Your Case Exists. Your Argument Doesn't.

The Real Problem Isn't Fake Cases

By now, most attorneys have heard the cautionary tales. In Mata v. Avianca, counsel submitted a brief citing cases that didn't exist—fabrications generated by ChatGPT, complete with plausible-sounding party names and reporter citations. The resulting sanctions and disciplinary referrals were predictable.

Most lawyers are now on guard against fake cases. But the more dangerous problem isn't the citation that doesn't exist. It's the citation that does.

The Semantic False Positive: a real case, with a real citation, containing the exact words you searched for—that doesn't actually support your argument. The case exists. The quote exists. But the proposition you're advancing? The court never held that. It might have rejected it, hypothesized it, or attributed it to the losing party before dismantling it entirely.

This is where AI can fail silently. Large language models are pattern-matchers, not legal analysts. They excel at finding text that semantically resembles your query. But in long, structurally complex opinions, the full shape of the argument gets lost—a rejected proposition surfaces as established fact, a dissent reads like a holding. The model cannot distinguish between what a court said and what a court ruled—between binding precedent and the straw man the court constructed before knocking it down.

The "Easy" Problem vs. The "Hard" Problem

The Easy Problem: Hallucination

The AI invents a case. Smith v. Jones, 342 F.3d 891 (9th Cir. 2019). It sounds real. The citation format is correct. But the case doesn't exist.

• Detection: Easy. A Westlaw or Lexis search returns nothing. Even a free search on CourtListener or Google Scholar reveals the fabrication.
• Consequence: Sanctions, disciplinary referrals, and potential Rule 11 violations. But usually caught before filing if the attorney performs basic verification.

The Hard Problem: Misinterpretation

The AI returns a real case. The citation is accurate. The quoted language actually appears in the opinion. You can pull it up on Westlaw and Ctrl+F directly to the sentence.

But the case doesn't support your argument. It might actively undermine it.

• Detection: Hard. Requires reading the full opinion and understanding the context of the specific sentence—where it falls in the opinion's structure, what argumentative work it's doing, whether it represents the court's view or someone else's.
• Consequence: You file a brief citing real authority for a proposition the court never endorsed. Opposing counsel reads the actual case and eviscerates you in the reply. Or worse—no one catches it until the judge does.

Problem Type	What AI Returns	Detection	Consequence
Hallucination	Fabricated case	Easy (citation lookup)	Sanctionable
Semantic False Positive	Real case, wrong meaning	Hard (requires full read)	Sanctionable

The recent Cassata decision from Suffolk County included sanctions for two cases that existed but "do not support the proposition advanced by defendant." These weren't hallucinations. They were real cases, really cited, for claims they didn't actually establish.

Why Opinion Structure Defeats Pattern Matching

You already know the difference between holding and dicta, between a court's rule and the rejected argument it dismantled. The question is why AI struggles with these distinctions at the scale of real judicial opinions—and the answer lies in how these systems process text.

To a language model, a judicial opinion is a sequence of tokens. To you, it's a structured argument where position determines meaning. The same sentence carries completely different legal weight depending on whether it appears in the holding, the factual background, a jurisdictional survey, or the court's characterization of the losing argument.

In sufficiently long contexts—and nearly every opinion where this distinction matters is long enough—AI loses this structure. It sees text. It matches patterns. It cannot ask "what role does this sentence play in the opinion's architecture?"

Consider how often courts articulate the losing position clearly before rejecting it. "Landlords owe a duty of care to protect tenants from third-party criminal acts" might appear in an opinion—followed three sentences later by "We disagree." The AI found your keywords. It missed the negation. You cited the losing argument as the rule.

The Technical Root Cause

Tokenization: The Model Doesn't See What You See

Before a language model processes any text, it breaks the input into tokens—fragments that might be whole words, word pieces, or individual characters. The model doesn't "read" a judicial opinion. It processes a sequence of token IDs: numbers in a very long list.

The phrase "We hold that" and "Plaintiff argues that" look structurally similar to a tokenizer. Both are short phrases followed by a legal proposition. The model has no built-in representation of "this phrase signals a holding" versus "this phrase signals someone else's argument." It learned statistical patterns about when these phrases appear, but it cannot verify the distinction in a document it's never seen before.

Worse, legal terms often tokenize unpredictably. "Attorney-client privilege" might become three or four tokens. The model sees fragments, not concepts. It has learned that these fragments often appear together, but it doesn't "understand" privilege as a legal doctrine—it understands that certain token sequences have high probability of co-occurrence.

The Needle in a Haystack Problem

Modern language models can accept enormous context windows—some exceeding 100,000 tokens (roughly 75,000 words). A judicial opinion might be 10,000 words. In theory, the model can "see" the entire document.

In practice, attention degrades. The model's ability to correctly weigh information decreases as context length increases, and the location of critical information matters. Models perform worse when the relevant information ("the needle") is buried in the middle of a long document ("the haystack") rather than near the beginning or end.

Retrieval Doesn't Solve This

Many legal AI tools use Retrieval-Augmented Generation (RAG): they search a database for relevant passages, then feed those passages to the model. This helps with hallucination (the model can only cite what the retriever found), but it doesn't help with misinterpretation.

The retrieval step typically uses semantic similarity—finding passages whose meaning is close to your query. But semantic similarity has the same blind spot as the generation model. A passage discussing negligence elements is semantically similar to your query about negligence elements, whether that passage states the rule, rejects the rule, or attributes the rule to the losing party.

