2.1 What they offer and how they work

As outlined earlier, archival AI assistants can now be used by the public to explore major archives online, including the European Parliament Archives and the Luxembourg National Library (Archives of the European Parliament, 2024; eLuxemburgensia, 2023). Similar systems are also being planned or piloted at the National Archives and Records Administration (NARA) in the United States and the Arquivo Histórico Ultramarino de Lisboa of Portugal (Koebler, Reference Koebler2024; NARA 2025; Mapping Atlantic Portuguese Empire (MAPE), 2025). Notwithstanding some differences in design, their basic architecture is highly similar: applying Retrieval Augmented Generation (RAG) to existing LLMs (such as Anthropic’s Claude or Google’s Gemini), the AI assistants draw on indexes of the relevant archives’ documents to answer, in natural language, users’ questions pertaining to these archives. In particular, addressing well-known problems that off-the-shelf LLMs face in regard to producing outputs lacking factual accuracy and making “reasoning” errors, RAG techniques seek to anchor LLMs’ outputs more closely in specific sources of knowledge, such as indexes of digitised documents from an archive, to ensure that generated outputs stay closer to original source material (although not always successfully), and enable AI assistants to pinpoint specific sources on which answers are based (Lewis et al., Reference Lewis, Perez, Piktus, Petroni, Karpukhin, Goyal, Küttler, Lewis, Yih, Rocktäschel, Riedel and Kiela2021; Magesh et al., Reference Magesh, Surani, Dahl, Suzgun, Manning and Ho2024). In the context of archives, then, RAG-based AI assistants can only access those sources which have been prioritised for digitisation (as opposed to the entire physical holdings), and expertise in LLM engineering, whether sourced internally or externally, is necessary for the development of these new tools. Some of the archival AI assistants following these design principles have been given names: “Archibot” at the European Parliament Archives, “Archie AI” at NARA, and simply “Chat” at the National Library of Luxembourg (Archives of the European Parliament, 2025; eLuxemburgensia, 2025; Koebler, Reference Koebler2024). While this article engages with these existing, pioneering archival AI assistants, it is not intended as a critique of these specific systems. Rather, the discussion draws on aspects of these existing systems as jumping-off points to engage the larger, ongoing project of developing similar (including more advanced) systems likely to be introduced at further archives, and to initiate epistemological and practical exchange on their risks, benefits, and appropriate design.

2.2 Framing and marketing

3.1 Conceptual disruptions

What effects do such framings have, in the minds of users? When interfaces suggest that AI assistants can produce answers to historical questions based on the contents of archives, pressure is placed on established understandings on how, if at all, we can extract reliable information about the past from a collection of documents. In particular, pressure is placed on our understanding of and expectations for a range of roles (“archivist,” “historian,” “expert”); forms of information (“document,” “archive”); and outputs (“history”) involved in that process. We suggest that one way this pressure can be productively understood is through the lens of what has been termed conceptual disruption (Hopster and Löhr, Reference Hopster and Löhr2023; Löhr, Reference Löhr2023). As explained by Löhr, a technology can be considered conceptually disruptive if “it challenges or prompts an overturning of entrenched conceptual or classificatory norms and practices, i.e., practices pertaining to language and thought” (Löhr, Reference Löhr2023, pp. 2–3). In this case, the disruptions caused by archival AI assistants may generate conceptual uncertainty, as the largely unknown scope of these systems’ abilities induce uncertainty around the application of entrenched concepts traditionally used in relation to archives.

While archival concepts are not as uniformly understood or entrenched in popular thinking as those discussed in the conceptual disruptions literature thus far (for example, disruptions to the notion of “author” by the widespread uptake of LLMs such as ChatGPT; see Khosrowi et al., Reference Khosrowi, Finn and Clark2024), we believe this framework to also be useful here, especially in a heuristic way. First, a conceptual-disruptions lens allows us to focus on the important role of entrenched concepts within the historical and archival sciences, even if these do not always have well-defined, shared meanings across broader society. Second, it facilitates clear, organised discussion of the conceptual, as well as methodological and practical, uncertainties and questions that are raised by these new archival tools for both lay and expert users. For instance, the lens invites us to consider: are archival AI assistants “assistants,” comparable in competence to a human archival assistant or archivist? Can they competently draw on archival sources to synthesise “histories” that attend to user queries, like a “historian” or “expert” could? How should users interpret what kind of outputs archival AI assistants produce (i.e. do they constitute “history”? (cf. Khosrowi and Finn, Reference Khosrowi and Finn2025)) and in what sense, exactly, do they provide “access” to “sources” and “archives”? These issues are conceptually and practically unclear, despite deployers’ more definitive framings. As we can see, archival AI assistants hence induce a range of conceptual and practical disruptions and uncertainties that affect crucial concepts we use to understand and organise epistemic practices revolving around archives. Let us further unpack them, highlighting concerns that arise around specific disruptions.

