The United States’ Intelligence Community1 (IC) has released Interim Guidance about the acquisition and use of foundation AI models (FMs).
This guidance has been provided by inter-agency lawyers and concerns the application of AG guidelines and policies to the IC’s acquisition and use of FMs. As well as clearly suggesting that IC elements will use FMs for intelligence purposes, the document provides additional guidance concerning model acquisition, modification, augmentation, prompts and outputs.
- Model acquisition
The document defines FM acquisition by IC elements as when it is hosted or accessed in a manner different from that available to the public for the model.
The document defines the criteria on which training data is considered to have been collected. It affirms that acquiring a FM does not in itself constitute collection of training data; this depends on whether the IC element is able to access the data in its initial form, and whether it is authorised and intends to do so.
- Model modification and augmentation
IC elements may modify or augment an acquired FM through fine-tuning – a form of modification that involves additional training on data collected by IC elements, which alters its model weights, or, in the case of augmentation, they may choose to connect it to other data and/or tools, without altering the model weights. The guidance states that covered information cannot be used by IC elements to modify or augment a FM unless the purpose and use complies with Attorney General (AG) approved guidelines. In addition, IC elements are not permitted to modify or augment a FM using data obtained or derived from the Foreign Intelligence Surveillance Act (FISA)2 without coordinating with the Department of Justice and Office of the Director of National Intelligence.
- Model prompts and outputs
The document defines the circumstances in which information, returned from a prompt, can be considered as “collection”:
- If the data is understood as not currently held (collected or acquired) by the IC element. However, it would be considered as “collection” if the information is copied, saved, supplemented or used.
- If the data is understood to have already been collected by the IC element, it should not be considered “collection” unless an articulable and/or case-specific reason is provided.
Finally, the guidance recommends using the opportunity to explore mechanisms to mark or identify any U.S. citizen information returned by a prompt, provided they are already known and retained by the IC element, so that the information can be traced back to the appropriate model and activity.
One question that the Interim Guidance does not address concerns the sharing of resources and models, once they have been modified or augmented, between agencies within or beyond the intelligence sphere.
However, it is reasonable to expect some clarification and updates on this subject in the coming months, either due to revised Interim Guidance, scheduled to be published within the next 6 months, or from work done since the issuance of the National Security Memorandum on AI in October 2024. In section 4.2.j, the Memorandum mentions the need to improve internal coordination to ensure interoperability and resource sharing, and calls for guidance to be issued to ensure that the United States Government can “coordinate and share AI resources effectively”.
To stay informed, visit our website at AI-Regulation.com and follow us on LinkedIn, Twitter and Facebook.
S.P
- Elements here refers to the departments within the intelligence community such as the CIA, NSA, FBI etc. ↩︎
- FISA-obtained or derived data are data collected from foreign actors under the Foreign Intelligence Surveillance Act with the assistance of U.S. telecommunication companies. ↩︎