Rep. Anna Paulina Luna (R-FL) is pushing back on claims that her staff used artificial intelligence to help draft language for a major defense bill amendment—insisting instead that AI was limited to a narrow, non-substantive task.
The dispute began after posts on X circulated screenshots of an amendment summary tied to the 2027 National Defense Authorization Act (NDAA). In the screenshots, the amendment description appears to include text that looks like it came from an AI assistant, including a line attributed to “Claude responded.” The implication—at least in the way the screenshots were shared—is that AI may have been used not just to polish wording, but to generate or shape policy content that ultimately became part of the legislative record.
Luna’s response, posted after the screenshots gained traction, draws a bright line between using AI as a writing aid and using it to draft legislation. She said her staff used AI only for “spellcheck” in an amendment summary, while denying that AI was used for the bill text itself. In her statement, she emphasized that “NO Legislation is ever drafted with AI,” framing the issue as one of misunderstanding: that the presence of AI-generated phrasing in a summary does not mean AI wrote the underlying legal language.
That distinction matters, because in Congress the difference between drafting statutory text and drafting explanatory or administrative materials can be significant. Amendment summaries are often treated as supporting documentation—useful for communicating intent, scope, and context—rather than as the operative language that governs law. Still, the public controversy underscores a growing concern: even when AI is used only for “helping” with documents, the outputs can blur into the kinds of sentences that resemble policy language, especially when summaries are written in a style that tracks the structure of legislation.
What exactly was shown in the screenshots?
According to the reporting described in the Verge article, the amendment summary screenshot included language that appeared to be “Identical to H.R. 100 (118th Congress)” followed by a reference to a time stamp and then an apparent AI interaction. The screenshot text included something along the lines of “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…” The phrasing is striking because it reads like a directive—precisely the kind of sentence that resembles legislative requirements.
When such language appears in a document labeled as an amendment summary, it can be easy for observers to assume that the AI output is more than cosmetic. After all, the line “Requires the Secretary of Defense…” is not merely a grammar correction; it is the kind of substantive formulation that could be interpreted as policy drafting.
Luna’s counterargument is that the AI involvement was limited to spellchecking in the summary, not drafting the amendment’s actual content. That claim, if accurate, would mean the AI output seen in the screenshot is either (1) an artifact of how the summary was prepared internally, (2) a misread of what the screenshot represents, or (3) a case where AI was used in a way that produced text resembling legislative language even though the underlying policy decisions were made by human staff.
But the controversy also raises a more general question: what does “spellcheck” mean in practice when AI tools are involved?
In everyday usage, “spellcheck” implies correcting typos and obvious errors. Yet modern AI writing tools often do more than fix spelling—they can rewrite sentences, adjust tone, and restructure phrasing. Even if a user intends to use an AI tool only for minor edits, the tool may respond with a fuller rewrite. If staff copy-pasted AI output into a document, even inadvertently, the result can look like AI authored the substance.
This is where the Luna case becomes a window into a broader shift in government workflows. As AI tools become embedded in common productivity software, the boundary between “editing” and “drafting” can become less clear—not because officials want to hide AI use, but because the tools are designed to be helpful in ways that go beyond strict spellchecking.
Why the NDAA context makes the issue combustible
The NDAA is one of the most consequential pieces of legislation Congress passes each year. It sets broad policy direction for the Department of Defense, authorizes spending, and often includes provisions that touch everything from procurement and readiness to personnel policy and national security strategy. Because it is so large and so politically salient, the NDAA also attracts intense scrutiny over how language is written and how amendments are described.
Amendments to the NDAA can be highly technical, and they often involve cross-references to existing bills and prior versions of legislation. That’s why the screenshot’s reference to “Identical to H.R. 100 (118th Congress)” is important. It suggests the amendment summary is tracking a prior legislative text or a previously introduced version. When AI appears in that context, it can be perceived as an attempt to speed up or streamline drafting—something critics may interpret as undermining transparency or accountability.