3.2 “Sources,” “archives”

Using archival AI assistants to answer historical questions and framing them as tools that allow one to “speak directly” with the archives could obscure well-established expert beliefs that “documents,” “sources” and “archives” are not unmediated, clearly navigable, or necessarily comprehensive. Historians and archival theorists have long shown that records find their way to us via many intermediaries and can contain multiple points of view. For instance, Ketelaar has argued that “archival fonds, archival documents, archival institutions, and archival systems contain tacit narratives which must be deconstructed in order to understand the meanings of archives” (Ketelaar, Reference Ketelaar2001, p. 131). Indeed, by suggesting that an AI tool can immediately extract an answer to a historical question based on the most relevant documents it identifies in an archive, we risk encouraging or facilitating a strange sort of reversal on the very established belief, most famously asserted by Derrida (Reference Derrida1995) and Foucault (Reference Foucault2002), that archives are not to be approached as mere repositories of records that offer a clear window onto the past, but as complex systems that warrant careful examination themselves (Callahan, Reference Callahan2024).

With this in mind, it becomes clear that the “documents” and “collections” that these AI tools are drawing on to provide their answers to users are not unmediated sources that can instantly “unlock decades of democratic history,” as Amodei of Anthropic phrased it in her announcement, nor consistently provide one singular, “true” account of past events in the form of an “answer.” Rather, these digitised documents have been selected and shaped by the actions and decisions of various human actors, along with the biases, institutional norms, archival practices, and sometimes sheer luck that governs the survival of historical material. As various scholars have demonstrated, even what seem to be the driest administrative documents can be unreliable and contain biases, being the products of what an institution wanted to show and preserve (Prescott, Reference Prescott, Greengrass and Ashgate2008; discussing Hunnisett, Reference Hunnisett, Bullough and Storey1971; Galbraith, Reference Galbraith1964). What is more, the selective and political processes of digitisation have often only perpetuated these issues, further narrowing the scope of materials most examined by users and entrenching the absences of marginalised groups in archives (Thylstrup, Reference Thylstrup2019; Milligan, Reference Milligan2022; Ortolja-Baird and Nyhan, Reference Ortolja-Baird and Nyhan2022; Zaagsma, Reference Zaagsma2023).

Moreover, significant epistemic gaps exist between historical phenomena/events, physical sources recording these phenomena and digital representations of those sources. Acts of encoding, translation and re-presentation underpin the transformation from one form to the other, and these require careful scrutiny and acknowledgement (Jordanova, Reference Jordanova2019). In particular, awareness of what information is often lost when sources are digitised and/or transcribed (e.g. a telling change in ink colour or hand; differences between versions of a text; or redactions and edits) and that errors may be introduced (e.g. through OCR) is vital (Prescott, Reference Prescott, Greengrass and Ashgate2008; Smith and Vine, Reference Smith and Vine2024). These necessities are ever more pressing in the case of archival AI assistants, as these tools insert yet another intermediary layer and shift in medium, providing users with a new re-presentation and indeed interpretation of the underlying sources.

