Digital collections & discovery

Discovery platforms have used machine learning for relevance ranking, query expansion, and faceted refinement for well over a decade. Primo, EBSCO Discovery Service, Summon, and WorldCat Discovery all use algorithmic relevance models that have been refined through usage data. When a patron's search returns results ranked by relevance rather than date or call number, that ranking reflects a machine learning model, even if the platform does not describe it that way. To understand where your discovery layer stands, distinguish features that have been present for years from genuinely new generative AI capabilities. The appropriate professional actions for each are quite different.

The newer generation of discovery AI includes semantic search, natural language query processing, and AI-generated summaries of search results. Semantic search interprets the meaning of a query rather than matching its exact terms. A patron searching for "how stress affects memory" retrieves conceptually relevant results even when sources use different vocabulary, such as "cortisol and hippocampal function." Harvard Library's Collections Explorer, launched as a public beta in September 2025, provides a concrete institutional example. It lets researchers explore Harvard's archival and special collections through natural language queries, combining semantic search with generative AI to surface relevant records even when the catalog description does not use the researcher's words. This shift in patron experience, from constructing search terms to conversational inquiry, is not merely an interface change. It reflects a fundamental change in how the discovery layer interprets patron intent.

OCLC added AI-assisted classification and subject heading suggestions to its WorldShare Record Manager and Connexion cataloging applications in December 2025. The feature draws on WorldCat's hundreds of millions of bibliographic records to suggest Dewey Decimal Classification numbers, Library of Congress Classification numbers, and Library of Congress Subject Headings at the point of cataloging. It illustrates the direction all major platform vendors are moving: AI embedded directly in the professional workflow rather than offered as a separate tool. Ex Libris, for its part, launched Specto, an AI-powered platform for digital collections management that combines AI metadata creation, bulk editing, and discovery enrichment in a single integrated workflow.

Specto is now generally available, alongside a companion tool that deserves the same attention: the AI Metadata Assistant for Alma. It works inside Alma's cataloging interface to draft descriptive metadata, suggest subject headings, and flag likely duplicate records during import and ingest. Both tools sit under Ex Libris's broader Academic AI platform, the umbrella under which the company is consolidating the AI capabilities it is building across Alma, Primo, and Rosetta. For digital collections librarians at institutions running Ex Libris products, general availability changes the relevant professional question. Specto and the AI Metadata Assistant for Alma are no longer features to watch for in a future release. They are features to request access to, configure for a pilot collection, and evaluate against the audit and pilot framework this module returns to later. A meaningful evaluation requires knowing what the tools actually do with a specific institution's records, and only a pilot on local data can answer that.

The professional limitation that matters most in this landscape is the difference between vendor-controlled AI features and locally configurable ones. Many discovery platform AI capabilities are set by the vendor and applied uniformly across subscribing institutions. A digital librarian cannot adjust the semantic model or the relevance weights directly. What can be configured locally varies by platform and contract: relevance tuning profiles, query expansion settings, local boosting rules, and which AI features are exposed to patrons. To learn what is actually available at your institution, start with your platform's current release notes. Ask your vendor representative specific questions about the AI features enabled in your subscription tier. Request a demonstration of any semantic search or AI summary capabilities that are available but not yet turned on. Advocating for specific features, rather than passively accepting the platform's defaults, is a genuine professional function. It improves patron experience in ways the vendor cannot achieve without institutional input.

The practical difference between keyword search and semantic search shows up in patron behavior, even when patrons cannot articulate it. A patron who types a natural language question into a semantically capable discovery layer and gets relevant results is experiencing something fundamentally different from Boolean search, even if the interface looks similar. That patron does not need to know controlled vocabulary, translate research questions into search terms, or apply Boolean operators. For patrons who have always sought information through conversational search engines, dropping that translation step is significant.

Yale Library has developed a prototype Digital Collections AI application that applies large language models directly to digitized texts. It performs summarization, translation, named entity extraction, and question answering on items from Yale's digital collections. For example, a researcher working with digitized historical correspondence could ask the system about the letters' content and receive a synthesized answer drawn from the transcribed text, instead of reading every letter to find a specific reference. That capability transforms the researcher's relationship to a large, underdescribed collection. Instead of navigating finding aids and reading item by item, the researcher can interact with the collection's content directly. Yale is actively seeking instructor partnerships to test the application, so libraries investigating similar capabilities have a peer example to examine and a potential collaboration to pursue.

