Case Law Research

From Keywords to Meaning

Traditional legal research works like searching a dictionary — you type exact words and hope the opinion you need uses those same words. If the court said "promissory estoppel" but you searched for "detrimental reliance," you miss it. This is the fundamental limitation of keyword search, and it is the reason attorneys spend hours on Westlaw refining search queries.

Semantic search changes this entirely. Instead of matching exact words, AI understands the meaning behind your query. Search for "employer's duty to accommodate religious practices" and AI finds opinions discussing "undue hardship," "reasonable accommodation," and "Title VII" — even if your exact phrase appears nowhere in those opinions.

This chapter explains how semantic legal search works, which databases and AI tools offer it across US and UK jurisdictions, and — critically — how to verify that the cases AI finds actually exist.

How Semantic Legal Search Works

Think of it as a two-step process:

Step 1: Understanding. When you type a query, AI converts your question into a mathematical representation of its meaning — a vector of numbers that captures concepts, not just words. "Landlord evicting tenant for non-payment" and "lessor seeking possession for rent arrears" end up as similar vectors because they mean similar things.

Step 2: Matching. AI compares your query vector against pre-computed vectors for millions of opinion paragraphs. The closest matches by meaning — not by keyword overlap — appear first. This is why semantic search finds relevant precedents that keyword search misses entirely.

Open data/case-law-search-examples.json in the code panel. This file contains 15 sample queries with their keyword search results vs semantic search results, showing the difference in relevance and completeness.

Citation Formats by Jurisdiction

Before using any AI research tool, you must understand citation formats so you can verify what AI returns:

US Citations (Bluebook Format)

UK Citations (OSCOLA Format)

EU Citations

AI tools sometimes return citations in non-standard formats or mix up citation styles. If AI returns "Supreme Court, 2013, voting rights case" instead of "570 U.S. 529 (2013)," that is a sign the AI is summarizing rather than citing — and the citation needs verification.

Court Hierarchies

AI research tools must account for court hierarchy, and you must verify that AI weights precedents correctly:

US Federal Court System

US Supreme Court
  ↓ Binding on all federal and state courts (on federal questions)
US Circuit Courts of Appeals (13 circuits)
  ↓ Binding within circuit
US District Courts (94 districts)
  ↓
Magistrate Judges / Bankruptcy Courts

State courts operate in parallel, with their own three-tier hierarchy (trial court → intermediate appellate → supreme court). A California Supreme Court decision is binding in California state courts but merely persuasive in federal court — unless it interprets California state law.

Circuit splits are critical: when the Second Circuit and the Ninth Circuit reach opposite conclusions on the same federal question, AI may not flag this conflict. Always check whether the circuit where your case is pending has addressed the issue.

UK Court System

UK Supreme Court (formerly House of Lords)
  ↓ Binding on all UK courts
Court of Appeal (Civil and Criminal Divisions)
  ↓ Binding on lower courts; bound by own previous decisions (with exceptions)
High Court (King's Bench, Chancery, Family Divisions)
  ↓ Persuasive on other High Court judges; binding on lower courts
County Courts / Crown Courts

A common AI error is treating a High Court judgment as equivalent to a Court of Appeal judgment on the same point. AI may also fail to distinguish between *ratio decidendi* (binding) and *obiter dicta* (persuasive) within a judgment.

Legal Research Databases

Here are the primary databases where you verify AI-found citations:

Westlaw (Thomson Reuters)

The dominant US legal research platform. Comprehensive coverage of federal and state case law, statutes, regulations, and secondary sources. KeyCite citation service tracks the history and treatment of every case. Now integrated with CoCounsel AI. Best for: US federal and state research, citation verification, regulatory tracking.

LexisNexis

The other major US legal research platform. Comparable coverage to Westlaw with Shepard's Citations for case treatment analysis. LexisNexis+ AI integrates generative AI into the research workflow. Best for: parallel research to Westlaw, news and public records, international coverage.

PACER / RECAP

Public Access to Court Electronic Records — the official repository for US federal court filings. RECAP (via CourtListener) provides free access to many PACER documents. Best for: docket searches, verifying case filings, accessing motions and briefs.

Google Scholar (Case Law)

Free access to US case law (federal and state). Limited compared to Westlaw/Lexis but useful for quick verification. Does not track case treatment (overruled, distinguished). Best for: quick citation checks, preliminary research, budget-constrained research.

Practical Law (Thomson Reuters) / LexisPSL

Practice-oriented resources with precedent documents, practice notes, and checklists. Strong UK coverage. Best for: transactional practice, understanding market-standard positions, cross-jurisdictional comparison.

BAILII

British and Irish Legal Information Institute — free access to UK case law. Best for: verifying UK citations, reading full judgments, preliminary research.

Open data/legal-databases-comparison.json to see a detailed comparison of these databases by coverage, cost, AI features, and suitability for different types of legal research.

The Hallucination Problem

This is the most dangerous aspect of AI legal research: AI can fabricate citations that look completely real but do not exist.

Here is an actual pattern of AI hallucination in legal research:

You ask: "Find federal court decisions on the enforceability
of non-compete agreements in the technology sector."

AI responds: "In Nitro-Lift Technologies v. Howard, 568 U.S.
17 (2012), the Supreme Court held that state courts must
give effect to the FAA when considering non-compete
agreements with arbitration clauses..."

This citation is REAL and the holding is broadly accurate.

But AI might also say: "In TechVentures Inc. v. Morrison,
847 F.3d 562 (9th Cir. 2019), the Ninth Circuit held that
overbroad non-competes in technology employment are per se
unenforceable..."

This citation may be FABRICATED — the case name sounds
plausible, the citation format is correct, but the case
may not exist.

The danger is that fabricated citations are formatted correctly and sound authoritative. An attorney who cites a non-existent case in a brief faces sanctions, professional embarrassment, and potential disciplinary action. The Southern District of New York sanctioned attorneys in *Mata v. Avianca, Inc.* (2023) for submitting a brief containing AI-generated fictitious case citations — a cautionary tale for every legal professional.

Verification Protocol

Every AI-found citation must go through this verification:

Prompt Engineering for Legal Research

Here are effective prompts for using general AI tools for case law research:

Prompt: "I need to find US federal court decisions on
[legal issue]. For each case:
1. Provide the exact citation (Bluebook format)
2. Identify the court and year
3. State the key holding in one sentence
4. Note if this case has been overruled or distinguished
   by a later decision (if you know)
5. Indicate confidence level — are you certain this case
   exists or is this a reconstruction from memory?

Important: If you are not certain a case exists, say so
explicitly. Do not fabricate citations."

The last instruction is critical. Explicitly asking AI not to fabricate reduces (but does not eliminate) hallucination. Always verify regardless.

Building a Research Workflow

The most effective legal research workflow combines AI speed with human verification:

Total: 1.5-2 hours for research that would take 6-8 hours with purely manual methods.

Key Takeaways