Making the distinction that the AI assistants are likely providing an interpretation (even if mindlessly and unintentionally), and not merely raw, unmediated information, is important – and stems from the understanding that sources are not “quarries of raw factual material” (Prescott, Reference Prescott, Greengrass and Ashgate2008, p.13).Footnote ¹ Rather, as Ludmilla Jordanova has highlighted, sources “stem from the messiness of lives as they are lived – including institutions and arms of government” and “it is worth reminding ourselves of the anachronisms of taking sources and making them work for the discipline of history” (Reference Jordanova2019, p. 43). As has been argued in other disciplines in relation to the misleading use of the term “data literacy” to describe the transformation of data into comprehensible narratives (as if the insights were waiting, fully-formed, within the data all along) (Veel, Reference Veel2018, section 3), current framings of AI assistants producing accounts of the past based on the digitised contents of archives obscure the complexity of the underlying process, as well as the role of the human user (e.g. through their query or “prompt”) (see also Khosrowi et al., Reference Khosrowi, Finn and Clark2024). With their promise of fast answers, archival AI assistants could encourage lay users to fall into the trap of assuming that “chatting” with “documents” and “collections” can quickly and easily produce a window onto past events, and of imagining a misleadingly neat, straight line between historical accounts and their evidential base. Experts with historical training, too, might be susceptible to these misconceptions, especially in terms of failing to appreciate the extent to which AI assistants affect their engagement with a collection. Indeed, historians are unlikely to be immune, as supported by various uncritical approaches taken by scholars to new digital affordances (see, for example, calls for greater methodological reflection made by Hill, Reference Hill2016; Hitchcock, Reference Hitchcock2013; Milligan, Reference Milligan2013, Reference Milligan2022; Putnam, Reference Putnam2016). While concerns about both the public’s ability to exhibit relevant skills in interpreting historical documents and historians’ uses of new digital tools are neither novel nor distinctive of interactions with AI systems, our point here is that framings around archival AI assistants may further exacerbate such concerns in significant ways.

3.3 “History,” “historian,” “archivist”

Oversimplifications of how historians come to formulate accounts of the past – which, in reality, is a multi-step and complex process – brings us to our next set of disruptions, as our understanding of “history” and the roles played by those involved in its production come under pressure. Given users are encouraged to ask these chatbots historical questions about the past, the implicit suggestion is that the output is some form of “history,” but is it? By jumping directly from a limited number of documents, predicted to be most relevant to the query, to automated reports, archival AI assistants skip and/or skim over fundamental steps of the historical process – and any distinction between information/data/sources/evidence and, then, interpretation/analysis/argument/historical knowledge becomes blurred. In order to produce the latter, the process includes, but is not limited to, wide-ranging research (including, perhaps, engagement with serendipitous or less obvious finds which have fallen under the radar of e.g. keyword searches); thoughtful selection of sources according to articulated criteria; decisions of foregrounding and backgrounding; source criticism; intensive contextualisation; the disentangling of the many layers of information present in sources; and reflection on one’s findings in light of others’ research (Currie and Walsh, Reference Currie and Walsh2019; Jordanova, Reference Jordanova2019; Prescott, Reference Prescott, Greengrass and Ashgate2008).

What is more, the process by which historical accounts are produced also arguably involves applying human empathy to historical topics, and showing an awareness of a topic’s relation to the present, both of which an archival AI assistant or any LLM cannot currently do (Götter, Reference Götter2024; discussing Droysen, Reference Droysen1875). As this brief description of at least some of a historian’s process suggests, archival AI assistants could encourage (especially nonexpert) users to overlook the complexity of producing accounts of the past, and to view history in a more simplistic, one-dimensional way. Indeed, these tools could also obscure well-established understandings that the practice and outputs of history are often disputable, malleable, opinionated, and so on.

At the same time, it is generally considered important that those engaging with historical knowledge are able to assess the possible biases or influences which may have shaped the work of its author (e.g. personal beliefs, intellectual heritage, etc.), and while this is at least somewhat possible with a human author, it would be a far greater challenge with an AI assistant (see also Clavert and Muller, Reference Clavert and Muller2024, p.15). Despite being an overused term at this point, these AI systems are indeed “black boxes” and it is therefore difficult for users to understand why they might, for example, privilege or hone in on some archival documents over others (see, for instance, Makhortykh et al., Reference Makhortykh, Vziatysheva and Sydorova2023a; De Ninno and Lacriola, Reference De Ninno and Lacriola2025).

