1. Introduction: approaching the data wall

Today’s artificial intelligence (AI) technologies have the potential to render all cultural heritage preserved in archives into training data by ingesting digitized and digital collections. As AI systems consume larger and larger caches of digital data to fuel their training and development for deployment, firms like Anthropic and OpenAI are confronting what tech evangelists, journalists and researchers have called “the data wall” (Roose Reference Roose2024). For AI firms, the data wall is the boundary of high-quality, machine accessible, digital data published and openly available to be scraped from the internet (Villalobos et al. Reference Villalobos, Anson, Sevilla, Besiroglu, Heim and Hobbhahn2024). For archives, libraries and scientific data repositories, the scarcity of online machine accessible data resources augurs a deeper tension to be reckoned with. Simply put, most cultural heritage and historical records remain out of reach to AI systems because they are not yet digitized, machine-accessible or published online. By the same token, most contemporary institutional collections preserved by archives, libraries and repositories are discoverable online with digital surrogates by actively positioning their collections as data under the banner of open access (MacDougall and Ruediger Reference MacDougall and Ruediger2024). Further, institutions now document collections and provide access points via representational information like metadata, catalogue records or finding aids. All are available online and can be used as training data, too.

Speaking in response to Harvard’s recent training dataset release of public-domain books for model development (Knibbs Reference Knibbs2024), Meta’s former Chief AI scientist and Advanced Machines Intelligence Labs founder Yann LeCun posted, “Every institution, library, foundation, cultural group, and government around the world that possesses cultural content should make it available for training free and open AI foundation models” (LeCun Reference LeCun2024). LeCun’s proposal urging institutions to release content squarely illustrates how the largest tech firms and AI startups position cultural heritage collections like those preserved in archives as potential sources of data to drive AI development. However, the rush to make collections available as training data may risk subordinating traditional archival access models to serve tech firms’ interests, enclosing and inhibiting the documentation of AI for future archives as well.

Once archival collections are conceived as data, their digital representation information becomes potential resources to both power AI as training data and to leverage it for future applications, including processes that result in more outputs, as both records and data. This article proposes a framework for understanding archival materials at the conceptual boundary of machine accessibility. In my argument, much of this data (and metadata) exists before, beyond or outside a euphemistic “training window.”Footnote ¹ I adapt this concept from machine learning to describe the boundaries of what AI systems can access, process and produce. Throughout, I use the training window as a conceptual metaphor for the threshold between archival materials and AI systems that computationally consume records and those that remain beyond their reach.

This approach intentionally echoes the historical and critical archival practice of reading archives against the grain to uncover how power is asserted in the preservation and access of records. Here, I use the preposition against as a heuristic to illustrate the problem of how AI writ large (that is, big tech firms, systems, even well-meaning researchers and policy implementations) conceive of archives as extractable data resources, accessed to train models in coercive directions for deployment in AI applications and services. Below, I propose a three-part conceptual framework to investigate how archival records are subsumed by AI through the lens of reading data against the training window: first, by understanding records as processual, and how they are subsumed when accessed as training data; second, examining the infrastructural power that determines the preservation and access of records in AI; and third, focusing on how AI data acquisition processes obscure provenance information, destabilize metadata and create new forms of archival silence in producing new records and data.

To read data against the training window is to investigate archives and representation information that has yet to be published to the internet or that which has not already been scraped for AI deployment, such as large language models (LLMs). That which is against the training window draws our attention to oppositional forces at play in subsuming archives as data for machine access and actionability, including institutional, technical or epistemological constraints. While libraries, archives and museums (LAMs) have been digitizing collections and making them accessible online for decades, transforming collections by reformatting them into datasets for bulk download or API access is still underway (Cordell Reference Cordell2020). The need for more available sources of training data for AI has begun to put pressure on institutions providing open access to collections online. Robust metadata and structured data access points from digital collections and catalogues are attractive resources for LLM development. However, once ingested into training corpora, collections’ data are separated from their original contexts that carry rights statements or persistent identifiers.

