm of individual document classification and the contextually aware mode in which archivists appraise, describe and represent archival materials (Cook Reference Cook2011; Fintland Reference Fintland2016; Yakel Reference Yakel2003). As Rolan et al. (Reference Rolan, Humphries, Jeffrey, Samaras, Antsoupova and Stuart2019) explicitly note, part of the difficulty of automating archival processes is in approximating the holistic means by which archivists make decisions about individual documents while thinking of collections as aggregate wholes. This situation is exacerbated by the networked and nonlinear nature of contemporary communication technologies (Acker Reference Acker2017). Construction of a set of reasonable “series” out of, e.g., millions of social media posts is far from straightforward.Footnote ⁹

Second, the choice to endow an archive with one context or another can significantly impact user interpretations as well. Douglas (Reference Douglas2016) emphasizes that the traditional mode in which finding aids’ descriptive text is written is that of “the view from nowhere” – obscuring the process by which a collection was constructed while narrativizing a seemingly static and universal vision of a given collection’s history. This is despite the sometimes contentious processes by which materials are acquired. Recalling the earlier discussion of paradata (Davet et al. Reference Davet, Hamidzadeh and Franks2023), Douglas might encourage the production of a finding aid wherein the automation of descriptive choices is motivated and explained within the archive’s social context.

5.2. Transparency

Most NLP models used today take the form of neural networks involving many layers of nonlinear functions. Language carries both compositional meaning and other types of semantic content (referential, metaphorical, etc.). As a result, it is frequently difficult to attribute outputs to specific elements of a text and almost always difficult to attribute them to a model mechanism. This poses a particular challenge for using NLP in archives, given the emphasis on explicating processing choices to users (Caswell Reference Caswell2021; Douglas Reference Douglas2016). In the context of appraisal specifically, the invisibility of the discarded materials in a typical archival setting further exacerbates this challenge; only the selected records are visible to the user.

Bunn (Reference Bunn2020) argues that machine learning in archives is best undertaken with a plan for articulating how a given model functions. This means both making basic facts about the model available (e.g., performance on a representative dataset) in a form that a user can understand and conducting an analysis of the model using methods from the machine learning subfield of interpretability. Transformers, however, are sufficiently complex that it is difficult both to visualize mechanisms within the model and to simplify them for the benefit of users (Abnar and Zuidema Reference Abnar, Zuidema, Jurafsky, Chai, Schluter and Tetreault2020; Antverg and Belinkov Reference Antverg and Belinkov2022). We might also hope to provide some accessible description not just of the process by which an LLM was used, but also of the data used to train that model. (Desai et al. Reference Desai, Pasquetto, Jacobs and Card2024) and Schoenebeck and Conway (Reference Schoenebeck and Conway2020) have argued that archival thinking and practices are themselves applicable to the data used in machine learning systems. Proper description of archival processes relies on documentation – ideally archival – of the data used to construct those processes.

5.3. Resource limitations

Successful adoption of NLP systems in general and LLMs in particular at a given archival institution relies on a number of material factors. These include the experiences of archivists there, financial constraints with respect to costs like API fees and the nature of the processing backlog. The limitations of available data in historical contexts – especially labeled historical data – can similarly impede system development and evaluation, given the data requirements of large neural models (Hedderich et al. Reference Hedderich, Lange, Adel, Strötgen and Klakow2021). Annotation in a given archive may need to be undertaken by people with a specific experience, such as expertise or community membership, in order to produce representative training data. Indeed, the annotation projects that support frontier LLMs have been critiqued on the basis of exploitative employment models, the injury done to annotators when they contend with harmful content and other ethical concerns (Crawford et al. Reference Crawford, Miltner and Gray2014; Muldoon et al. Reference Muldoon, Cant, Boxi and Graham2024); these issues are especially salient in the context of archives that may relate to marginalized populations. LLMs appeal to some of the practitioners whose work we review because of a perceived lesser need for expert annotations, although validating their use also requires human judgments. Practical use of LLMs requires either expensive hardware (with open models) or per-use cloud computing or API fees – whereas earlier methods can be run locally on standard computers. Zhang et al. (Reference Zhang, Peng, Wang and Hou2024) is somewhat unique because of how it involves developing an archive-specific LLM; outside of state archives with significant resources, pretraining an entirely new model is likely infeasible. However, we recognize the largest factor in archival resource limitations to be time.

Time is a central concern for archives. Many archivists simply do not have the time to acquire skills that are generally required to use NLP tools seamlessly, such as Python programming (Jaillant and Aske Reference Jaillant and Aske2024; Poole and Todd-Diaz Reference Poole and Todd-Diaz2022; Rasmussen Reference Rasmussen2019). There can be a high upfront cost to adoption of NLP systems (Jaillant and Aske Reference Jaillant and Aske2024) even if they have the potential to speed up processing in the long run. Archivists at state archives especially tend to be occupied by the massive backlogs of unprocessed (and undigitized) materials that have accumulated over decades, absent a plan for addressing them (Greene and Meissner Reference Greene and Meissner2005). As Mordell (Reference Mordell2019) observes, the practical outcome of many efforts to digitize and “datafy” archives is frequently much time wasted by archivists and few benefits to users. AI is hardly a panacea for some of the most pressing content-related problems facing archives; the limitations of many collections – especially from the perspectives of marginalized communities – originate in preservation and appraisal decisions made a hundred years ago (Jaillant et al. Reference Jaillant, Mitchell, Ewoh-Opu and Urbaneja2025).

5.4. Maintenance

Maintenance entails both the labor that updating infrastructure and schema involves and the high-level ideation that arises from revising archival processing practices. The manifesto (Acker et al. Reference Acker, Arnold, Castro, Galvan, Hswe, Meyerson, Nowviskie, Lassere, Olson, Parsons, Russell, Vinsel and Wright2019) produced by the archival professional collective The Maintainers argues that “moments of upgrade and reimagining can also resource and provide justification for maintenance processes.” Ideally, adoption of more sophisticated NLP technologies and the new schema they make possible is also an opportunity to participate in maintenance projects. Conversely, Jackson (Reference Jackson2014) contends that new technology’s adoption relies on a continuous project of repair which articulates that technology in the context of older infrastructure and addresses its breakdown. LLMs may at first seem inherently flexible in a way that makes maintenance less necessary; intuitively, the prompt templates undergirding systems using LLMs may make them easier to understand and modify. However, the polysemic nature of fluent text may obscure the limitations of an LLM-based system (Shah and Bender Reference Shah and Bender2022), both intrinsically and as interfaces with other components of archival processing infrastructure. As seen in Potter and Saccucci (Reference Potter and Saccucci2024), LLMs are also rarely the optimal choice for tasks involving well-defined yet complex schemas. Adapting any NLP system, but LLMs especially, to function within a preexisting ontology requires its own maintenance and articulation work.