Most AI systems operate on text: they retrieve documents and generate responses based on textual similarity. Legal knowledge graphs add a layer that text alone cannot provide: structured relational information about how legal authorities relate to each other.
This distinction matters profoundly for legal practice. A case citation in isolation is legally ambiguous — the case may be binding precedent, persuasive authority from another circuit, overruled on the specific point, or distinguished to near irrelevance. A knowledge graph encodes all of these relationship facts, allowing AI to surface not just relevant cases but legally valid and appropriately weighted authority.
The most familiar legal knowledge graphs are the citator systems that lawyers already use daily: Westlaw's KeyCite and LexisNexis's Shepard's Citations. These systems encode exactly the relationship information that characterizes a knowledge graph: this case has been cited by 47 subsequent decisions; these three decisions limit the holding to its specific facts; one subsequent decision has questioned the reasoning. Modern legal AI systems extend this citator model into a more comprehensive knowledge graph that also encodes relationships between statutory provisions, regulatory frameworks, secondary sources, and case law — creating a richer and more complete relational map of the legal landscape.
Understanding that certain legal AI tools use knowledge graphs — and what that means for the quality and reliability of their outputs — helps lawyers evaluate vendor claims about legal research accuracy and understand why grounded legal AI performs dramatically better than ungrounded text-generation models.
How It Works
Nodes and edges — the basic structure:
A knowledge graph is formally a collection of nodes (entities) connected by edges (relationships). In a legal knowledge graph:
Node types include: court opinions (identified by citation), statutory provisions (by code section), regulatory rules (by CFR citation), courts (by jurisdiction and level), judges (by name and appointment), parties (by name and role), and secondary sources (by publication and section).
Edge types (relationships) include: - Cites: Case A cites Case B in its opinion - Overrules: Case A overrules Case B on a specific legal point - Distinguishes: Case A distinguishes Case B because of factual differences - Limits: Case A limits Case B's holding to narrower circumstances - Applies: Case A applies Statute X to the facts before the court - Interprets: Case A interprets the meaning of Statute X or Regulatory Rule Y - Amends: Legislative Act Y amends Statute X, effective as of a specific date - Supersedes: Regulation A supersedes Regulation B upon effective date
From citators to knowledge graphs:
Westlaw's KeyCite represents a targeted knowledge graph focused specifically on citation relationships — the citing references for any case, with editorial annotations characterizing the nature and strength of the citation treatment. This is a graph of cases connected by citation relationships with editorial labels on the edges.
A more comprehensive legal knowledge graph extends this model to include statutory and regulatory relationships (statute amended by act, act implemented by regulation, regulation interpreted by agency guidance), doctrinal relationships (legal test A applies to fact pattern B in jurisdiction C), and procedural relationships (court D has jurisdiction over subject matter E in geographic area F).
How knowledge graphs enhance legal AI:
When a legal research AI incorporates knowledge graph data, it can provide qualitatively richer outputs:
-
Precedent weighting: Instead of returning a list of cases matched by text similarity, a knowledge graph-enhanced AI can weight results by precedential authority — mandatory authority in the specific court and jurisdiction appears prominently; overruled cases are flagged or excluded.
-
Negative treatment detection: The AI automatically identifies cases that have been overruled, limited, or questioned, preventing the presentation of superseded authority as valid law — a failure mode that ungrounded LLMs frequently produce.
-
Statutory context: The AI understands that a statute referenced in a case has subsequently been amended, and can advise the user whether the amendment affects the case's current authority.
-
Jurisdictional filtering: Knowledge graph data about court hierarchies and jurisdictional coverage allows the AI to distinguish mandatory from persuasive authority for a specific practice context.
Legal knowledge graphs in major platforms:
Westlaw Precision AI integrates Thomson Reuters' knowledge graph — built from decades of editorial work on the KeyCite citator and the West Key Number System — into its AI legal research functionality. The West Key Number System is itself a knowledge graph: a hierarchical taxonomy of legal concepts with thousands of nodes encoding relationships between legal doctrines and the cases that apply them. Lexis+ AI integrates LexisNexis's Shepard's Citations knowledge graph and its broader legal concept taxonomy into its AI research capabilities. Casetext uses case law graph data to power its CARA AI semantic research, enabling the system to surface cases based on factual and legal concept similarity with precedential relationship awareness.
The cost of building and maintaining knowledge graphs:
Legal knowledge graphs require enormous editorial investment to build and maintain. Every case must be read and annotated; every statutory amendment must be tracked; every overruling decision must be reflected in updates to the graph. This is why the most comprehensive legal knowledge graphs are maintained by the major legal database vendors (Thomson Reuters, LexisNexis) rather than by AI startups — the editorial infrastructure to build and maintain an authoritative legal knowledge graph represents decades of investment and thousands of editorial staff.
Key Considerations for Law Firms
Knowledge graph coverage determines AI reliability for jurisdiction-specific work: The breadth and depth of the knowledge graph underlying a legal AI tool determines how reliably it can identify valid authority and flag negative treatment for specific jurisdictions. A tool with comprehensive federal case law coverage may have less complete state court coverage or may miss administrative law decisions. Evaluate knowledge graph coverage specifically for the jurisdictions and practice areas most relevant to your work.
Freshness matters for knowledge graphs: A knowledge graph that is not updated to reflect recent case law developments, statutory amendments, and regulatory changes will give the AI an inaccurate picture of current legal validity. Ask vendors about their knowledge graph update frequency and how long it takes for a new court decision to be reflected in the graph.
Knowledge graphs vs. retrieval accuracy: Even with a knowledge graph, the AI must correctly interpret the relationships encoded in the graph and apply them to the specific legal question being asked. A knowledge graph reduces but does not eliminate the risk of presenting overruled authority or missing mandatory precedent — the AI must correctly use the graph data it has access to.
Proprietary knowledge graphs as competitive moats: The major legal research platforms' competitive advantage is substantially built on their proprietary knowledge graphs — the West Key Number System and Shepard's Citations represent intellectual property developed over decades that is not publicly replicable. When evaluating legal AI tools, the depth of their underlying knowledge graph is often the most important differentiating factor for legal research quality.
Limitations and Risks
Knowledge graph construction is never complete: Legal knowledge graphs are built incrementally and always have coverage gaps — older decisions that predate systematic annotation, niche jurisdictions with limited coverage, emerging regulatory areas where graph relationships are still being established. A legal AI tool is limited by the completeness of its underlying knowledge graph.
Editorial error in knowledge graph construction: Human editorial judgment determines how case citations are characterized in a knowledge graph. An incorrect characterization — classifying a case as "distinguished" when it was actually "overruled" — will propagate through any AI analysis that relies on that relationship. Knowledge graph error rates, while low for the major vendors' well-established systems, are not zero.
Dynamic legal landscape creates maintenance challenges: The law changes constantly. Every legislative session, every regulatory update, every new court decision potentially changes relationships in the knowledge graph. Keeping the graph current is an ongoing editorial challenge, and there will always be a lag between a legal development occurring and that development being reflected in the knowledge graph.
Complexity of legal relationship types: The edges in a legal knowledge graph represent relationships with legal significance that is often context-dependent. A case that "distinguishes" another case on its facts may still be highly relevant to a matter where the factual distinction does not apply. Knowledge graph relationships are labels that simplify complex legal reasoning, and AI systems that treat those labels as determinative may miss important nuance.