Recently, LAMs have reported an uptick in traffic overwhelming their online services from AI bots scraping institutional collections to build training datasets (Maiberg Reference Maiberg2025; Weinberg Reference Weinberg2025). Duke University Libraries has told its campus community that “the overwhelming number of simultaneous requests that these AI bots make on our websites has sometimes rendered them inaccessible to our human library patrons” (Duke University Libraries 2025). Similar experiences have been reported by the University of North Carolina Libraries (Panitch Reference Panitch2025), where library IT staff now limit request rates in order to preserve the stability of their online access systems. As we come closer to facing the data wall, more bottlenecks are predicted by AI influencers like Leopold Aschenbrenner (Reference Aschenbrenner2024). Moreover, a divide is emerging between online archival collections and the extensive holdings in archives that are not yet machine accessible as data (Jaillant et al. Reference Jaillant, Mitchell, Ewoh-opu and Urbaneja2025; Jaillant and Rees Reference Jaillant and Rees2023). As AI firms’ scraping comes under more legal scrutiny, the tension between open access ideals and extractive data use is coming to the fore (Field Reference Field2025).

These institutional disruptions are at the level of collections infrastructure, but the consequences for archival records and knowledge production are equally concrete. Consider what it means to encounter one’s own published scholarly record as part of an output from AI training. When peer reviewing a manuscript recently, I discovered a number of citations to my own scholarship that were inaccurate in some peculiar ways. Much of the manuscript’s prose appeared to be presenting relevant analysis with recognizable throughlines to my research, but nearly half of the citation elements listed in the bibliography were wrong. Facts were misattributed to specific publications, familiar sounding but fake journal titles were transposed, articles (that I would have liked to have published) were invented and listed in the reference list next to correct citations. The pastiche is legible once you witness the two processes built around the training window at play. First, a body of published work is scraped from open repositories and ingested as training data, where individual publications lose the provenance information that distinguishes one source from another. AI systems then generate combinatory outputs that may format citations properly but camouflage inaccuracies by blending in reliable information sources and severing the original context.

The task at hand for critical archival studies scholarship is to trace how archival collections travel (or fail to travel) across the threshold of training windows into AI systems, and once records have been incorporated as data, to trace how they pass and carry evidential meaning. Following archival records as data through thresholds of AI systems to the infrastructure that supports AI means examining their transmission from archives to online access points and, once scraped, studying training datasets by examining the transformations and recontextualizations that occur in using this technology.

Examination of records in AI should also involve gathering evidence of their erasure, decontextualization, or the cascading, powerful silences well known in archives and documented by scholars (Burton Reference Burton2008; Han et al. Reference Han, Han and Wickett2024). First, what and when is a record in AI systems? For both archivists and scholars who study the work of preserving human culture, the record is an unstable appellation and hard to define, especially in a time of networks and near-constant data transmission. In previous work, I have argued that the question “when is a record?” offers a more productive approach than defining what a record is in systems, particularly for understanding how born networked digital records like emails that move through infrastructure during their creation, transmission and receipt (Acker Reference Acker, Gilliland, McKemmish and Lau2017). This approach emphasizes examining the layers of infrastructure that shape records formats, the networked recordkeeping practices that govern their movement across systems and the information retrieval systems that condition their access through time.

Each of these elements of transmission and reception remains relevant when considering how AI systems access, acquire and process archival materials, incorporating them as information sources and using them as training data. In defining what and when a record is for archives-driven AI, we must extend our analysis to thresholds or boundaries that decompose records, destabilize metadata and subsume collections as data. Specifically, I argue that the training window as a scene of origination determines what data, information and archival materials become legible to machine learning systems and those which remain incommensurable or out of reach.