Knowledge graphs are a related but distinct approach to discovery enhancement. Tools like Yewno connect concepts across resources through semantic relationships rather than keyword matching. A patron can follow a concept through related ideas, disciplines, and sources in ways that traditional subject heading navigation does not support. Primo's "Topics" feature uses a similar logic, surfacing conceptual relationships between a search result and connected ideas based on how concepts are linked across the literature. Relational discovery is particularly useful for interdisciplinary research, where a concept in one discipline connects meaningfully to literature in adjacent fields that may not share controlled vocabulary.

The professional responsibility that semantic search creates is instructional rather than technical. Patrons who use natural language discovery tools successfully may conclude that AI search is more comprehensive than it is. They reason that if the tool did not surface a source, the source does not exist or is not relevant. To keep this misconception from affecting research quality, instruction should address the difference between semantic discovery and comprehensive bibliographic retrieval. AI can surface likely relevant results, but it cannot guarantee completeness the way a systematic database search with explicit inclusion criteria can. Semantic search results also require the same evaluation as keyword search results. Relevance ranking does not confer authority, and a result at the top of an AI-ranked list is not more credible than one lower down. These points belong in any session that introduces patrons to a semantically capable discovery platform.

The shift from keyword search to AI-mediated discovery is not only a usability improvement. It changes who, or what, decides which materials a patron sees and in what order, which makes it an intellectual freedom question as much as a technical one. ALA's AI guidance names Intellectual Freedom as a core value precisely because AI systems threaten it when they, in the guidance's words, "obscure how information is ranked, extract public or library-created data without clear public benefit, or optimize engagement at the expense of attention." A discovery layer whose relevance model is a vendor-controlled black box is doing exactly the first of these. It ranks the collection by criteria neither the librarian nor the patron can inspect. A patron who does not know why certain results rose to the top cannot reason about what the ranking might be burying. Taking this value seriously means treating the discovery layer's ranking not as neutral plumbing but as an editorial act that requires accountability.

Three concrete obligations follow from the guidance, and each is actionable at the level of discovery configuration and instruction. The first is transparency. The guidance calls for "clear, accessible documentation" of an automated system's "function, purpose, accountability, and influence on user outcomes." A library should therefore be able to tell patrons, in plain language, that discovery results are algorithmically ranked and personalized. It should also press its vendor for enough explanation of the ranking model to make that disclosure honest. The second is the preservation of predictable, non-personalized paths. The guidance asks libraries to "preserve predictable, transparent, non-personalized, and explainable access paths alongside AI-enhanced discovery features." In practice, that means keeping a straightforward, non-AI way to browse and search the collection, rather than replacing it entirely with a personalized, engagement-optimized interface. For example, a patron who wants to see everything the collection holds on a topic, in a stable and reproducible order, has a legitimate research need. A purely semantic, personalized discovery layer can actively frustrate that need, and preserving a non-personalized path protects it. The third is opt-out. The guidance holds that "users should be able to opt out of AI-automated products when possible" and that libraries "will require vendors to offer equivalent services that do not include AI assistance, or document the limitation and identify alternatives." That turns the availability of a non-AI route into a procurement question to raise before adoption, not a feature the library hopes the vendor happens to provide.

The most distinctly intellectual-freedom obligation, and the easiest to overlook, is the audit for suppressed visibility. The guidance asks libraries to review "AI-enhanced discovery, ranking, recommendation, and personalization features" on a scheduled basis "for patterns that reduce visibility, distort representation, or limit access for particular topics, authors, viewpoints, languages, or communities." This is a different exercise from evaluating relevance quality. A relevance audit asks whether the top results are good. A visibility audit asks whether certain materials, viewpoints, or communities are being pushed so far down the ranking that they are effectively invisible, however good the top results look. For example, a discovery layer that consistently ranks recent, high-usage, English-language materials first may, without anyone intending it, bury older holdings, minority-language materials, and low-circulation works documenting marginalized communities. Only a deliberate audit that looks for what is hidden, rather than what is surfaced, will catch it. This audit connects directly to the bias-auditing practice that Module 05 describes, and it shares the same premise: a pattern that disadvantages certain materials is invisible in any single search and detectable only across many. The guidance does not turn these obligations into a checklist a system can satisfy on its own. Whether a discovery configuration honors intellectual freedom remains a professional judgment: whether its ranking is accountable, its non-AI paths genuine, and its visibility patterns fair. That judgment belongs to the librarians responsible for how their community reaches the collection.