Even if AI was used only for editing, the public expects that defense policy language will be handled with careful human oversight. That expectation is not just political; it is procedural. Legislative text is supposed to be traceable to accountable actors—members of Congress, committee staff, and official legislative processes. If AI tools are used in ways that produce policy-like sentences, the public may worry that the chain of accountability is being weakened.
Luna’s statement attempts to restore that chain by asserting that AI did not draft legislation. But the screenshots complicate the narrative by showing AI-like phrasing in a document associated with an amendment.
The “summary” problem: when explanatory text starts to look like law
One of the most interesting aspects of this story is that it centers on an amendment summary rather than the final bill text. That detail matters because it changes what is being alleged. If AI wrote the bill text, the allegation would be about the operative legal language. If AI only touched a summary, the allegation becomes about communication and internal drafting practices.
However, summaries are not always purely descriptive. In many cases, amendment summaries are written in a way that closely mirrors the structure of the underlying legal requirement. They may use “requires” language, specify who must do what, and describe the scope of an obligation. That style is meant to help readers understand the amendment’s effect quickly. But it also means that if AI is used to generate or refine those sentences, the output can look indistinguishable from legislative drafting.
So even if Luna is correct that AI was used only for spellcheck, the public may still see the results as substantive because the summary is written in substantive language. This is a design issue as much as a technology issue: the format of legislative summaries invites policy-like phrasing, which makes AI artifacts more visible.
There is also a practical point: staff may use AI tools to speed up rewriting tasks, including converting rough notes into polished summaries. If the tool is asked to “fix” a paragraph, it may produce a rewritten version that is more than a spelling correction. The staff might then treat the output as a “spellcheck” result because it was used for editing, even though it involved more extensive language generation.
In other words, the dispute may not be about whether AI was used at all—it likely was. The dispute is about what role AI played and whether that role crossed the line from editing to drafting.
What Luna’s denial signals about the political stakes
Luna’s response is unusually categorical: “NO Legislation is ever drafted with AI.” That kind of absolute language suggests she believes the core accusation is that AI was used to create legislative text. By denying that outright, she is trying to prevent the story from evolving into a broader narrative about AI replacing human lawmaking.
But absolute denials can also invite skepticism when screenshots appear to show AI-generated policy-like language. Critics may argue that if AI output appears in a document connected to an amendment, then the denial is incomplete or misleading unless it explains how the AI output got there and why it does not reflect drafting of the operative text.
That tension is likely to keep the story alive. In the absence of a full document trail—such as the original draft, the exact AI prompt, the tool settings, and the final version of the summary—public debate will continue to hinge on interpretation.
Still, Luna’s statement provides a framework: AI can be used for minor editorial tasks, but not for drafting legislation. Whether that framework holds up depends on details that are not fully visible from screenshots alone.
The bigger question: should AI use in legislative communications be disclosed?
Even if Luna’s specific claim is accurate, the incident highlights a policy gap. There is no single, universally enforced standard for how members of Congress should disclose AI use in drafting communications, summaries, or internal documents. Some offices may voluntarily disclose AI assistance; others may not. And because AI tools can be used in subtle ways—like rewriting a sentence, suggesting alternatives, or correcting grammar—the line between “AI used” and “AI not used” can be hard to define.
This is particularly relevant for legislative materials. The public expects that official documents are produced through transparent processes. If AI tools are used, even for editing, disclosure could help maintain trust and reduce confusion.
At the same time, requiring disclosure for every minor AI-assisted edit could be burdensome and could lead to over-disclosure that dilutes meaningful transparency. The challenge is to find a disclosure standard that distinguishes between trivial assistance and substantive generation.
The Luna case may push that conversation forward by making the issue visible to a wider audience. Screenshots that include “Claude responded” are attention-grabbing, and they make the invisible visible. Once the public sees AI traces in documents, it becomes harder for institutions to treat AI use as a purely internal matter.
A unique angle: the risk of “AI residue” in documents
Another insight from this story is the concept of “AI residue”—the leftover traces of AI interactions that remain in documents. Many AI tools produce responses that include conversational markers, attribution lines, or formatting that