Indeed, AI infrastructure creates unique junctures of transmission where the integrity, provenance and contextual meaning of records may be compromised or lost entirely. Such concerns parallel related accounts from critical archives scholars examining how digital data, algorithmic systems and web archives challenge traditional notions of evidence and record-ness in internet infrastructure (Maemura Reference Maemura2023; Summers and Punzalan Reference Summers and Punzalan2017; Trace and Hodges Reference Trace, Hodges, Huvila, Andersson and Sköld2024). Reading how archival records become data resources for AI requires close study of the infrapolitics that enable (or resist) records’ transformation to digital data (Poirier Reference Poirier2021; Thylstrup Reference Thylstrup2019). As such, I argue that investigating collections that exist “against the training window,” or archival collections that are poised to become data for machine actionability, stands as a critical tool for understanding AI’s limits in handling cultural heritage. This conceptual framework also offers a means of historicizing our current moment of massive AI adoption. While AI technology may fail to integrate foundational archival concepts concerned with documentation such as provenance, authenticity and appraisal, some have argued that AI designers and researchers should turn to archival first principles (Jo and Gebru Reference Jo and Gebru2020), so we can better understand the implications of treating archival collections as data as they become potential resources for machine action and training data.

The notion that archives can be data (or perhaps have always been data) stems from a longer history of rendering institutional collections of information as potential sources of big data (Acker Reference Acker2025; Kitchin Reference Kitchin2014). Beginning in the early 2010s, LAMs began to characterize their collections as digital data. As digital collections programs’ offerings began to mature, a more formal “collections as data” paradigm emerged in the United States as a Mellon Foundation-funded project network for computational analysis and access to LAM holdings (Lampert and Lapworth Reference Lampert and Lapworth2020; Padilla et al. Reference Padilla, Allen, Frost, Potvin, Roke and Varner2019). The collections as data approach promotes standardization and interoperability of digital collections while aiming to democratize public access online, but it can also create asymmetric forces that delimit access in unanticipated ways. Science and technology studies scholars have noted that as cultural institutions invest in making their collections machine-readable and digitally accessible as knowledge infrastructure, large technology firms stand to benefit most from such public investments in open data infrastructure (Acker et al. Reference Acker, Struett, Tian and Finn2026; Mirowski Reference Mirowski2023; Plantin and Thomer Reference Plantin and Thomer2023).

In the US context, open data initiatives from government and research agency policies occurred simultaneously alongside the Obama administration’s efforts at engaging with corporate technology firms to spur economic innovation and digital solutions for government bureaucracy. Achieving open data access to drive innovation and increase collections engagement has been recognized by archives and libraries as an opportunity for partnering with private firms (Kriesberg Reference Kriesberg2017; Shankar et al. Reference Shankar, Eschenfelder, Buchholz, Cullen, Taylor, Christian-lamb, Martin and Nardi2019; Susha et al. Reference Susha, Boriana, Gil-garcia, Tan and Gasco2019). But public-private partnerships for digital data access can also be a challenge. Companies developing digital collections and AI applications can use and reuse newly accessible collections as training data without contributing proportionally to the substantial costs of digitization, preservation, maintenance or responsible stewardship (Murray-rust Reference Murray-rust2012). Major technology companies like Microsoft, Google and Amazon have positioned themselves to profit from both cultural heritage and scientific research made accessible through public institutions’ digitization and open access initiatives (Chalmers and Edwards Reference Chalmers and Edwards2017; Kriesberg Reference Kriesberg2017). Notably, these firms leverage institutional investments in public access infrastructure while implementing their own restrictive access controls through APIs, proprietary platform services and commercial licensing that then create asymmetric relationships between public institutions, researchers and corporate interests (Acker and Kriesberg Reference Acker and Kriesberg2020; Nesmith Reference Nesmith2023; Plantin et al. Reference Plantin, Lagoze and Edwards2018). This asymmetry is a key distinction between institutional open access to archival collections of records and the corporate reuse of data drawn and assembled from records preserved in archives and scientific repositories.

