Can AI Interpret Statutes? How Statutory Interpretation Works With AI

Short answer: AI can read statutes, but it cannot interpret them the way a court does. An LLM will find the operative text, summarize it, pull the amendment history, and surface the cases that construed a term in seconds. It does not weigh the canons, the legislative history, the structural arguments, or pick the single best reading. After Loper Bright killed Chevron deference, that judgment call sits squarely with the lawyer and the court, not the model. Use AI statutory interpretation as a research accelerant and own the reading yourself.

A federal judge asked ChatGPT what "landscaping" means

In May 2024, Eleventh Circuit Judge Kevin Newsom did something that made half the legal-tech world nod and the other half wince. In a concurrence to Snell v. United Specialty Insurance Co., 102 F.4th 1208 (11th Cir. 2024), he admitted that while puzzling over whether installing an in-ground trampoline counted as "landscaping" under an insurance policy, he had quietly typed the question into a large language model and asked it for the ordinary meaning of the word.

He was careful about it. He did not pretend the model decided the case. He framed it as one tool sitting next to dictionaries, usage corpora, and the interpretive canons, a possible new instrument for the old job of figuring out what words ordinarily mean.

But the concurrence kicked off exactly the argument it should have: can statute interpretation AI actually interpret a statute, or is it just doing a very confident impression of interpretation?

This is the question that matters more than any benchmark score, and most people get it backwards. They treat interpretation as a lookup problem. It is not.

Interpretation is an argument about the best meaning of a text, and an argument is not something you retrieve. This post is about where AI genuinely helps with that work, where it quietly overreaches, and why the Supreme Court's decision to bury Chevron deference made the whole question hotter, not cooler.

TL;DR

• AI does not "interpret" statutes the way a court does. Interpretation is an argument about the best meaning (ordinary meaning, structure, canons, precedent), not a fact you can fetch.
• A real federal judge (Newsom, CA11) floated LLMs as one interpretive tool. A real counter-camp (and a dissent citing ChatGPT in Ross v. United States) argues LLMs encode their makers' choices, not a neutral census of language.
• Loper Bright v. Raimondo, 603 U.S. 369 (2024) overruled Chevron and told courts to find the single best meaning themselves. That raises the stakes on getting the text exactly right.
• AI's honest job is surfacing the raw material faster: exact text, every cross-reference, amendment history, and how courts have construed a term. That is research, not adjudication.
• The overreach risk is the same one that produced sanctions in Mata v. Avianca: confusing a fluent answer for a legally operative one.

Part of our corporate and transactional lawyer playbooks.

Interpretation is an argument, not a fact

Start with what a court is actually doing when it "interprets" a statute. It is not looking up a definition. It is building a case for the best reading among plausible competitors, and the materials it uses are layered.

There is ordinary meaning, what a normal speaker would understand the words to mean. There is structure, how the provision fits with the rest of the statute and the surrounding code. There are the canons: lenity, which resolves genuine ambiguity in a criminal statute in the defendant's favor; noscitur a sociis, where a word takes color from its neighbors in a list; the major questions doctrine, which demands clear congressional authorization before an agency claims sweeping economic power. And there is precedent, the binding gloss courts have already put on the same words.

None of that is retrievable as a single answer. A statute does not have a meaning sitting in a database waiting to be pulled. It has a text, a history, and a contested space of readings, and interpretation is the work of arguing one reading is better than the others.

When Carpenter v. United States, 585 U.S. 296 (2018) wrestled with whether the Fourth Amendment reached cell-site location records, the disagreement was not about what the words said. Everyone could read them. The disagreement was about how a centuries-old text governs a technology its drafters never imagined. No model "knows" that answer because the answer did not exist until the Court argued it into being.

This is the conceptual error underneath the whole "can AI interpret statutes" debate. People assume that a big enough corpus plus a fluent enough model equals a correct reading. But coverage is not comprehension.

I made the coverage-versus-grounding case in our piece on AI legal research across all 50 states and federal sources. This post goes one level deeper: even with perfect grounding, even when you can click every citation, interpretation is still a reasoning task that the lawyer or the judge owns. That post asked "can you click the cite?" This one asks "even when you can, who decides what it means?"