The description gap in archival collections is one of the most persistent practical challenges in digital library work. Collections digitized under grant funding, often on tight timelines with minimal staff, frequently have finding aids with collection-level scope notes and series descriptions but no item-level description at all. These collections are technically accessible but practically invisible. A patron who cannot find an item through description cannot know it exists, however thoroughly it has been digitized. AI description assistance addresses this gap at a scale that manual remediation, at existing staff and resource levels, cannot match.

The workflow for AI-assisted finding aid enhancement builds on existing content rather than generating from scratch. For a collection with a collection-level scope note, series descriptions, and digitized item images or transcriptions, AI can do three things. It can draft plain-language summaries of each series, generate patron-accessible overviews of complex intellectual arrangements, and produce item-level descriptions from provided transcriptions. For example, an archivist working with a digitized correspondence collection can give AI a letter's transcription and ask for a three- to five-sentence catalog description identifying the sender, recipient, date, primary subject, and any significant details a researcher would find relevant. That description is a draft. It requires archivist review for factual accuracy, appropriate terminology, and adherence to archival description standards. But it is produced in seconds rather than minutes, a substantial per-item gain across a collection of several hundred items.

AI's particular strength in finding aid work is translating archival language into patron-accessible prose. Finding aids written in traditional archival conventions, with terms like "accruals," "extant," "bulk dates," and "provenance," are opaque to most patrons without archival research training. For example, giving AI an existing scope and content note and asking it to rewrite the note for an undergraduate who has never visited an archive produces a genuinely more accessible document. The information is the same, presented in a different register. Plain-language rewriting preserves the intellectual content of the description while greatly expanding the number of patrons who can interpret it.

Two categories of archival material require heightened professional care. The first is collections containing records about living individuals, including recent institutional records, community archives, and personal papers with donor-imposed restrictions. These need careful review to ensure AI-generated descriptions do not disclose information beyond what the access policy permits. AI does not inherently know which details in a document are sensitive; the archivist's judgment about appropriate description remains the only reliable filter. The second is collections documenting Indigenous communities and cultural materials. Their description should reflect the originating community's preferred terminology and the collection's access conditions, which may be governed by the community rather than the institution. AI trained on mainstream archival conventions will not automatically apply Indigenous-centered description principles. Those principles must be provided explicitly in the prompt and verified in the output. Any finding aid content generated with AI assistance should also be disclosed in the collection's processing note, both for professional transparency and as documentation for future revision.

Most digital collections programs carry substantial accessibility debt, largely invisible until a patron with a disability encounters it. Audio and video materials without transcripts are inaccessible to deaf and hard-of-hearing users. Images without alt text are inaccessible to blind and low-vision users who rely on screen readers. Digitized text with poor OCR is inaccessible to users who depend on text-to-speech tools. For public colleges and universities, the Department of Justice's 2024 rule under Title II of the Americans with Disabilities Act makes WCAG 2.1 Level AA the required standard for web content, digital collections included. An April 2026 interim final rule moved the compliance dates to April 26, 2027, for public entities serving populations of 50,000 or more and April 26, 2028, for smaller ones. Among its reasons, it cited the limits of using generative AI to remediate content at scale. Private institutions fall outside that rule but carry accessibility obligations under Section 504 of the Rehabilitation Act and Title III of the ADA. Despite good intentions, most digital collections programs have not produced the accessible formats compliance requires, because the manual labor exceeds available staff capacity.

AI transcription has become reliable enough to be a practical tool for production workflows, not an experimental one. OpenAI's Whisper model, which is open source and free to run locally, produces transcripts from audio files accurate enough for most library applications after human review. Oral history interviews may run several hours each, and manual transcription would require paid professional services or substantial volunteer labor. For these collections, AI transcription turns an economically prohibitive task into a review-and-correction workflow. For example, a digital librarian managing fifty oral history interviews, each averaging ninety minutes, faces approximately seventy-five hours of manual transcription. Running the recordings through Whisper and reviewing the transcripts takes considerably less time, particularly for recordings with good audio and standard English speech patterns. Lowering the labor barrier this way is what makes accessibility compliance achievable rather than permanently deferred.

