What are the records and classification risks of using generative AI tools with sensitive or classified information?

Generative AI tools can summarize, draft, and analyze text quickly, but feeding them sensitive, Controlled Unclassified Information (CUI), or classified material raises serious recordkeeping and security concerns. The convenience of these tools does not change the underlying legal and policy obligations that govern how protected information must be handled, captured, and safeguarded.

Spillage and unauthorized disclosure

The most immediate risk is spillage — placing classified or controlled information into a system not authorized to hold it. Many AI tools transmit prompts to external servers, retain inputs, and may use them to train future models. Once classified or CUI data leaves an accredited environment, it can be exposed to unauthorized parties, and the disclosure may be difficult or impossible to fully remediate. Even unclassified-but-sensitive details can become a compilation risk when aggregated.

Classification and derivative classification

AI output can blur classification lines. A model may combine fragments in ways that reveal classified relationships, or it may strip classification markings, banners, and portion marks that controlled documents require. Relying on AI-generated text can also undermine derivative classification discipline, where authors must carry forward markings from authorized sources. Mismarked or unmarked output can lead to over-classification, under-classification, or improper handling.

Recordkeeping and accountability gaps

Prompts and AI responses created in the course of agency business may themselves be federal records subject to capture, retention, and disposition. If these interactions occur in tools outside official systems, they can escape proper recordkeeping, creating gaps that harm accountability, FOIA responsiveness, and litigation readiness. Lack of provenance also makes it hard to verify the authenticity and reliability of AI-assisted content.

Practical safeguards

• Never input classified or CUI material into AI tools that are not accredited for that information’s level.
• Follow your organization’s AI use policy and consult security and records officials before adopting any tool.
• Capture official AI interactions in approved recordkeeping systems and apply correct markings.
• Treat AI output as a draft requiring human review, verification, and proper classification.

Strong governance — combining classification rules, records requirements, and clear AI-use policy — keeps these tools useful without compromising protected information. See the declassification topic hub for related guidance.