Federal AI Prompts Rarely Qualify as Records Under New NARA Rules

Federal agencies face minimal requirements for preserving AI interactions

The National Archives and Records Administration published guidance in August 2026 that defines when federal employees' use of generative AI tools creates a federal record—but the framework is surprisingly narrow, potentially leaving significant gaps in government transparency.

According to analysis first reported by Just Security, NARA's interpretation means AI prompts and outputs only become federal records if employees manually download or capture them in official agency systems. Conversations that remain solely within third-party platforms like ChatGPT, Claude, or Copilot are not considered "received" by the agency and therefore fall outside Federal Records Act requirements.

Why it matters

Federal employees are using AI for consequential decisions—from State Department visa revocations to VA contract cancellations to military operations. Without comprehensive records of these AI interactions, investigators cannot determine whether decisions were based on sound reasoning or AI hallucinations. Courts lose the ability to conduct "arbitrary and capricious" review under the Administrative Procedure Act when the rationale—the AI conversation itself—has been erased. Future historians will lack insight into how this technological shift transformed government operations.

A departure from previous technology guidance

NARA took a different approach with AI than it did for earlier technologies. When agencies adopted email, cloud computing, and social media, NARA guidance recognized that records might reside on third-party servers and needed retrieval. The 2015 electronic messaging guidance clearly stated that messages created during agency business are federal records "regardless of whether federal employees use a personal account."

The AI guidance reverses this logic. It focuses exclusively on information already "received and preserved" rather than information "made" or "appropriate for preservation"—categories explicitly included in the Federal Records Act's statutory definition.

What the guidance includes

NARA identifies several categories likely to constitute federal records: user prompts and AI outputs captured in agency systems, meeting summaries, chat histories (audit trails), procurement documents, copies of records used for AI training, and agency-created software code.

The guidance also addresses record disposal schedules, distinguishing between transitory records (needed under 180 days), intermediary records (used to create subsequent records), and permanent records.

Critical gaps remain

The guidance lacks procurement language that previous NARA directives included for cloud and social media contracts. Earlier guidance provided draft terms requiring vendors to ensure "secure storage, retrievability, and proper disposition of all federal records."

The definition of "significant modification" for off-the-shelf AI products remains unclear, leaving common use cases unresolved. The guidance also omits chain-of-thought reasoning, attachments, metadata like model versions and timestamps, and other contextual data that enterprise AI accounts capture.

These omissions echo a 1990s court case where judges ruled that email printouts were insufficient records because they lacked structural and contextual information including recipient and sender names.

What happens in practice

Under this framework, federal employees are unlikely to routinely copy entire ChatGPT conversations with full metadata into Word documents on shared drives. More probably, they will use AI to generate content for documents without any trace of AI involvement—exactly the scenario that limits transparency and oversight.

The analysis was published by Asad Ramzanali and Margaret Mullins at Just Security.