When Stanford researchers put specific, verifiable questions about federal cases to leading language models, the models fabricated an answer between 58 and 88 percent of the time. Tax citations are vulnerable to the same problem.

Ask a general-purpose model about the qualified business income deduction and you may get a fluent paragraph anchored to something like §199A(d)(4)(C). The citation looks ordinary. But §199A(d) stops at paragraph (3). There is no (d)(4).

That distinction matters, because a misremembered fact and a manufactured one fail in different ways, and only one of them can be fixed.

A citation is just a string the model finishes

A language model writes one token at a time, choosing a likely continuation from the context. This works well for prose because grammar and meaning constrain what comes next.

A citation is different. The surrounding sentence gives the model little help in choosing between (d)(2) and (d)(4)(C). It has seen many strings shaped like §[number]([letter])([number]), so it can reproduce the form even when the particular subsection is wrong.

Unless the assistant retrieves the citation from a source, a plausible section number may simply be part of the generated answer.

Why tax is a worst-case domain for this

Three properties of tax authority stack the deck.

Section numbers are unforgiving. A wrong subsection letter can look perfectly normal while pointing nowhere.

The real structure is deep and irregular. §199A defines the specified service trade or business test at (d)(2), puts the phase-in mechanics at (d)(3), and the threshold amount at (e)(2). Remembering the topic is not enough to reproduce that structure accurately.

CitationReal?What is actually at that address
§199A(d)(2)YesDefines the specified service trade or business
§199A(d)(3)YesPhase-in of the SSTB limitation
§199A(e)(2)YesThe threshold amount
§199A(d)(4)(C)No§199A(d) ends at paragraph (3). There is no (4).

Tone does not reveal accuracy. The model can present (d)(4)(C) as confidently as (d)(2). The reader has to check both.

The fix is not a better prompt

A prompt such as “only cite real sections” may reduce errors, but it does not give the model a copy of the Code. You still cannot know which citation is right without checking it.

The better approach is to retrieve the section from the Code and give that source to the model. This makes the citation traceable and lets the model focus on explanation and drafting. It still does not prove that the retrieved provision is complete or correctly applied, so professional review remains necessary.

Four habits that catch it today

Until your workflow is grounded in retrieval, these four checks are what stand between a fabricated cite and a client deliverable.

  1. Do not accept a bare citation. Open it on uscode.house.gov, the eCFR, or another authoritative source.

  2. Compare the quoted text with the source. When the quote and the citation disagree, rely on neither until you resolve the mismatch.

  3. Check deep subsection paths carefully. Every added level is another place for a letter or number to be wrong.

  4. Separate the drafting layer from the authority layer. Let the model write. Do not let it be the source of record for what the law says. The section numbers, the quoted text, and the effective dates should come from something that actually holds the law, with a URL you can open.

A general model can still be useful for explanations and drafts. Just give it a real source to work from, and open that source before the answer reaches a client.