Introduction
The United States Intelligence Community has released interim guidance on the acquisition and use of Foundation AI Models [1], clarifying protocols and emphasizing adherence to existing frameworks. This guidance [1] [3], developed by inter-agency lawyers and aligned with Biden administration directives, outlines the conditions under which foundation models and their training data can be acquired and utilized for intelligence purposes.
Description
The United States’ Intelligence Community has issued the “Common Intelligence Community Interim Guidance Regarding the Acquisition and Use of Foundation AI Models,” clarifying protocols for the acquisition and utilization of artificial intelligence tools. Developed by inter-agency lawyers and aligned with directives from the Biden administration, this guidance emphasizes that acquiring a foundation model (FM) does not automatically include the associated training data; this is contingent on specific facts and circumstances [3]. It addresses the application of Attorney General (AG) guidelines to the Intelligence Community’s (IC) use of FMs for intelligence purposes, detailing aspects such as model acquisition, modification [1], augmentation [1], prompts [1], and outputs [1], while underscoring adherence to existing frameworks, including the Framework to Advance AI Governance and Risk Management in National Security [2] [3], which provides essential protections and principles for the IC [2].
The acquisition of FMs by IC elements is defined as hosting or accessing models in a manner not available to the public [1]. The guidance specifies that acquiring a FM does not equate to collecting training data unless the IC element can access the data in its original form and is authorized to do so [1].
IC elements are permitted to modify or augment an acquired FM through fine-tuning [1], which involves additional training on data collected by IC elements [1], or by connecting it to other data and tools without altering model weights [1]. However, covered information cannot be used for modification or augmentation unless it complies with AG-approved guidelines [1]. Furthermore, data obtained under the Foreign Intelligence Surveillance Act (FISA) cannot be used for these purposes without coordination with the Department of Justice and the Office of the Director of National Intelligence [1].
If an IC element does not currently hold data [1], any copying [1], saving [1], supplementing [1], or use of information is considered “collection.” Data already collected is not deemed “collection” unless a specific reason is provided [1]. The guidance also suggests mechanisms to mark or identify US citizen information returned by prompts [1], ensuring traceability to the appropriate model and activity [1].
The Interim Guidance does not address the sharing of resources and models between agencies [1], either within or outside the intelligence community [1]. Clarifications on this matter are anticipated in the near future [1], potentially through revised guidance or updates stemming from a national security memorandum on AI released by the White House in October 2024, which emphasizes the need for improved internal coordination and resource sharing within the US Government [1]. Concurrently [1] [3], the US Department of Commerce’s Artificial Intelligence Safety Institute has established a taskforce to research and test AI models relevant to national security and public safety [3], indicating a broader commitment to enhancing AI governance in these critical areas. As changes in administration may influence AI governance [2], companies supplying AI models to the Intelligence Community may face a more complex procurement landscape as developments progress.
Conclusion
The interim guidance on Foundation AI Models marks a significant step in formalizing the protocols for AI use within the US Intelligence Community. By emphasizing adherence to existing frameworks and clarifying the conditions for model acquisition and data use, the guidance aims to ensure responsible and secure AI deployment. Future updates and clarifications, particularly regarding inter-agency resource sharing, will further shape the landscape of AI governance and procurement, potentially impacting companies involved in supplying AI technologies to the government.
References
[1] https://ai-regulation.com/interim-guidance-on-foundation-ai-models/
[2] https://www.huntonak.com/privacy-and-information-security-law/agencies-focus-on-national-security-and-ai-directives-pursuant-to-executive-orders
[3] https://natlawreview.com/article/agencies-focus-national-security-and-ai-directives-pursuant-executive-orders