The next section offers some more background on definitions of records from archival science, including important related concepts. Following this theoretical background, I outline the framework with some case studies drawn from generative AI (GenAI). The selected examples are meant to both illustrate challenges and opportunities for critical researchers concerned with archival issues of evidence and accountability. Given the conceptual aims of this paper, these examples are not meant to be exhaustive, but modular and interoperable. Following the “supply chain” analogy used by critical AI researchers (cf. Crawford Reference Crawford2021; Posner Reference Posner2021; Widder Reference Widder2024), these expository cases can be easily swapped to shift attention towards alternative logistics components. I conclude by discussing how this heuristic can be applied to critical studies of AI access regimes and digital cultural heritage. A broader goal is to stimulate awareness of issues that cultural heritage institutions face in maintaining their access and preservation missions as they are engaged by tech firms that rapaciously rely on public access to archival collections configured and used as data.

2. Theoretical background: archival records, thresholds and training windows

Apprehending archival records in contexts of AI requires first acknowledging that most records created today are pushed and pulled across digital networks. Email, social media, digital health records and mobile payments are all digital and born in networks. As born networked records move through systems, they simultaneously acquire new layers of materiality and information governed by algorithms that transform the evidence they carry forward. Records in digital environments have always challenged archivists in terms of their boundaries, because as they age and move through infrastructure, they acquire layers of context and more meaning (McKemmish et al. Reference McKemmish, Upward and Reed2009). In addition to acquiring layers of context through transmission, software-dependent records can also be technically transformed during initial access contexts and during preservation via reformatting, migration or emulation.

The archival threshold has traditionally marked the point where records enter an archive’s custody and control (Duranti Reference Duranti1997). This threshold has historically functioned as a site where authority is asserted through physical custody, where archivists govern what records are preserved and how they are accessed. But such thresholds have become more complex in digital environments where custody over digital and digitized collections can be distributed across cloud infrastructure, multiple places and stakeholders (Borgman Reference Borgman2015; MacNeil and Eastwood Reference MacNeil and Eastwood2017, p. 67).Footnote ² Thus, the perspective of records always becoming and having processual qualities of evidence is helpful for understanding when records acquire dimensions of context by passing across thresholds, becoming archival data. Tracing how archival records cross (or fail to cross) over thresholds like training windows requires reading both the technical infrastructure of AI systems and the social, economic and political forces that determine what materials become datasets (Poirier Reference Poirier2021; Stevens and Keyes Reference Stevens and Keyes2021; Thylstrup Reference Thylstrup2022). When records in networks simultaneously exist in multiple places and durations, traditional archival notions of authenticity and provenance are challenged (consider webpages published to the internet, social media posts or text messages sent via cellular networks). These boundary issues are amplified in AI as the transformation of records into training data may be stripped of layers of evidential context, combined with other data sources, reassembled and processed through algorithmic systems that distance records from their original creation and preservation contexts. The access regimes leveraged by AI firms bypass this protective function of archival custody when extracting records and using them as resources for training models. Training windows within AI models introduce a threshold of custody that parallels the archival threshold of accessioning. However, while archival epistemology centres on maintaining provenance and original order of records across time, as discussed below, AI epistemes operate with different relationships to preserving context and chains of evidence.