Accuracy varies predictably with audio quality, speaker accent, technical vocabulary, and background noise. Recordings with poor audio, multiple simultaneous speakers, strong regional accents, or specialized disciplinary terminology will need more substantial review and correction. That variation is a reason to pilot transcription on a representative sample before committing to a production workflow. A pilot lets the digital librarian calibrate the review time a specific collection requires before planning a full-scale project. For collections in languages other than English, both Whisper and commercial tools like Otter.ai and Rev AI perform variably, so evaluate them against the specific language before adopting the workflow.

AI alt text generation addresses a second major accessibility gap in digital collections. For photographs, maps, illustrations, and digitized artworks, AI can generate descriptive alt text that screen reader users receive in place of the visual content. The practical workflow is to give the image to a vision-capable AI model and request a description suited to the image type and patron use. For example: "Write alt text for this photograph from a 1950s university yearbook. Describe the image content precisely in one to two sentences. Do not interpret or editorialize; describe what is visible." The digital librarian then reviews the draft alt text for accuracy and an appropriate level of detail. The review step is essential. AI image description can misidentify individuals, misread text within images, or give descriptions that are accurate but not very useful for the research context. Human review reliably catches these errors, and catching them before publication matters both professionally and institutionally.

Institutional repositories accumulate metadata inconsistency the way physical collections accumulate dust: gradually and invisibly. It becomes expensive to address only once it is large enough to affect discoverability measurably. Author name variants arise from inconsistent formatting across departments and submission periods. Subject terms reflect practices from multiple eras and catalogers. Abstracts are missing from older deposits that predate the current deposit workflow. Institutional affiliation fields contain every version of the institution's name used over its history. These inconsistencies reduce discoverability both within the local IR interface and in aggregated indexes like BASE, CORE, and OpenDOAR, which rely on metadata quality for reliable syndication.

AI batch processing addresses IR metadata inconsistency at a scale manual remediation cannot reach. It follows the same information-in logic that makes AI batch processing reliable in cataloging. Given a clear normalization task and a representative set of records, AI applies rules consistently across the full set, and the digital librarian reviews the result before committing it to production data. For example, a digital librarian managing an IR with five thousand records with inconsistent author name formatting can export a CSV of the affected name fields. They give AI the target format and three to five examples of the transformation, and receive a corrected dataset for review. Done manually, the task might take days of careful review and editing. With AI assistance and post-processing review, it can typically be done in hours. The important caveat is that the review step must be taken seriously, not treated as a formality.

Subject term enrichment for poorly described IR deposits is a second high-value application. Records deposited without subject terms, or with only local, nonstandard terms, are hard to discover through standard subject facets. They also syndicate poorly to aggregated indexes that rely on controlled vocabulary. Give AI a deposit's title, abstract, and any available keywords, and ask it to suggest subject terms from a specified vocabulary. The digital librarian evaluates the candidates against the actual controlled vocabulary and the item's content. For example: "Suggest five to seven subject terms for the following dissertation abstract, using Library of Congress Subject Headings. Note any terms you are uncertain about." This produces a working set of suggestions that substantially reduces the per-record effort of enriching a large backlog.

Ex Libris has implemented AI metadata enrichment directly in Alma for ProQuest EBook Central records. It automatically adds language, summary, and subject heading fields, aligned with Library of Congress standards, for records in the Alma Community Zone. This vendor-implemented example shows where the field is moving: AI metadata enrichment is becoming a standard feature of IR and library management platforms rather than a locally built workflow. For digital librarians evaluating platform upgrades or new IR systems, AI metadata capabilities are now a relevant evaluation criterion. It is a question worth asking any vendor whose platform manages institutional repository deposits.

Duplicate records are one of the quieter forms of metadata debt, and one of the most consequential for discovery. In a shared cataloging environment like WorldCat, duplicate records for the same item split holdings information across multiple records. A library's holding may attach to a record that other patrons and other libraries never see. In a local digital collection or institutional repository, duplicates arise from re-deposits, from harvesting the same item from multiple source systems, or from records created independently by different staff before a backlog was addressed. Duplicates inflate result counts and confuse patrons who meet the same item twice in a results list. They also complicate any downstream use of the collection's metadata, including syndication to aggregated indexes.

