Rep. Luna’s Staff Left Claude AI Output in a Defense Bill Summary

Rep. Anna Paulina Luna (R-FL) found herself defending her office's use of AI after screenshots circulated on X showing visible Anthropic Claude output inside an amendment summary for the 2027 National Defense Authorization Act. The text included a timestamp and the label "Claude responded:" before substantive legislative language. Luna says her staff used AI only for spellcheck in the summary document, and insists no bill text was drafted by AI. The incident is a clean example of what happens when AI tools are used in professional workflows without basic review steps.

What happened

An amendment summary tied to the 2027 National Defense Authorization Act (NDAA) began circulating on X after accounts posted screenshots of its contents. The document contained a line reading: “Identical to H.R. 100 (118th Congress).11:25 AM????Claude responded: Requires the Secretary of Defense to designate Department of Defense activities, support, and operations at the southwest land border as a named operation with…”

That text is not legislative boilerplate. It is the raw output of a conversation with Anthropic’s Claude, including a timestamp, pasted directly into what became a public-facing document. Someone on Luna’s staff copied the AI’s reply and did not remove the header before publishing.

What the congresswoman said

Luna issued a statement after the screenshots spread. She acknowledged that her staff used AI for “spellcheck” on the amendment summary. She denied that AI wrote any portion of the actual legislation, stating plainly: “NO Legislation is ever drafted with AI.”

Her initial response reportedly created some confusion before she clarified the position. The office drew a line between the summary document (where AI was used in some capacity) and the bill text itself (where she says it was not).

Why it matters

This is not really a story about whether AI should assist legislators. Plenty of offices use AI tools for research, drafting, and editing, and that is a reasonable thing to do. The problem here is process.

A raw Claude reply, timestamp and all, ended up in a published government document. That means at least one of these things happened:

• The output was copy-pasted without review.
• No one proofread the final document before it was shared.
• There was no checklist or approval step to catch obvious errors.

Any one of those is a workflow failure, not an AI failure. The model did exactly what it was asked. The staff did not verify what they published.

For anyone running a business, this matters because the same risk exists in your own processes. AI tools produce output fast. Fast output is easy to paste without reading. Client proposals, legal documents, support emails, ad copy, all of these can carry AI artifacts if you skip the review step.

Our take

The “we only used it for spellcheck” explanation is doing a lot of work here. A spellchecker does not produce a timestamped conversational reply with its own name in the header. Someone asked Claude a substantive question about the amendment and then pasted the answer, more or less directly, into the document.

That is not inherently scandalous. What is worth noting is how quickly the explanation shifted from the screenshots to the statement. The story also shows something we see often with clients who are new to AI tools: there is a gap between how people describe their AI use (“just for spellcheck”) and what they actually do with it. Both can be true at the same time. The gap matters because it shapes how carefully people review the output.

If you tell yourself you are only using AI for minor edits, you may not feel the need to read the result closely. That is exactly when something like this happens.

What to do about it

If your team uses AI in any document that will be shared externally, put one rule in place now: the person who pastes AI output is responsible for reading it before it leaves the building. No exceptions for summaries, headers, or “minor” sections. A simple final review step catches nearly all of these problems before they become screenshots on social media.

Source: The Verge · AI