A second way of defining records in archives is through the modes and means of accessing them after they have been designated for preservation to support future access. The Society of American Archivists’ Dictionary of Archives Terminology defines “access” as the ability to locate relevant information using online catalogues, search indexes, finding aids or other access points for retrieval (Society of American Archivists 2025). Since the rise of automated recordkeeping systems and electronic records in the 1980s, archivists have recognized that when we separate the storage of records from their representation information in systems that support online access, we begin to create a tenuous distance between records, metadata and their descriptive context. This is a distinction in point: while archival collections themselves may be analogue, full of paper and material artefacts, their representation information and descriptive metadata that is used for retrieval and online access is digital, structured data. Consider the notebooks of Octavia Butler housed in her archive, for example. While the physical materials are stored at the Huntington Library, the access points that describe these documents for retrieval have been published on the open web for decades and in different forms (Jamieson Reference Jamieson2017). This distance between the physical collections and their digital access points online is often obscured but should be attended to closely. Such distancing is technical, but it is also ideological, reflecting a fundamental limit in how we preserve digital information and create representation information to support access that is increasingly reused by AI firms as training data.

What AI systems produce is shaped by the records acquired and incorporated during their model design. LLMs and AI-driven tools are trained on finite corpora, constrained by copyright, data availability and the epistemological assumptions underlying their curation and assembly for access and reuse (Desai et al. Reference Desai, Pasquetto, Jacobs and Card2024). By design, these models operate on preselected datasets that omit records that have not yet been digitized, structured or otherwise made machine-actionable and published online. This limitation creates a reality where AI is disproportionately trained on more contemporary, born-digital materials and digitized works that are out of copyright or publicly available on the internet. Moreover, such context windows define the span of tokens involved in generating responses from LLMs. To be sure, the context window is distinct from archival context, but both set boundaries on relevance and interpretation. Instead of preserving evidence and original order across time, AI context windows function within narrower temporal frames than archives, often without reference to provenance information, or reliable metadata, thereby obfuscating context even further. While publicly available data are known to be leveraged by AI firms, documentation about how it is accessed and reused is not easily available downstream (Cole Reference Cole2024; Mehrotra and Couts Reference Mehrotra and Couts2024).

Training windows introduce another threshold for records that complicates traditional archival thresholds of accessioning collections and asserting custody. Thresholds for training data are different from archive thresholds because AI systems themselves tend to obfuscate the provenance of their assembly and restrict information about data acquired for training, where preserving provenance and context are the foundational archival functions of representation and access. For example, when GenAI platforms are asked to share more information about the data sources found in their model architecture, answers range from facile to the evasive: “The training data for my model, GPT-4, consists of a mixture of publicly available text, licensed data, and data created by OpenAI to be safe and useful. However, OpenAI has not publicly disclosed the full details of the datasets used, including the specific sources, proportions, or licensing agreements.”Footnote ³ To evasive: “Anthropic (the company that created me) hasn’t publicly disclosed the exact composition or specific sources of my training data. This approach is common among AI technology development, as training data details are often considered proprietary information.”Footnote ⁴

Another consideration is the various records manifested as outputs from AI systems themselves. Common records from AI systems include search answers, hallucinations, citations, summaries of sources and probabilistic outputs (Sundin Reference Sundin2025). A gallimaufry of records, to be sure. Both types – records that drive AI as data sources and records produced from AI – inhabit two separate forms of documentation, each with distinct challenges for archivists, researchers and technologists. I belabour the distinctions between archival records used as data for AI and records produced from AI to illustrate the challenge of defining evidential relationships without firm thresholds, metadata and provenance information that are paramount to archival science.

3. A framework for reading archival records as data in AI

GenAI tools like ChatGPT introduce grammars of action that distance users from the records by destabilizing their boundedness. AI firms like Google and Anthropic extend this destabilization by obscuring provenance information, too. Such barriers can be found in algorithmic processing, proprietary access controls and the flattening of contextual information that occurs when records are transformed as data sources to power AI tooling. What we are witnessing is a kind of trick mirror where data training processes in AI subsume records as data while separating records from original contexts such as archives or the public internet, then transform them into proprietary resources within black-boxed technical systems (Heikkilä Reference Heikkilä2024; Knibbs Reference Knibbs2024). The following framework aims to provide a set of strategies for how critical researchers might unpack and investigate such distancing and destabilization of archival records in AI.

