By: Tamir Borensztajn, EBSCO Senior Vice President of Marketing; and Andrew White, Interim Vice President, Enterprise Computing; Director, Library Information Services, Rensselaer Polytechnic Institute

Artificial intelligence has introduced a paradox of progress: the more capable our models become, the more they risk producing confident errors. This tension between speed and trust defines the current moment in information systems. As large language models (LLMs) increasingly mediate how people access and interpret knowledge, the question shifts from “Can AI generate?” to “Can AI be trusted?”

From Bigger Models to Better Foundations 

The path toward trustworthy AI lies not in ever-larger models but in stronger foundations, specifically through the use of unique identifiers (UIDs), verified statements, and linked data standards that anchor AI outputs.

This idea isn’t new. It echoes the long-standing information science problem of ambiguity. Whether in cataloging, metadata, or semantic web architecture, ambiguity has always been a challenge to information retrieval. 

The same principle applies to AI. When an LLM encounters a single keyword search like “Peanuts,” it lacks the contextual scaffolding to distinguish between a legume, a comic strip, or a corporate entity. If left to its own logic, the AI could pick the wrong context and thus create misinformation or hallucinations. This confusion mirrors early cataloging challenges before the advent of controlled vocabularies and persistent identifiers. The difference now is scale: AI replicates and amplifies ambiguity at machine speed. Without grounding, even the most advanced models can misclassify, misinterpret, or hallucinate.

Building “Ground Truth” with Linked Data

So how do we fix it? We start by building reliable and trustworthy sources for the AI to consult.

By connecting entities to resolvable identifiers (e.g., Wikidata IDs, schema.org URIs, etc.), systems can move from statistical plausibility to semantic precision because the UID helps retrieve the correct information for the precise entity the user is looking for. 

Having a UID also allows for retrieval of metadata and other data the entity relates to, facilitated as a knowledge graph, that can be used in the retrieval-augmented generation (RAG) process to further ground AI responses on more trustworthy data. These knowledge graph statements separate verified statements from vague claims, enabling models to produce responses that are both relevant and verifiable.

Prior to the wider public availability of AI platforms, the Libraries at Rensselaer Polytechnic Institute began exploring ways to leverage open-linked data (OLD) architectures. The initial foray into OLD was through research projects with faculty experts in data visualization and cognitive science, developing knowledge graphs from hand-coded XML as a mechanism for creating interactive agent-driven narratives with data visualizations. 

These projects could be considered precursors of agentic AI in their intended functionality, which demanded that the metadata supporting the narrative engine contain referential contextual indicators and terminology to establish information coherence and accuracy. The figure of the XML snippet from this project illustrates how to leverage the flexibility of metadata to establish relationships (namesake, designer, geolocation, construction date) with specificity that help to assure accuracy in the information.

The evolution of technology and of ontologies has further enabled the RPI Libraries to take advantage of authoritative metadata in the description of archival digital assets. RPI Libraries are one of numerous members of the Archipelago Commons, which supports a digital asset management platform built with the principles of open linked data in mind. The ability to utilize various authoritative ontologies and schema simultaneously can be seen by accessing online an archival record of the design for a tower foundation of the East River Bridge (now known as the Brooklyn Bridge), and then following the various Library of Congress and Wikidata Subjects mapped on the left side of the tower foundation webpage. 

Trust is the Infrastructure

The broader argument extends beyond AI to the entire ecosystem of digital scholarship and discovery. Curation, durable metadata, and interoperability remain essential to the infrastructure of trust. Discovery depends on connecting people with relevant, credible, and contextual information. The same principles that guide scholarship - clarity, verification, and consistency - must guide how AI represents and retrieves knowledge.

From Information to Understanding 

UIDs and linked data do not merely refine how AI interprets information; they reshape how humans and machines collaborate to construct meaning. They enable dialogue rather than error, explanation, rather than vague claims. 

By building on what does not change - verified facts, standard vocabularies, and trusted sources - we create a stable foundation for discovery.

In short: the path to trustworthy AI isn’t bigger models. It’s better, more trustworthy and verifiable data.