What Newsom actually proposed (and what the critics fired back)

It is worth being precise about Newsom's concurrence because it gets flattened in both directions. He did not say judges should outsource meaning to a chatbot. He said LLMs are trained on staggering volumes of ordinary human writing, so they might serve as a rough proxy for how words are actually used, a supplement to dictionaries that freeze a single editor's choices in time.

He explicitly rejected mechanistic deference. The model is a witness to usage, not an oracle.

The pushback landed fast and it is serious. In "Judges Shouldn't Rely on AI for the Ordinary Meaning of Text" (Lawfare, 2025), the argument is that LLMs do not give you the median of human language. They give you the discretionary choices of their creators: what data went in, how it was filtered, how the model was tuned, what got reinforced. The piece notes ChatGPT's well-documented overuse of "delve" as a small sign of how far post-training can pull a model away from ordinary usage.

A dictionary at least tells you who the editor was. An LLM hides the editorial hand inside billions of parameters, which means a litigant with the resources to shape training data or prompt framing could quietly tilt the "ordinary meaning" a court relies on. That is not a neutral language census. That is a black box with opinions.

This is not a hypothetical academic spat. In Ross v. United States, a judge on the D.C. Court of Appeals cited ChatGPT in a dissent. Judges on the same benches now openly disagree about whether the technology belongs in the interpretive toolkit at all.

That live disagreement is the most honest signal in the whole field. The people closest to the work have not decided, and anyone selling you certainty is selling you something.

The empirical work backs up the caution. A team of linguists and law professors tested the idea head-on in "Large Language Models for Legal Interpretation? Don't Take Their Word for It" (Georgetown Law Journal, 2026). Their finding: model answers about ordinary meaning are sensitive to prompt phrasing and model choice in ways that make them unreliable as a neutral language census. Reframe the same question and the "ordinary meaning" can move.

Why Loper Bright raised the stakes

For forty years, Chevron v. NRDC, 467 U.S. 837 (1984) told courts that when a statute was ambiguous and an agency's reading was reasonable, the court deferred to the agency. Ambiguity was a feature. It handed the hard interpretive call to the experts at the agency, and courts could stop short of deciding the single best meaning.

Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024) ended that. The Court held that under the Administrative Procedure Act (the relevant hook being 5 U.S.C. 706), courts must exercise independent judgment and decide questions of law themselves, using the traditional tools of statutory construction. No more reflexive deference to the agency's reading. The court finds the best meaning, full stop.

Sit with what that does to the demand for careful interpretation. Under Chevron, "the statute is ambiguous, so the agency wins" was often the end of the analysis. After Loper Bright, ambiguity is the start of the hard part, not the escape hatch.

Every regulated party, every litigator, every agency lawyer now has to build the actual textual argument, ordinary meaning, structure, canons, history, because there is no deference backstop to catch a lazy reading. The post-Loper Bright world is one where getting the text exactly right matters more than it has in two generations.

There is a trap in this. When the demand for rigorous interpretation goes up, the temptation to shortcut it with a fluent machine goes up with it.

A model that can produce a polished three-paragraph statutory analysis in four seconds is exactly the kind of tool an overworked associate reaches for under deadline. That is precisely the moment the overreach risk bites hardest.

Try Vaquill AI on your matterBook a demo

A worked example: where the plain-text reading breaks

Take a real one. Fischer v. United States, 603 U.S. ___ (2024), turned on 18 U.S.C. 1512(c)(2), which makes it a crime to "otherwise obstruct, influence, or impede any official proceeding." Read that clause in isolation and the ordinary meaning is sweeping. Almost any disruptive act "otherwise obstructs" a proceeding.

Ask a general-purpose LLM "what does 1512(c)(2) cover" and you will likely get that broad reading back, fluent and confident. The plain words support it.

The Supreme Court read it the other way on June 26, 2024. The decisive move was structural. Subsection (c)(1) lists evidence-tampering acts (altering, destroying, concealing a record or document), and the canon noscitur a sociis says the catch-all "otherwise" clause in (c)(2) takes its color from the specific acts next to it. So the Court limited (c)(2) to conduct that impairs the availability or integrity of evidence, not any obstruction at all.