Further pressure is placed on how we should conceptualise archival AI assistants’ workings and outputs by the fact that LLMs are not carrying out “analysis” or “reasoning” in the way humans do: they are, at bottom, statistical models of the distribution of word forms, designed to produce likely and plausible sequences of words (Shah and Bender, Reference Shah and Bender2024; Stone et al., Reference Stone, Goodlad and Sammons2024).Footnote ² At this point, more far-reaching and fundamental questions also arise, such as: what are we losing when we hand over the process of formulating historical knowledge to AI systems? Is the “point” of the practice of history simply the end result, therefore arguably justifying faster routes to that – or does its value also lie in the process, and the fact that humans undertake the bulk of the labour? Although not within the scope of this article, these are important and relevant questions which warrant future discussion.

3.4 Public responses: disruptions evident

Emerging public discussions around archival AI assistants on social media platforms illustrate how the conceptual disruptions and uncertainties we describe here are unfolding amongst lay audiences and users. Consider, for instance, the comments made under the announcement of the European Parliament’s Archibot on LinkedIn. One comment notes:

Such uncritically enthusiastic responses underwrite concerns about the potential for lay users setting misplaced confidence in archival AI assistant’s abilities, being drawn towards epistemic short cuts that bypass the role of expert intermediaries and instil illusions of comprehensiveness. At the same time, some voices also highlight reasonable concerns about censorship and disinformation (e.g. “I hope it’s not going to ‘rewrite history’ then. ‘For safety’” (LinkedIn, 2024)), demonstrating how AI assistants could also cause erosions of trust and confidence in archival institutions (see also Makhortykh et al., Reference Makhortykh, Zucker, Simon, Bultmann and Ulloa2023b).

Other comments proceed to highlight how Archibot facilitates “unlocking 2.1 million documents, spanning seven decades, for the world? That’s not just nice to have, that’s a fundamental shift in transparency and accessibility” (LinkedIn, 2024). While it seems plausible to think that Archibot indeed helps mitigate barriers to archive accessibility for citizens (though less comprehensively than the discussion suggests), such comments again raise concerns about misplaced trust and confidence.

Beyond social media commentary, other, journalistic responses similarly fall short of distinguishing the outputs of archival AI assistants from concepts like “history” or “historical knowledge.” For instance, an article about the European Parliament’s Archibot equates the outputs of Archibot to “history” without distinguishing “history” from “historical sources” or considering or acknowledging the significant differences presented by an AI-mediated process of gathering, learning and engaging with “history.” “This initiative is not merely an archival project; it’s a democratic endeavour aimed at enhancing the public’s understanding of the European Union’s legislative history,” the article reads; “It enables citizens, researchers, and policymakers to dive into the rich legislative history of the European Union, offering insights into the decision-making processes and debates that have shaped current policies and regulations” (Kimaid, Reference Kimaid2024).

As these impressions from public reception suggest, it is important, especially for lay audiences, to conceptually situate archival AI assistants on more solid foundations, especially their capabilities, outputs and the nature of our interactions with them. Crucially, we believe that doing so helps users avoid collapsing distinct steps of historical practice into a misguided vision of streamlined, AI-facilitated conversational processes that directly extract historical knowledge from archival contents.

The concerns outlined here are particularly pressing in view of the similarities of archival AI assistants to popular conversational systems built on LLMs like Claude, ChatGPT and others. Recent years have seen rapid growth in the user base of such systems, with users increasingly exploring use-cases, getting accustomed to prompt-based interactions with AI tools, and integrating these systems into everyday workflows to optimise and partly automate knowledge-related tasks. While a significant portion of users show scrutiny and caution in their use of these systems, public discussions on platforms like Reddit suggest that many people also use these systems in ways that “skip steps” associated with manual research processes; for example by moving straight from prompt to answer; employing chatbots as teacher-like entities; cementing the belief that there is one “correct” answer to complex questions (that one can reach with the right prompting); and reporting relief in speeding up tedious aspects of traditional research (e.g. Reddit, 2024). A growing number of studies are making similar claims about lax LLM usage, and how it may atrophy various critical thinking and research skills (Lee et al., Reference Lee, Sarkar, Tankelevitch, Drosos, Rintel, Banks and Wilson2025; Shah and Bender, Reference Shah and Bender2022). Given the strong similarities between archival AI assistants and familiar conversational AI systems, it seems reasonable to think that the public may engage with archival AI assistants in similar ways, thereby importing modes of interaction that (even more so than in the most common knowledge-retrieval settings) threaten to induce simplistic views on how to acquire knowledge of the past.