RAG systems also chunk documents into smaller pieces for retrieval. If the system retrieves a chunk containing "Landlords owe a duty of care to protect tenants," but the chunk boundary falls before "We disagree," you get the proposition without the negation. The model never saw the rejection because the retriever didn't include it.

No Structural Awareness

A lawyer reading an opinion knows where they are in the document's architecture. The "Background" section describes facts. The "Discussion" section analyzes law. The "Holding" section announces the rule. Headings, formatting, and position all signal what kind of text you're reading.

Language models flatten this structure. To the model, a sentence from the "Factual Background" and a sentence from the "Holding" are both just sequences of tokens. The model may have learned that certain phrases correlate with holdings ("We hold," "It is therefore ordered"), but across the length of a real opinion, it cannot reliably parse the document's structure and say "this section contains the binding rule."

When you ask an AI for authority supporting a proposition, it searches for semantic matches. At the length of actual case law, it has no reliable mechanism to filter for "only return sentences that function as holdings."

The Epistemic Status Blind Spot

Consider two sentences that might appear in the same judicial opinion:

1. "Negligence requires a showing of duty, breach, causation, and damages."
2. "The defendant argues that negligence requires a showing of intent, but this misstates the standard."

Both sentences are about the elements of negligence. Both contain relevant keywords. A semantic search for "elements of negligence" might return either one. But their epistemic status is completely different:

• Sentence 1 states the law.
• Sentence 2 describes a rejected misstatement of the law.

To the language model, these are just sequences of tokens with high semantic relevance to your query. In a short passage, the model might catch the distinction. Buried in a twenty-page opinion, the epistemic status of each sentence—holding versus rejected argument—is exactly the kind of structural signal that degrades with context length.

The Semantic Match Fallacy

If you ask an AI tool to find authority for the proposition that "employers may terminate at-will employees for any reason," it will scan for text matching that pattern. It might return a case containing:

The keywords match. The semantic content is relevant. The citation is real. And it says the opposite of what you need.

The Danger Zones

Certain patterns in judicial writing reliably mislead AI tools:

1. The Dissenting Opinion

Dissents are often more eloquent than majority opinions. The dissenting judge, freed from the burden of building consensus, can write with passion, clarity, and rhetorical force. These qualities make dissents more "quotable"—and more likely to surface in AI-generated research.

You cite a beautiful articulation of the principle you're advancing. It appeared in the case you cited. But it was written by the judge who lost the vote. The actual holding went the other way.

2. The Jurisdictional Survey

Before ruling on an issue of first impression, courts often survey how other jurisdictions have handled it. This creates paragraphs explaining the rule in California, the different rule in Texas, the split among the circuits.

The AI extracts a clear statement of the rule—but it's the rule from another jurisdiction that the court discussed before declining to follow it. You've cited a New York case for Texas law that the New York court explicitly rejected.

3. The Straw Man

Effective judicial writing often requires articulating the losing position clearly. The court states the rejected argument in its strongest form before explaining why it fails.

You search for authority that "consequential damages are recoverable in contract." The AI returns a case containing exactly that sentence. But the full paragraph reads: "Plaintiff contends that consequential damages are recoverable in contract without proof of foreseeability. This argument fails under settled law."

The AI found your keywords. It missed the word "fails."

Practical Defenses

The Ctrl+F Test Is Not Enough

Verifying that the quoted language appears in the opinion is necessary but nowhere near sufficient. You need to know what work those words are doing.

Contextual Reading

For any case you cite for a proposition, read the paragraph before and the paragraph after the quoted language. Look for signaling words that indicate argumentative structure:

• Negation signals: "However," "We decline to adopt," "This argument fails," "We disagree"
• Attribution signals: "Plaintiff argues," "The dissent contends," "Defendant maintains"
• Limitation signals: "In the specific context of," "Under the facts presented here," "We do not reach the question of"

If you're citing a case for a legal rule, find where the court actually announces that rule—not where it appears in factual recitation or argument summary.

Verify the Proposition, Not Just the Citation

Shepardizing or KeyCiting tells you whether a case is still good law. But it doesn't tell you whether the case stands for the proposition you're citing it for. A case can be valid, unchallenged precedent and still not support your argument—because your argument was never the holding.

For critical citations, verify not just that the case exists and hasn't been overruled, but that it actually establishes what you claim it establishes. This requires reading the case. (Before selecting any AI legal research tool, conduct thorough AI vendor due diligence to understand how the tool handles retrieval and citation.)

Conclusion

AI is a powerful retrieval tool. It can search vast databases and surface relevant documents faster than any human researcher. What it cannot reliably do—at the scale of real judicial opinions—is understand those documents: distinguish holding from dicta, rule from rejected argument, majority from dissent.

This limitation grows with document length. Models find patterns in text and match semantic content. In a short passage, they can often identify negation and attribution. Across the full architecture of a complex opinion, that ability degrades—and legal research rarely involves short passages.

The attorney who uses AI for legal research is using a librarian who can find every book on your topic—but cannot tell you which books are authoritative, which chapters are relevant, or whether the passage you're quoting is the author's thesis or a position the author is refuting.

The hallucinated citation is easy to catch and inexcusable after Mata. The misinterpreted citation is the deeper danger. The case is real. The quote is real. And it doesn't say what you think it says.

Can you use ChatGPT for legal research? Yes—but only as a starting point, never as the endpoint. Using AI for legal research without rigorous verification isn't just risky; after Mata and Cassata, courts have made clear it can constitute sanctionable conduct. The tool that finds your cases cannot reliably tell you what those cases mean.