On February 11, 2025, OCLC ran a test of AI-assisted de-duplication against approximately 500,000 record pairs of print English-language books in WorldCat. The test evaluated how well an AI model could tell true duplicates of the same edition from distinct but closely related records, such as different printings, different editions, or multi-volume sets with similar titles. A test at that scale matters to digital collections librarians for the shape of the problem it illustrates, not only its size. Candidate duplicate pairs in a collection this large fall naturally into two categories. In the first, two records unmistakably describe the same edition of the same book. In the second, harder category, title, author, and publication information are similar, but the records describe a different edition, a different printing, or a multi-volume set that should remain distinct. Review matters most in the second category. A merge that collapses two genuinely different editions into one record destroys information a researcher might need, such as which edition contains a particular revision or which printing a specific provenance note refers to.

The same logic transfers directly to local digital collections and institutional repository work, at a smaller scale but with the same stakes. For example, a digital librarian may suspect that an institutional repository contains duplicate deposits of the same thesis, submitted once at deposit and again during a later metadata cleanup. AI can compare records across the repository and flag pairs that share enough metadata to warrant review. It produces in an afternoon a candidate list that manual comparison of several thousand records would take far longer to build. What AI-assisted de-duplication does not do, in the WorldCat-scale test or a local repository, is make the merge decision itself. Whether two records describe the same intellectual object, or two objects that merely resemble each other in their metadata, depends on examining the actual items, not only their descriptions. That judgment belongs to the cataloger or digital librarian reviewing the AI's candidate list, every time.

The most common mistake in AI adoption for digital collections work is choosing the tool before defining the problem. A digital librarian who reads about AI image description and immediately starts generating alt text for the most prominent collection may get inconsistent results. The workflow may need more review time than expected and be hard to hand off, because the process was developed informally. To build a workflow that scales and lasts, move from audit to pilot to documented process, and treat the documentation step as essential rather than optional.

The audit phase begins with a discerning assessment of where the most significant description, accessibility, and discoverability gaps exist in the current program. For most institutions, this means answering three questions. Which collections have the most serious description deficits? Which materials lack accessible equivalents that are required or most needed? Where is metadata inconsistency most affecting patron discoverability? The audit does not require a comprehensive collections inventory. It requires enough knowledge of the collection landscape to identify the highest-priority targets for AI assistance. A digital librarian who has managed a repository for several years usually already knows where the biggest problems are. The audit formalizes that knowledge into a prioritized list that can guide a phased implementation.

The pilot phase applies AI assistance to a small, bounded scope, such as one collection, one record type, or one batch of fifty items, before committing to a full-scale workflow. The pilot serves two purposes. It reveals the practical quality of AI output for the specific collection and prompt design, and it produces realistic estimates of review time for project planning. For example, a digital librarian piloting AI transcription for an oral history collection might run ten recordings through Whisper and time the review and correction of each. That data then supports an estimate of the full collection's transcription timeline. Empirical data like this is more reliable than vendor estimates or peer institution benchmarks, because it reflects the actual recordings (audio quality, speaker patterns, technical vocabulary) rather than generic performance claims.

Documentation is where sustainable workflow development most often breaks down. A workflow that works for one person and has never been written down is not an institutional workflow. It is a personal practice that disappears when the person changes roles or institutions. To build a workflow that persists and can be handed off, document four things: the prompt or prompt template used, the quality review criteria applied, the record-keeping conventions for flagging AI-assisted records, and the conditions that require escalation to additional professional review. This documentation is the difference between AI adoption that builds institutional capacity and AI adoption that builds individual efficiency. For digital collections programs managing long-term collections, institutional capacity is the appropriate goal. Sharing workflow documentation with peer institutions also helps. Networks such as the Digital Library Federation, the Society of American Archivists, and OCLC communities build the shared knowledge base that makes adoption easier for institutions earlier in the process.

This module has treated AI as something the library applies to its collections. The March 2026 ARL and CNI workshop report raises the question from the other direction. The library's collections are themselves training material, and in many cases they are already being consumed that way without anyone in the library having decided anything. Research libraries hold digitized texts, archives, oral histories, and curated corpora that commercial AI systems cannot replicate, and the report identifies these as a distinctive institutional asset. Assets attract attention, and this question has arrived faster than most libraries have built frameworks to answer it.

The first task is reviewing what the library has already agreed to. Many institutions signed digitization agreements with third-party partners years before generative AI existed, and those contracts were not written with model training in mind. ARL recommends reviewing every agreement with a company or organization that has digitized material from the collection, to establish what those agreements actually permit regarding AI training. For example, a mass digitization contract from 2011 may grant a partner broad reuse rights in language that nobody at the time understood to cover training a language model. The library will not know whether it does until someone reads the contract with that question in mind. This review is unglamorous, and it is the kind of work that surfaces obligations before they become disputes.