3.5 Conceptual disruptions: the right lens?

As we have argued so far, the lens of conceptual disruptions seems helpful to draw out how archival AI assistants put pressure on how we apply familiar concepts in new contexts where novel technologies may assume roles previously occupied by humans and generate outputs that may, at face value, resemble those produced by human experts. Even so, we may ask what, exactly, the lens of conceptual disruptions adds to understanding these emerging issues. Let us elaborate and respond to some potential reservations about it.

One way to push back on our suggestions is to insist that at least some of the issues we highlight here are simply new forms of familiar problems: unreliable historical accounts have always existed; some people already lacked strong skills in source criticism; all new digital methods change the practice of and/or engagement with history in some way; and general forms of misunderstanding around how much we can know for certain about the past have been pertinent well before the advent of archival AI assistants. With these points in mind, one might worry that such a lens is misleading because it presupposes a relatively fixed pre-existing conceptual fabric that is then disrupted by archival AI assistants. As the objection goes, it may seem that some of the terms discussed here (e.g. “history,” “archives”) did not, and do not, have well-defined, shared conceptual meanings across society – and thus the putative disruptions we highlight do not cut as deep as, say, how generic LLM-based conversational agents like Claude and ChatGPT disrupt the public’s notion of what it is to be an “author” (Khosrowi et al., Reference Khosrowi, Finn and Clark2024), or how the invention of the mechanical ventilator interrupted our understanding of “death” (de Boer and Hoek, Reference de Boer and Hoek2020; Löhr, Reference Löhr2023). Unlike disruptions affecting such household concepts, we may think that, even before the advent of archival AI assistants, many members of the general public might not have considered archival documents to be heavily mediated remnants of the past, so it remains unclear in what sense archival AI assistants induce novel misunderstandings, as such misunderstandings have recognisably existed before.

Pushing back on these reservations, we point out that while some concerns we highlight here are familiar, they are aggravated by the fact that archival AI assistants are, for now, considerably more prone to error and opaque than human experts whose roles they may conceptually and practically encroach on. It is not just that laypeople may fail to appreciate the considerable amount of analysis, interpretation and expertise necessary to bring archival materials to bear fruitfully on queries about the past, but that the novel roles played by archival AI assistants make these epistemological tenets even more difficult to recognise, in light of the lure of automation and straightforward-seeming interaction with sources they provide to users. Whereas laypeople may have falsely, but somewhat safely, disregarded the crucial role of expertise in interrogating archival materials in the past, they may now be drawn to rely similarly on systems that are likely to be significantly more brittle, opaque, and prone to error than human experts are (Borji, Reference Borji2023).

Moreover, at least within the specific domain of historical research itself, there are indeed more stable conceptual fabrics that archival AI assistants disrupt, e.g. concerning entrenched concepts such as “source,” “evidence,” “history,” and so on. Historical researchers, perhaps even more than the general public, are pressed to make conceptual choices about how to understand and frame the (appropriate) roles of archival AI assistants, and these choices are complicated by considerable conceptual uncertainty that, in turn, spurs further methodological uncertainties downstream around how to use these systems in epistemically responsible ways.

Beyond capturing some of these downstream street-level issues (of, e.g. how exactly to divide labour between humans and machines in archival contexts, how much credibility to assign to archival AI assistants’ outputs, etc.), it is important to appreciate that the lens of conceptual disruptions also helps us in two other crucial ways. First, it encourages us to focus on the bigger picture. Concepts are crucial for understanding and organising the world: they are the currency that underlies and shapes our thinking and our norms. By taking a step back and focusing on concepts, we are able to consider the effects of AI in archives not just in terms of archival and historical theory but within a broader landscape of how AI technologies are disrupting “entrenched conceptual or classificatory norms and practices” across different industries and areas of society (Löhr, Reference Löhr2023, p. 3). This broader focus helps us better attend to seemingly mundane but cognitively significant choices, such as the rhetoric we use in framing archival AI assistants and the implicit messages regarding their roles, capabilities and affordances that are embedded in that rhetoric.

