Algorithms at the Reference Desk: What Artificial Intelligence Can and Cannot Do for LCSH's Unresolved Vocabulary Challenges

Artificial intelligence has entered nearly every domain of information management, and library cataloging is no exception. Yet despite promising demonstrations of machine learning applied to subject heading assignment, fundamental limitations in both the technology and the underlying controlled vocabulary suggest that automation alone will not resolve the structural tensions that have long defined LCSH's relationship with emerging knowledge. The question is not whether AI can assist catalogers, but what kind of assistance it can realistically provide.

The Promise That Arrived with the Hype

Over the past several years, a cohort of library technology vendors and research institutions has invested seriously in machine learning approaches to automated subject cataloging. Projects at institutions including the Library of Congress itself, the National Library of Medicine, and several major research universities have explored whether neural network models trained on existing catalog data can reliably predict appropriate subject headings for new materials.

The results have been genuinely encouraging within defined parameters. Systems trained on large corpora of bibliographic records with high-quality human-assigned headings have demonstrated meaningful accuracy rates for well-established subject areas with stable, consistently applied vocabulary. The Library of Congress's own Multilingual Subject Heading Project and related computational initiatives have shown that automation can meaningfully accelerate throughput for certain categories of materials—particularly monographs in heavily cataloged disciplines where training data is abundant.

For institutions facing growing cataloging backlogs—a near-universal condition among US academic and public libraries following years of budget pressure—even partial automation represents a genuine operational benefit. If a machine learning system can correctly assign the primary topical heading for seventy or eighty percent of incoming materials, catalogers can redirect their attention toward the more complex cases that genuinely require expert judgment.

Where the Model Meets Its Limits

The performance picture changes substantially when AI systems encounter the categories of material that pose the most significant challenges for LCSH itself: works at disciplinary intersections, materials on emerging research areas not yet represented in the authority file, and resources that engage with communities or perspectives historically underrepresented in the training data.

Machine learning models learn from patterns in existing data. When the existing data reflects LCSH's well-documented biases—its Anglocentric assumptions, its tendency to center Western academic frameworks, its historical inattention to certain communities and knowledge traditions—the model does not correct those patterns. It replicates and potentially amplifies them. A system trained on decades of catalog records that assigned inadequate or inappropriate headings to Indigenous knowledge systems, for example, will not independently recognize that inadequacy. It will assign similarly inadequate headings to new materials on the same topics, with computational confidence.

The problem is not merely one of training data quality, though that is a significant factor. It reflects a more fundamental constraint: machine learning systems optimize for pattern recognition within a defined vocabulary, but they cannot recognize when the vocabulary itself is insufficient. LCSH's persistent lag behind emerging interdisciplinary fields—the decades it took for headings related to gender studies, environmental justice, and digital humanities to achieve stable authorized forms—represents exactly the kind of gap that algorithmic approaches cannot bridge. A model cannot propose a heading that does not yet exist in the authority file, and it cannot flag an absence as meaningfully as a knowledgeable human cataloger can.

The Structural Problem Beneath the Technical One

Some advocates for AI-assisted cataloging acknowledge these limitations but argue that they are primarily engineering problems—matters of better training data, more sophisticated model architectures, and improved feedback loops between automated systems and human reviewers. That framing, while not entirely wrong, tends to obscure the degree to which LCSH's challenges are institutional and political rather than technical.

The controlled vocabulary does not lag behind emerging knowledge because no one has thought to update it. It lags because the process of adding, revising, or retiring headings involves deliberation, consensus-building, and institutional review that necessarily operates on a different timescale than research itself. Those processes exist for legitimate reasons: uncontrolled vocabulary growth undermines the cross-institutional consistency that makes LCSH valuable as a shared standard. An AI system that could generate new headings autonomously would not solve that tension; it would dissolve the controlled vocabulary into something closer to a tag cloud.

Similarly, the biases embedded in LCSH's historical vocabulary are not primarily computational artifacts. They reflect choices made by human catalogers operating within specific institutional contexts and cultural assumptions. Correcting them requires not better algorithms but sustained human engagement with communities whose knowledge has been inadequately represented, followed by deliberate revision of the authority record. No amount of machine learning can substitute for that work.

What Thoughtful Integration Might Actually Look Like

The most credible visions for AI in cataloging treat automation as augmentation rather than replacement—systems designed to handle high-confidence routine assignments while surfacing ambiguous cases for human review, rather than systems designed to eliminate the cataloger's role entirely.

Several academic libraries have piloted hybrid workflows along these lines. In these models, an automated system processes incoming materials and assigns preliminary subject headings based on trained confidence thresholds. Records that meet a defined confidence level proceed with minimal human intervention; records below that threshold are routed to cataloging staff for review. The result is not a fully automated catalog but a more efficiently allocated human workforce.

This approach has genuine merit, particularly for institutions managing large backlogs with limited staffing. It also has an honest relationship with what the technology can currently deliver. The risk lies in institutional pressure to raise confidence thresholds over time—to accept lower-quality automated assignments as budgets tighten and staffing declines—gradually shifting from augmentation toward de facto replacement without a deliberate policy decision to do so.

The Irreducibly Human Dimension

At its core, the work of subject cataloging is an act of intellectual interpretation. It requires understanding not only what a work is about in the most literal sense, but how that work relates to existing knowledge structures, what communities of researchers are likely to seek it, and whether the available vocabulary adequately captures its intellectual content. Those judgments draw on domain expertise, cultural knowledge, and professional training that current AI systems do not possess.

This does not mean that machine learning has no role in the future of LCSH-based cataloging. It means that the role is genuinely supporting rather than transformative. The vocabulary problem—the structural mismatch between a controlled list and the perpetual expansion of human knowledge—will not be resolved by better algorithms. It will be resolved, to whatever extent it can be, by sustained human investment in the intellectual and political work of keeping that list responsive to the communities it serves.