NARA’s AI Pilots
Measured Innovation or Hesitant Modernization?
As AI becomes embedded in everything from email tools to enterprise governance, national archives worldwide are grappling with the same question: how can we responsibly use AI to improve access, efficiency, and transparency, without undermining trust or accountability?
The U.S. National Archives and Records Administration (NARA) has responded with a suite of pilot projects testing AI’s potential across redaction, metadata generation, semantic search, and internal knowledge assistance. It's a notable shift for an agency that has often taken a conservative approach to digital transformation. But how do these efforts compare to more advanced or differently scoped initiatives abroad?
Let’s take a closer look.
As the federal government sharpens its focus on responsible artificial intelligence, the U.S. National Archives and Records Administration (NARA) is taking deliberate steps to define its role in this new era. A number of AI pilot projects, ranging from metadata generation to semantic search, signal a growing curiosity about the value AI might bring to archival work. But how far do these efforts go? And how do they compare to more integrated or assertive AI strategies emerging from peer institutions around the world?
Recent internal documents offer a clearer picture of both NARA’s aspirations and its limitations. Together, NARA’s AI Compliance Plan for OMB M-24-10 (September 2024) and its High-Level AI/ML Goals articulate a vision: ethical AI grounded in archival values, supported by upgraded infrastructure, and aligned with federal mandates. But that vision remains largely formative, and the distance between principle and practice is significant.
A Snapshot of the Pilots
NARA’s publicly disclosed pilots include:
PII detection and redaction using both AWS and Google Cloud tools, under evaluation.
Metadata generation for federal and presidential records, using AI to ease descriptive backlogs.
Semantic search pilots exploring meaning-based retrieval to improve user access.
Chatbot development for internal use at the National Personnel Records Center.
Generative AI in productivity tools, with around 50 staff testing Google Gemini in Gmail and Docs.
Each of these reflects real operational pain points. Redaction, metadata creation, and information retrieval are resource-intensive and central to NARA’s access mission. What’s admirable is the transparency and restraint with which these technologies are being introduced. These are controlled pilots, not rushed deployments.
Strategy or Scoping Document?
But a review of NARA’s compliance and goal-setting documents reveals something else: these pilots do not yet add up to a coherent AI implementation strategy. The goals are expansive—personalized user experiences, anomaly detection for authenticity, cryptographic hash publishing, and NLP-powered virtual assistants—but they remain largely conceptual. NARA acknowledges that it is still building its internal guidance on generative AI, working to define responsible use cases, and creating foundational governance infrastructure such as an AI Governance Board and AI Ethics Review Team.
Compare this to The National Archives (UK), which has integrated sensitivity review tools powered by machine learning into its digital accession workflows. Or Library and Archives Canada, which is actively incorporating multilingual AI into its metadata enrichment pipeline. These institutions are already embedding AI into their systems of record and public services. NARA, in contrast, is still in the phase of exploratory scaffolding.
Governance on Paper vs. Governance in Action
NARA’s documents outline a thoughtful, multi-pronged governance structure: cross-functional working groups, ongoing risk assessments, and collaboration with legal, IT, records management, and privacy teams. The agency also plans to periodically review all AI use cases for rights- or safety-related impacts, although it currently classifies its systems as low risk.
This governance model echoes the NIST AI Risk Management Framework and reflects current federal guidance. But what’s missing is clear criteria for how pilots will be scaled or terminated. There is no published roadmap detailing thresholds for performance, user trust, or policy fit that would justify broader adoption. Nor is there a structured plan for public engagement beyond general statements about transparency and feedback loops.
Open Data and Model Sharing: Aspirational
NARA also nods toward open science principles, stating that it will share custom code, models, and datasets where appropriate. Yet it notes that no custom models are being developed—only interfaces to third-party LLMs. This is important: if NARA is not building its models, it is ceding control over their provenance, bias profiles, and change logs. Responsible archival use of AI depends not only on ethical deployment but also on interpretability and verifiability of the tools themselves.
A Federal Leader—On Pause?
NARA’s leadership on records management policy and information governance has long been recognized across the federal landscape. But when it comes to AI, its approach so far appears more cautious than catalytic.
Part of this is structural. NARA is not a CFO Act agency and faces resource constraints. Yet it is also a symbolic agency, charged with preserving public memory and fostering civic trust. The bar for responsible innovation is high, but so is the potential.
Conclusion: More Than a Compliance Exercise
What NARA has offered so far—through pilots, governance planning, and interagency coordination—is a sincere and increasingly mature response to federal AI mandates. But what it has not yet offered is a clear articulation of how AI will be woven into its core business operations, public services, and strategic posture.
To get there, NARA will need to move beyond episodic pilots and toward a long-range strategy that embraces AI not only as a tool for efficiency but as a framework for expanding transparency, enriching public access, and confronting the complexity of archival evidence in the digital age.