Traditional library discovery layers are built on exact-match retrieval. If a student searches for “digital library accessibility barriers for the blind,” the system scans metadata for those exact strings. If a seminal paper uses the terms “web accessibility” and “visually impaired” instead, that paper may fall to page five of the results or disappear entirely. For BVI users, this “keyword hunting” is more than an inconvenience. When navigating with screen readers, every irrelevant result or missed connection increases the cognitive load of the research process. The “discovery gap” is the distance between what a user needs and what a system provides based on their ability to master complex syntax. To explore this, I conducted a comparative analysis using the Stony Brook University Library SEARCH AI platform. I ran an inquiry through two lenses: a natural-language prompt and a manually constructed Boolean string. Since my research interest is in digital accessibility, I used “What are the barriers of digital library accessibility for the Blind and Visually impaired?” as the search query. For the traditional Boolean logic, the query was structured as: (digital librar* OR “online library”) AND accessibility AND (blind* OR “visually impaired”) AND barriers.

The result? A massive haul of 117 documents from the Boolean logic search. On the surface, this suggests high recall, where the system found nearly everything containing those strings. In contrast, in the natural language (SEARCH AI), the system did not just match words; it interpreted them. The AI agent expanded the query into a sophisticated multi-layered string: (barrier OR obstacle* OR challenge* OR difficulty*) AND (“digital library” OR “online library” OR “electronic library”) AND (blind OR “visually impaired” OR “visual impairment” OR “low vision”) AND (accessibilit* OR usability OR “user access”). One might expect that by introducing so many synonyms (like “obstacles,” “usability,” and “low vision”), the natural language search would return thousands of results. Surprisingly, it did the opposite: it returned only 31 results, compared to the 117 retrieved by the Boolean string. For researchers, lower recall is often viewed with suspicion that would need double-checking. Did the AI miss something important? While this reduction might initially look like a win for efficiency, a closer look reveals a significant trade-off. In evaluating the output, it became clear that the Boolean search, despite its “noisier” 117 results, captured highly relevant materials that the SEARCH AI simply missed. This suggests that while the SEARCH AI’s coded string was linguistically broader, its internal ranking and filtering logic may have been too restrictive.

For the BVI community and the researchers who serve them, this finding is a cautionary tale. If we rely solely on natural language discovery, we risk a high-low failure: high convenience, but low comprehensive recall. From a translational and accessibility perspective, the reduction from 117 to 31 results is a massive win for user experience. However, for critical research, where missing a single study on a specific accessibility barrier can stall progress, the natural language mode cannot yet stand alone. It serves best as a discovery starter, but the Boolean search remains the “safety net.” To ensure no relevant materials are missed, users must still be encouraged to compare both outputs.

The move toward natural language searching in libraries is a move toward Inclusion by Design, but it must not become a move toward information erasure. My findings suggest that AI does not yet replace the need for traditional information literacy. As we continue to test these tools at institutions like Stony Brook, the goal should be a hybrid model. We need systems that offer the ease of a conversation without losing the exhaustive reach of a logic-driven search. If we want a truly accessible digital library, we must ensure that the “AI concierge” doesn’t accidentally close the door on the very research we need most.