Second, focusing on concepts helps us think about new solutions to managing the impacts of novel technologies, and how we might negotiate their proper roles and facilitate users’ understanding of their limitations. We are not confined to applying only familiar concepts to emerging AI systems, their roles and their outputs; we can also come up with new, more suitable concepts. The philosophical literature on conceptual engineering (Cappelen, Reference Cappelen2018) offers a range of approaches for how we might engineer new, or better, concepts that respond to disruptions, minimise conceptual uncertainty and shape our interactions with archival AI systems (Hopster and Löhr, Reference Hopster and Löhr2023). In what follows, we explore several options for how conceptual, communicative and methodological work can help with these aims.

4.1 Clarity in disclaimers and instructions

Like search environments, AI assistant interfaces create an implicit model in the minds of users of what questions can be asked and what answers might appear (Ciravegna et al., Reference Ciravegna, Greengrass, Hitchcock, Chapman, McLaughlin, Bhagdev, Greengrass and Ashgate2008). A key guiding aspect of these interfaces is therefore their disclaimers and instructions. In order to mitigate the conceptual disruptions of archival AI assistants, detail and clarity should be provided to users on: the intended and appropriate uses of these tools; the extent of the documents they draw on; and the limitations of their outputs, especially in comparison to other historical outputs (see also Clavert and Muller, esp. p. 21). This information should be prominent, and equally present in interfaces, information pages, announcements, marketing content and training materials.

To elaborate on these points in turn, it would be beneficial to users to flag that, while useful in locating materials, the archival AI assistant’s ability to provide historical accounts is more limited and will likely necessitate significant fact-checking work. Moreover, the distinction between these types of queries (locating materials versus seeking historical knowledge) should be made clear. The source base used to formulate responses should also be clearly described at the outset, in terms of extent and type (e.g. a maximum of 10 documents deemed most relevant to the query from a subset of 75,000 quality-checked files out of the 1.5 million available at the European Parliament Archives (Kimaid, Reference Kimaid2024)), in order to facilitate users forming relevant expectations and understanding the depth of the response. It would also be beneficial to clearly emphasise that the AI assistant’s response is based centrally on the archive’s contents, rather than only a diffuse set of web-scale training data, as this might help to distinguish the assistants from other common and popular LLMs (ChatGPT, Claude, CoPilot etc.) and mark them more uniquely as a tool for conducting archival research.

In terms of limitations, users may benefit from clear communication in disclaimers and instructions of the risk of the AI assistant “hallucinating”/“confabulating,” but also other less popularly known possibilities, such as the effects of LLMs’ underlying safety mechanisms (which might e.g. censor certain sources or affect the accuracy of historical text translations (Tekgürler, Reference Tekgürler2025)); inaccurate referencing (Lopatto, Reference Lopatto2024); and the contemporary perspective of these models which, having been trained on modern texts from the internet, could imbue outputs with various anachronisms in terms of source criticism and translation (Neudecker, Reference Neudecker2024). In addition, clear and detailed documentation of how the archival AI assistant has been developed (especially in terms of any fine-tuning of the underlying off-the-shelf LLMs and any expert evaluations of model performance) should be made easily available from the main interface, so that users are encouraged to consider what limitations the AI assistant might have. Ideally, this would be made available in both accessible and advanced forms, so that users of differing technical abilities can engage with the documentation effectively.

4.2 Alternatives for marketing and design

Slogans such as “Ask the Archives” echo the language used in marketing for other similar technologies (such as “speak to the PDF” or “chat with the book”) and, as discussed in Section 2, can lead to misunderstandings. While these sorts of phrases are perhaps a predictable or unavoidable evolution of language in response to these new tools, they nevertheless require reflective consideration. Perhaps it is more productive for archival institutions to avoid replicating this form of (private sector) communication about its tools, given that straightforward and transparent descriptions of systems’ capabilities could better facilitate productive and critical engagement with their collections. In particular, consistency in describing the goals of the tool might be preferable, i.e. asserting that it either increases accessibility or provides analysis and interpretation; and if both, then a clearer distinction between these goals is needed.

Relatedly, the option of conceptual engineering could be considered, in which more appropriate terms and associated concepts that avoid uncertainty are created for new technologies (Cappelen, Reference Cappelen2018). In the context of archival AI assistants, terms that clearly categorise and speak more clearly to the tool’s intended use and capabilities could be deployed (e.g. terming the assistant’s output as a “document synthesis” rather than the more vague “answer,” and eschewing potentially misleading terms like “historical knowledge”).