The model could have found the text, the subsections, and even the Fischer opinion itself. What it could not do is supply the judgment that the structural canon should override the broader ordinary-meaning reading. That call is the interpretation, and it lived nowhere in the statute until the Court made it.

Where AI actually earns its keep

So if AI is not the interpreter, what is it good for in this work? Quite a lot, as long as you keep it on the research side of the line.

Surfacing the exact text. Interpretation starts with the operative words, and the operative words are not always where you think. A section cross-references three others, which incorporate definitions from a fourth, which was amended in 2019.

Pulling the precise current text of 18 U.S.C. 922 or 17 U.S.C. 107, with every internal reference resolved, is exactly the kind of mechanical-but-error-prone task software should own. A good statutes corpus indexes the U.S. Code, the full CFR, all fifty states' codes, the Constitution, and the Federal Rules with direct links to underlying HTML, PDF, and XML text. That is not interpretation. It is the raw material interpretation requires, delivered without transcription errors.

Tracing amendment history. What did the provision say before the 2018 amendment? When did the key phrase enter the statute? Ordinary meaning is time-bound, and the meaning of a word at enactment can differ from its meaning today. Pulling version history quickly is a genuine accelerant.

Showing how courts construed a term. The most valuable interpretive input is often binding precedent that already glossed the word. That lives in case law, not in the statute. A deep US case-law corpus (8M+ opinions) is the layer that lets you find every opinion that construed "landscaping," "use," or "instrumentality" in the relevant context. The value is in grounded retrieval from a real, inspectable source, not invention from a model's memory.

Grounded answers, not remembered ones. The whole point of retrieval-augmented generation, which we walk through here, is that the model answers from documents it just fetched, with citations you can open, rather than from a hazy recollection baked into its weights. That is the difference between a tool that helps you interpret and a tool that hallucinates a reading and dresses it in confidence.

In every one of these cases the AI is doing retrieval, organization, and summarization. The human is doing the interpretation. That division of labor is not a limitation to apologize for. It is the correct architecture.

The overreach that ends in sanctions

We already have the cautionary tale, and it is not subtle. In Mata v. Avianca (2023), lawyers filed a brief full of cases ChatGPT had simply invented, complete with fake quotes and fake citations, and got sanctioned for it. The failure was treating a fluent output as a verified fact.

Statutory interpretation invites a quieter version of the same mistake. The model will not usually invent a statute. What it will do is hand you a confident, well-written paragraph about what a provision "means," and that paragraph can be wrong in ways that are much harder to spot than a fake case.

It can give you the raw ordinary meaning while missing that the rule of lenity flips the reading in a criminal context. It can ignore a structural argument that the surrounding sections make unavoidable. It can reproduce a training-data bias and present it as a neutral language census.

The output looks like interpretation. It has the shape and cadence of legal reasoning. But it has not weighed the canons, it has not checked the binding gloss, and it does not know which of the competing readings a court would actually adopt.

The ABA addressed exactly this exposure in Formal Opinion 512 (July 2024), which makes a lawyer's duty of competence and the obligation to verify AI output explicit.

The Stanford RegLab study on legal AI (Magesh et al., 2024) found leading legal research tools hallucinating at meaningful rates even with real corpora behind them, in the range of one in six queries for one major product and roughly a third for another. An interpretive answer is far harder to fact-check than a citation. You can confirm a case exists. Confirming that a reading is the legally operative one requires doing the interpretation yourself, which is the whole task you were trying to speed up.

The discipline is simple to state and hard to maintain under deadline: use the model to assemble the material, never to deliver the verdict. If a tool gives you a confident statutory reading, treat it as a hypothesis to test against the canons and the case law, not as an answer to file. The same verification reflex catches fabricated cites; we cataloged the sanctions that follow when it lapses in AI hallucination sanctions: lawyer cases and how to avoid them.

So, can AI interpret statutes?