4. How to use this framework

Kate Crawford argues that AI is “a form of exercising power, and a way of seeing; it’s also a manifestation of highly organized capital backed by vast systems of extraction and logistics, with supply chains that wrap around the entire planet” (Crawford Reference Crawford2021, p. 18). The three-part conceptual heuristic I have outlined above serves as an investigative tool for mapping the archival power dynamics of AI systems by tracing how records move through technical infrastructures like a supply chain. The framework can be applied to describe technological and ontological changes from records that AI technology accesses as data, reuses, destabilizes and co-creates. Each transformation across the supply chain shifts the status of archives’ custody and control, demanding critical engagement from scholars concerned with technosolutionism enacted by AI supply chains moving archival records from the past and carrying them forward as data.

As AI systems acquire more and more archival collections for training data and beyond, they begin slowly to assert a kind of custody and authority over reuse of the human record in access regimes. The growing number of lawsuits against AI firms that allege improper use of data scraping speaks to these shifting access expectations (Grynbaum and Mac Reference Grynbaum and Mac2023). An epistemic mark of AI systems is the destabilization of metadata and the attribution of information sources accessed as data for combinatorial reuse. Records exist simultaneously in multiple states and locations, with varying degrees of contextual integrity maintained across these systems. In other words, for AI firms, questions about “what and when is a record?” can only then be answered by asking “who controls access to this record and how is it reused?”

AI firms intensify this distributed custody by introducing more layers of technological mediation in access and reuse. For example, DALL-E 3 and Midjourney generate images based partly on digitized cultural heritage, creating a form of distributed custody where the chain of control stretches across multiple actors: the creators of records collected for preservation, the originating archives that digitized the collections, the institutions that made them available online with representation information and structured metadata, the AI developers who incorporated them into training data and ultimately the users who prompt the system to generate derivative images.

In this extended chain of records co-creation, no single entity (creator or organization) maintains control over how cultural materials are represented with GenAI. And though original institutions may retain custody of physical artefacts, they lose control over collections’ reuse once they cross the threshold into training data for AI. Recent research framing GenAI as context machines acknowledges this distributed interpretive effect of meaning-making (Kommers et al. Reference Kommers, Ahnert, Antoniak, Benetos, Benford, Bunz and Caramiaux2026) but stops short of accounting for whose institutional responsibility it is (and who benefits) when such provenance is severed. This distributive effect transforms open access into a spectrum of control where reuse (sans provenance) dominates access goals that operate across the infrastructure of AI systems.

As technology companies like Meta, Google and OpenAI firms ingest, process and operationalize archival records as data resources, they become mediators of cultural memory, which is further leveraged for AI outputs. Just as archival records are being accessed as training data, they are (re)assembled into assets that can be extracted, transformed and monetized through AI applications and subscription services (Birch et al. Reference Birch, Cochrane and Ward2021; Weinberg Reference Weinberg2025). Such neoliberal priorities stand in sharp relief to the archival values of preservation to ensure authenticity, historical accountability and evidence of cultural memory.

Much of the current work in critical AI studies documents bias, harm, environmental impact, labour and policy efforts (Hao Reference Hao2025; Pasquinelli Reference Pasquinelli2023). For archival institutions navigating this shifting landscape of uptake and automation of collection into the AI supply chain, the framework I propose offers a means of reasserting agency by following the processes and junctures at which records move as data from archives into AI systems for reuse. Understanding records as processes rather than static objects allows archivists to trace how contextual information is transformed or lost when records cross the threshold and become AI training data. Even so, it remains unclear how we might document AI outputs as records for future archives. Analysing infrastructural components that drive such dynamics reveals leverage points where institutions might intervene to maintain greater control over how their collections are represented in AI systems and beyond (Lynch Reference Lynch2017). To be sure, studying the provenance challenges that AI supply chains manifest helps identify where traditional archival documentation practices can be extended to address the contemporary conditions of human–AI interaction. As above, these provenance challenges point to an AI episteme that privileges combinatorial data processes for inference over historical documentation of data journeys (Bates et al. Reference Bates, Lin and Goodale2016).

