Ethical AI and Librarianship
A Resource Guide
Artificial Intelligence (AI) Applications for Modernizing Declassification
|
Field
|
Description | |
|---|---|---|
| Title | Artificial Intelligence (AI) Applications for Modernizing Declassification | |
| Type | Videos | |
| Creator | National Archives | |
| Link |
|
|
| Creation Date | 06/24/2024 | |
| Last Updated Date | -- | |
| Summary | This is a video recording of an in-person public meeting hosted by Public Interest Declassification Board (PIDB) at U.S. National Archives, introducing two federal agency initiatives that apply artificial intelligence (AI) to support declassification. This resource may be particularly relevant to librarians interested in AI project implementation strategies and human-in-the-loop approach. It also offers insights for information professionals seeking to understand the technical and operational requirements of implementing explainable AI models. Presentation one, by Timothy Kootz of the State Department, introduces a machine learning (ML) project at the State Department to assist with declassification review. The State faces an unmanageable volume of diplomatic cables that require manual review to determine whether they should be declassified or exempt from declassification. To address this, the agency trained ML models on prior human-reviewed decisions and adopted a human-in-the-loop approach. Human reviewers continue to perform quality control, handle ambiguous cases, and monitor for data drift. This approach significantly reduced the volume requiring manual review and shortened the overall declassification cycle. Building on the success of the first pilot, State launched a second pilot for matching FOIA requests with existing reviews to improve processing efficiency. Presentation two, by John D. Smith of the Department of Defense, and Dr. Michael Brundage at Applied Research Laboratory for Intelligence and Security (ARLIS), presents a human-centric ML project to enhance declassification efficiency amid a surge of digital classified records. To avoid the pitfalls of opaque "black box" models, DoD implemented a "white box" approach, which includes: 1) Mapping each exemption rule to its own model for easier auditing and retraining. 2) Using contextual models (e.g., BERT) that analyze not just sensitive terms but also the surrounding words and sentence structure to determine if the context matches classified subject matter. 3) Implementing a human-in-the-loop process, where the AI system flags potentially sensitive content and human reviewers make final decisions to declassify, or exempt. These human decisions are then used to train and refine future models. The video is freely and openly viewable (with auto-generated transcripts) through web browsers. Links to the event summary and presentation slides are provided in the Link field. | |
| Topic | Ethical AI. Libraries. AI and librarianship. | |
| Source and Link | National Archives | PIDB “Transforming Classification” Blog. https://transforming-classification.blogs.archives.gov/ | |
| Access | Open. | |
| Accessibility | Auto-generated transcripts available. | |
| Audience | Librarian – General. Information Professionals. | |
| Platform or Format | Web – YouTube | |
| Length | 2:05:30 | |
| Geography | USA | |
| Language | ENG | |
| Description Date | 06/25/2025 |