The second task is deciding what to exclude, which is a values judgment rather than a legal one. ARL's guidance is explicit: libraries should identify collections that should not be used for AI training because of cultural and social obligations or contractual agreements. The report notes that explicit decision frameworks for these exclusions remain among the clearest gaps in current practice. For example, Indigenous materials held under community protocols, oral histories recorded with narrator agreements that specified particular uses, and archives of vulnerable populations all raise questions that a general openness policy does not answer. Consent given for preservation and access was not consent for computational ingestion. Treating the two as equivalent because a license technically permits both would be a failure of the stewardship under which the material was accepted.

One distinction, so far collapsed in this section, decides most of these cases in practice. The University of Virginia Library's Archival AI Protocol, version 1.1, was published in 2026 under a CC BY 4.0 license, so any institution can adapt it freely. It separates general-purpose model training from retrieval-based access. General training "absorbs data into its model, often without context, and reversal is not realistic," while retrieval-based and controlled internal models "look at data but don't absorb it, and source items remain stored and controlled by the archive." The protocol's core rule follows from that divide: "Irreversible models do not get access unless item-level provenance and meaningful attribution can be demonstrated in practice, and the archival organization retains contractually enforceable control to stop further use." For example, a request to index a digitized oral history collection for a retrieval system that cites and links back to the original recordings differs from a request to train a foundation model on the same audio. A library that treats both as "AI use" has given away the only distinction that lets it say yes to one and no to the other. A librarian who receives a training request also now has somewhere to point. The Statement of Shared Practice on AI Training Requests and Unique Cultural Collections launched on April 3, 2026, with twelve founding institutions, and more than twenty have signed since. It is a voluntary twelve-month commitment covering archival and special collections rather than widely held published works. Signatories agree to classify each request, to evaluate digitization and AI use as separate agreements, to require provenance and transparency, and to apply a presumption against approval for broad commercial training requests. They also gain access to a shared, anonymized ledger of the requests peers are receiving, which turns an isolated negotiation into a sector-wide one.

The third task is the affirmative one, and libraries are further along here than the previous two suggest. The flagship example is Institutional Books 1.0, released on Hugging Face in June 2025 by the Institutional Data Initiative, based at Harvard Law School's Library Innovation Lab. It comprises 983,004 volumes, about 386 million pages, roughly 242 billion tokens, and 254 languages, published from the fifteenth century onward. It comes with use policies and access-by-request protocols that balance openness against responsible reuse. Cornell University Library has published a primer on open bibliographic metadata that works through licensing choices and which fields or records are appropriate to share. ARL recommends considering CC0 licensing for library-created metadata where appropriate. These projects show that libraries can participate in this space on their own terms, not only as the source of somebody else's corpus. One fact belongs alongside them, because a curriculum that teaches critical evaluation of vendor relationships should not omit it. The Institutional Data Initiative is funded by OpenAI and Microsoft, a relationship reported when the project was announced in December 2024. Stating this makes the example more useful, not less, because it raises the question the module should be asking anyway. What does it mean for a library to set its own terms when the infrastructure behind the corpus is paid for by the companies that stand to benefit from it? The use policies and access protocols are genuinely the library's, and the funding is genuinely theirs. A practitioner evaluating a similar partnership at their own institution needs both halves of that sentence in view.

There is also an operational reality that arrives whether or not the library has made any of these decisions. AI training crawlers have been hitting library and cultural heritage sites at volumes heavy enough to cause slowdowns and outages, a pattern documented across the sector through 2025. In the moment it happens, this is not a policy question; it is a systems administrator watching a repository fall over. ARL's response is collective. It calls for shared technical standards, advocacy for responsible crawling practices by AI developers, and engagement with standards bodies on norms that protect open infrastructure without forcing libraries to abandon openness as their default. For a practitioner, the immediate implication is narrower. Know whether your repository and digital collections platform are being crawled, know who in your institution would notice, and know the library's position before an outage decides the question for it.

The decisions in this section are also not the systems librarian's alone. Which collections may be used, under what conditions, and with what attribution are questions about the library's relationship to the communities whose material it holds. The people who negotiated those relationships need to be in the room. The technical capacity to make a collection machine-readable says nothing about whether it should be made so, and that distinction is the one this module asks practitioners to hold.