Can AI interpret statutes? No, not in the sense that matters. It cannot decide the best meaning of a contested text, because deciding is an act of legal judgment that weighs canons, structure, and binding precedent against each other, and that judgment belongs to a lawyer and ultimately a court.

What AI can do, and do extremely well, is collapse the time between a hard interpretive question and the raw material you need to answer it. The exact text, every cross-reference, the amendment history, the cases that construed the term.

Newsom was right that the technology has a place at the table. The critics are right that the place is not at the head of it. After Loper Bright, with courts on the hook for the single best meaning and no deference to fall back on, the lawyers who win are the ones who use AI to get to the canons faster, not the ones who let it pretend to apply them.

AI is a research accelerant for interpretation. It is not the interpreter.

Build your workflow around that distinction and the technology is a genuine force multiplier. Forget it, and you are one confident paragraph away from the next Mata.

FAQ

Can AI interpret statutes?

Not in the way a court does. AI can read statutory text, summarize it, resolve cross-references, and pull the cases that construed a term, all fast. It cannot weigh the canons of construction, the legislative history, and the structural arguments to decide the single best meaning of a contested text. That judgment is the interpretation, and it belongs to a lawyer and ultimately a court.

What can AI do well with statutes?

The mechanical, error-prone parts of statutory research. It can pull the exact current text of a section with every internal reference resolved, trace amendment history to show what a provision said before a change, and surface the precedent that already glossed a key word. Those are retrieval and organization tasks, not interpretation.

Can AI read statutes accurately?

It can read and reproduce statutory text accurately when it retrieves from a real, structured source rather than answering from memory. The accuracy problem starts when a general-purpose model generates a reading from its training weights. The Stanford RegLab study (Magesh et al., 2024) found leading legal tools still hallucinating at meaningful rates even with real corpora behind them, so retrieval with openable citations matters more than fluency.

Did a judge really use ChatGPT to interpret a statute?

Yes. In a 2024 concurrence in Snell v. United Specialty Insurance Co., 102 F.4th 1208 (11th Cir. 2024), Eleventh Circuit Judge Kevin Newsom disclosed that he had consulted an LLM on the ordinary meaning of "landscaping." He framed it as one tool among dictionaries and usage corpora, not as the decider. A judge on the D.C. Court of Appeals later cited ChatGPT in a dissent in Ross v. United States.

Why is using LLMs for ordinary meaning controversial?

Critics argue an LLM does not give you a neutral census of how people use words. It reflects its makers' choices about training data, filtering, and tuning. A 2026 Georgetown Law Journal study found that model answers about ordinary meaning shift with prompt phrasing and model choice, which undercuts their use as objective evidence of meaning.

How did Loper Bright change AI statutory interpretation?

Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024) overruled Chevron deference and told courts to find the single best meaning of a statute themselves, using the traditional tools of construction. With no deference backstop to catch a lazy reading, getting the text exactly right matters more. That raises both the value of fast AI research and the cost of mistaking a fluent AI paragraph for actual interpretation.

Can I use AI to interpret a contract instead of a statute?

The same line holds. AI can extract clauses, compare them to a playbook, and flag deviations, which is genuinely useful. Deciding what an ambiguous term means against the rest of the agreement and the governing law is still legal judgment. Treat any AI reading as a hypothesis to verify.

Is there an API to pull statutory text for AI tools?

Not an official one for the full U.S. Code, but you can pull clean structured statute text programmatically. We cover the options, and why structured statute data is harder than it looks, in USC API: how to pull US Code sections programmatically.

If you want statutory research that stays on the right side of this line, Vaquill AI grounds answers in the actual U.S. Code, CFR, all fifty state codes, and the case law that construed a term, with citations you can open and check, so the model assembles the material and you keep the reading.

For related AI-in-practice and drafting coverage, see AI and Corporate Law in 2026 and Drafting the Schedule of Exceptions in an M&A Deal, and searching US statutes, regulations, and amendment history with AI.

Vaquill AI

Product & Content

Legal AI suite for US working lawyers: research, drafting, document comparison, document matrix, matters, and citation-verified answers, in one tool.