A Source of Authenticity: How the Library of Congress Approaches AI

Published by the Partnership for Public Service AI Center for Government with support from Civic Design Collaborative 

Lead Contributor: Amanda Starling Gould
Contributors: Nadine Foik and Sean Baker

As the steward of a collection containing over 178 million items in 470 languages and virtually every format imaginable, the Library of Congress is the largest library in the world, serving Congress, researchers and the American public. When it makes a claim, the public expects it to be true. That expectation shapes the Library’s approach to artificial intelligence. Natalie Buda Smith at the Library of Congress shared what that work looks like in practice. 

This is part of a series of interviews we are conducting with leaders across government to better understand how agencies are measuring the public impact of AI implementations.

Setting Up AI Base Camp 

Centralization to Enable Strategic AI. “Having technology centralized under the office of the Chief Information Officer was really key to our AI strategy,” Buda Smith said. “We were able to set up the base camp, start bringing people in and then deploy processes, resources and strategy throughout the Library.” Without centralization, working groups across service units may have developed parallel, uncoordinated approaches. With it, the agency can speak in one voice and act on shared principles. 

Grounded in Shared Values. A Library-wide strategic plan, completed for fiscal 2024-28, gave the institution a shared innovation and emerging technology vocabulary that staff across different service units contributed to and understood. The Library AI working group built on this foundation to develop six guiding principles for AI use. Representatives from across the institution contributed. Library leadership endorsed them, and it was clear that, if they ever had to prioritize, trust comes first.  

The Evaluation Framework: How the Library Makes Responsible Decisions 

The Library’s approach to evaluating AI is built around three interlocking structures: institution-wide guiding principles, a use-case inventory and a governance process embedded within existing technology review bodies. 

Principles as decision tools. Buda Smith is candid that the principles are qualitative commitments, not metrics. “We don’t necessarily have direct measures,” she said, “but we do have signals: incidents, training completion rates and network monitoring results.” For trust and authenticity, these signals run through culture and guardrails rather than data dashboards. The principles matter most in moments of tradeoff when there is pressure to deploy quickly. 

The use case inventory as a shared workforce tool. The Library maintains a living document that logs every AI tool in use across the institution: who is responsible for it, what data it touches, its risk level and the controls in place. “We’re making sure we have all the information we need to understand what people are doing, who is responsible and what’s the risk level,” Buda Smith said. The inventory serves as a coordination, risk management and governance tool in one. When a unit wants to adopt a new AI service, staff can check whether another is already using it. When patterns across the inventory signal a risk, they become the basis for institution-wide response. One recent example: inventory discussions surfaced questions about vendors using AI off the agency’s network for their deliverables. The Library receives the output but does not control the process. What standards should apply? That question became a new vendor requirement. 

Governance through existing structures. To reduce bureaucratic complexity, the Library chose not to build an entirely new parallel set of administration processes for AI. Instead, it embedded AI considerations into existing agency processes that already govern all technology adoption, with a Library AI working group member serving on the agency’s Technology Architecture Board. “Let’s build on that strong foundation of IT security and governance, not build a whole new separate process for AI,” Buda Smith said. Decisions stay close to the people with technical and operational knowledge rather than being handed off to a separate committee that may lack context. 

Experimentation through pilot projects. The Library runs experiments and pilots rather than adopting by default. This allows adoption to be targeted and differentiated: staff adopt tools that fit the mission and dismiss those that don’t. Getting the pilot evaluations right, Buda Smith argues, matters more than getting too many tools too quickly. 

Designing for change. One consideration Buda Smith raises that rarely appears in formal evaluation frameworks is the risk of lock-in. Federal AI preferences and market dynamics have both shifted quickly in recent years. The Library designs its AI-assisted workflows with substitutability in mind: when the underlying model changes, the workflow should not require a complete rebuild.

 

What Other Institutions Can Learn 

The Library’s context is not universally shared, but the underlying practices are transferable. 

Centralize before you scale. Coordination costs are invisible until they are enormous. A central function, or at a minimum a coordinating body, is what makes a coherent strategy possible. 

Ground your governance in mission values you already hold. The Library’s AI principles extend the institution’s existing beliefs about authenticity and public service. Principles developed from scratch and disconnected from existing values tend to produce documents that staff cannot apply when a real decision arrives. 

Use the inventory as a governance engine. The value is in the discussions the inventory enables and the patterns it surfaces, not (solely) in the list itself. 

Match the tool to the purpose. The pressure to adopt AI through the path of least resistance is real and constant. A convenient tool that is not well-suited to the institution’s actual needs creates more work than it saves and establishes adoption patterns that are hard to reverse.  

Design for substitutability from the start. In a market environment where AI providers and preferred tools can change quickly, building workflows that are tied to a single model or vendor creates fragility. Ask not just whether a tool works now, but whether and how you can replace it if necessary. 

Put humans at the center. The Library does not treat human review as a checkpoint in an automated pipeline. The human is the decision-maker. AI is a tool to support that decision. 

Questions That Remain

The Library’s confidence about what it values and reliance on well-developed IT processes helps with many of the questions it faces. Nonetheless, Buda Smith identified several questions that remain genuinely open: 

Explore other case stories