Avoiding replicating the choices of private technology firms is perhaps also beneficial in the realm of tool design. While the format of LLM-based chatbots may now be well-known and popular, this may not be the most appropriate choice for archival research; perhaps other ML/AI methods are better suited to preserving and/or replicating authentic, transparent human engagement with archives. As the wider dashboards which house some of the current archival AI assistants show, there are many more ways to use AI in the archives than LLM chatbots, and other exciting projects are underway (e.g. Widegren, Reference Widegren2024; Jaillant et al., Reference Jaillant, Mitchell, Ewoh-Opu and Urbaneja2025a; de la Rosa 2025 and other contributions in Jaillant et al., Reference Jaillant, Warwick, Gooding, Aske, Layne-Worthey and Downie2025b; Szatucsek, Reference Szatucsek2025). If LLM chatbots are to be built into archival information retrieval systems, then a rethinking of the question-answer format might aid in steering the tool away from participating in the wider, problematic trend of LLMs being used as the new Google Search, which arguably results in even greater commodification of knowledge and less critical engagement with the information generated (Karunaratne and Adesina, Reference Karunaratne, Adesina, Singh and Johnston2023; Lopatto, Reference Lopatto2024). The generation of articulated “answers,” which summarise a group of documents, rather than providing solely lists of sources may unintentionally give “the illusion of a full vantage point” (to borrow a phrase from the writer Ayşegül Savaş, in regards to the power of narrative), and it is worth considering what these outputs obstruct from a user’s view or consideration (Savaş, Reference Savaş2019). What is more, due to the sampling behaviours of LLM-based conversational systems like Claude or ChatGPT, outputs can display significant variability for one and the same prompt and may differ even more strongly between semantically similar (or identical) prompts that are phrased differently. As a result, outputs might change so that different users receive different “answers” for otherwise identical questions. The extent and significance of this variability and what measures may be taken to improve output consistency has yet to be researched in the context of archival AI assistants. It seems that society is now moving on from the “Googlisation” of knowledge (Hitchcock, Reference Hitchcock, Greengrass and Ashgate2008; Prescott, Reference Prescott, Greengrass and Ashgate2008; Vaidhyanathan, Reference Vaidhyanathan2012) to the LLM-isation of knowledge, and archival institutions have the opportunity to take an autonomous role in shaping that shift in ways that take user behaviour into thoughtful account and actively facilitate critical thinking.

4.3 Historians’ engagement

The training that historians have received will of course aid them in avoiding a number of the potential disruptions that these tools present, encouraging them to make appropriate sorts of queries; not take the archival AI assistants’ answers at face value; check and evaluate the sources provided; merely use the outputs as a launching pad for further thinking and research; and so on. That said, they may not be entirely immune to the risks posed (as indicated earlier) and it is still beneficial to consider how exactly historians might best use these tools, in ways that maintain rigorous historical practice. Standard transparency in describing methodological choices would of course extend to openness regarding the use of archival AI assistants (whether for locating sources, generating interpretations, or anything else), and historians should be conscious of how their use of an archival AI assistant might affect their research’s route and contents. Just as catalogues, keyword searches, OCR searchable text, recommendation systems (e.g. content-based filtering), and specialised digital methods (such as topic modelling) can currently impact the pathways of researchers’ findings and thought processes, archival AI assistants will now also play a guiding role, pointing the researcher towards some records and topics, and implicitly obscuring others. While this is not a problem in and of itself, it requires awareness and intentionality on behalf of the researcher. In similar ways to how researchers came to make productive use of the fresh perspectives topic modelling provides while being clear about the limitations of the method – being aware, intentional and transparent will also allow historians to make the most of the advantages of archival AI assistants (e.g. navigation, categorisation, distillation) while not losing sight of the complexity of the sources they find.Footnote ³ Moreover, facilitating researchers’ ability to accurately document their use of AI assistants, such as by making chat histories downloadable as txt files (as included in the MAPE Engine’s AI assistant; Mapping Atlantic Portuguese Empire (MAPE), 2025) will be crucial to executing this transparent approach.