Applying this framework and reading data against the training window does not mean opposing AI technologies altogether in archives and cultural heritage collections. Instead, it suggests an orientation that recognizes the limitations of AI’s temporal and contextual boundaries and works to guard that which exists beyond those boundaries.

For decades, archivists and cataloguers have explored the possibilities of machine learning to support appraising collections for archival preservation and automated metadata generation (Fishbein Reference Fishbein1972; Greenberg Reference Greenberg1998).Footnote ⁵ Digitization programs in LAMs have already incorporated machine learning in their work processing collections for description, preservation and access (Jaillant and Caputo Reference Jaillant and Caputo2022; Neudecker Reference Neudecker2022). Optical Character Recognition transcription is perhaps the most widely adopted to transcribe both printed and handwritten documents as part of large-scale digitization efforts (Nockels et al. Reference Nockels, Gooding and Terras2024), while natural language processing techniques are used for sensitivity review to redact personal information when processing collections like email or government documents (Hutchinson Reference Hutchinson2020). Following on many practitioners’ calls, the application of AI in archives should involve the development of collaborative governance models that integrate archivists into AI development processes, the establishment of documentation standards that preserve contextual integrity across AI supply chains and concerted efforts at the articulation of archival practices that deliberately designate records that should never be accessed as data by machines.

The archival profession faces a complex tension between adopting AI technologies and maintaining custody by controlling access to collections. As Cliff Lynch eloquently argued, the question is not whether AI will transform archival practice; instead, it should be how archivists will respond to this transformation – either by accepting the terms set by tech firms in valuing accessible data or by asserting archival values in the infrastructure of AI itself (Lynch Reference Lynch2017). By examining real-world examples of scraping abuses for training data and AI-infused archival processing for automated metadata description, we can see that AI technologies are already reshaping how cultural heritage is accessed, interpreted and reproduced with GenAI offerings.

This challenge is exemplified in current AI policies at the US National Archives and Records Administration (NARA), which restricts employees from using some GenAI tools while co-developing others as part of a broader AI pilot project initiative (U.S. National Archives and Records Administration 2024). In May 2024, NARA announced a ban on employee use of commercial GenAI tools such as ChatGPT across agency networks and devices. This prohibition stems from legitimate access concerns about data security and sovereignty. In a memo to staff, “NARA has determined that ChatGPT’s unrestricted approach to reusing input data poses an unacceptable risk to NARA data security” (U.S. National Archives and Records Administration 2025). Specifically, inputs of records and archivists’ prompts could be incorporated into training data by GenAI tools, potentially exposing sensitive or non-public information without any mechanism for subsequent removal or redaction. The archives’ restriction policy on GenAI indicates a growing awareness of how commercial AI systems function as data extraction mechanisms without limitations on reuse. Paradoxically, NARA has partnered with Google to develop and launch ArchieAI, a public-facing chatbot, at the same time (Koebler Reference Koebler2024). This initiative is part of the federal agency’s strategic effort to harness AI to improve accessibility to NARA’s vast holdings of federal information. This contradiction illustrates the pressure of AI adoption that most LAM institutions are facing.

As archives and cultural heritage institutions explore AI’s potential for enhancing discovery and access, they recognize that unrestricted data flows to commercial systems likely undermine their stewardship missions and could risk violating existing data security laws by allowing unrestricted reuse. A critical perspective on AI and archives requires understanding both the potential and limitations of AI in archival contexts of labouring practitioners, preserving control over how collections are made accessible. An opportunity for archival practitioners moving forward is to develop preservation frameworks that can leverage AI’s capabilities while maintaining autonomy and protecting the integrity of collections in the face of AI firms’ acquisitive drive for subsuming archives as data.