Opening — Why this matters now

Legal interpretation used to belong to humans in black robes, law libraries, and late-night arguments about commas. Now it increasingly happens in chat windows.

As large language models (LLMs) enter legal practice—drafting contracts, summarizing judgments, and proposing interpretations—the question is no longer whether AI will assist legal reasoning. It already does. The real question is whether machines can interpret law in any meaningful sense.

This question sits at the intersection of two powerful trends: the digitization of legal knowledge and the rise of generative AI. Understanding where we are requires a look at how artificial intelligence has historically approached legal interpretation—and why each generation of technology solved one problem only to uncover another.

The evolution tells a story: from rigid rule engines, to structured argumentation systems, and finally to probabilistic language models. Each stage captures a different aspect of how humans reason about law.

Background — Context and prior art

Legal interpretation refers to the process of determining the meaning of legal texts—statutes, regulations, contracts—and applying them to real-world situations.

Even seemingly simple rules can create ambiguity. Consider the famous example often used in jurisprudence: a rule stating “No vehicles in the park.” Does it include bicycles? Electric scooters? A child’s tricycle?

Humans resolve these questions through interpretive reasoning that combines language, context, purpose, and precedent. AI systems have historically tried to replicate this reasoning using three different paradigms:

AI Approach Period Core Idea Limitation
Expert Systems 1970s–1990s Encode legal rules directly in a knowledge base Rigid and difficult to adapt
Argumentation Systems 1990s–2010s Model competing legal arguments and evaluate them Complex formal modeling
Machine Learning / LLMs 2020s– Generate interpretations through linguistic pattern recognition Lack grounded reasoning

Each approach reflects a deeper philosophical question: is legal interpretation primarily about rules, arguments, or language?

Analysis — What the paper shows

1. Expert Systems: When Law Became If–Then Logic

The earliest wave of AI and law research treated legal reasoning as a form of rule-based inference.

Expert systems stored legal knowledge in structured rules such as:


IF X is a vehicle THEN X is prohibited in the park

The challenge was not computing the rule but encoding the interpretation. Someone had to decide what counts as a “vehicle.” That responsibility typically fell to a knowledge engineer translating legal understanding into machine-readable rules.

This created an unexpected shift in interpretive authority.

Actor Role in Interpretation
Judge or lawyer Traditionally interprets the law
Knowledge engineer Encodes interpretation into system rules
AI system Applies interpretation consistently

The result was efficient but brittle systems. Once encoded, interpretive choices became rigid. If the system defined “vehicle” as including bicycles, it might also prohibit a child riding a tiny Christmas bicycle—an outcome a human officer might reasonably ignore.

Expert systems therefore demonstrated a paradox: formalizing law improves consistency but reduces flexibility.

The next generation of AI research moved away from rigid rules toward argumentation frameworks.

Instead of forcing a single interpretation, these systems represent competing arguments and evaluate which one prevails.

A simplified argument structure might look like this:

Argument Reasoning
Bicycle counts as a vehicle Ordinary language includes bicycles
Child’s bicycle is not a vehicle Purpose of rule is safety, not restricting children’s play

Argumentation systems represent these competing claims as graphs where arguments attack or support one another. The final interpretation emerges from the strongest surviving argument.

This approach mirrors how legal reasoning actually works in courts:

  1. Parties propose competing interpretations
  2. Arguments invoke interpretive canons
  3. Judges decide which reasoning prevails

Argumentation models therefore captured something earlier systems missed: law is inherently dialectical.

But they also introduced complexity. Representing legal reasoning required formalizing interpretive canons, argument strengths, and conflict resolution rules—an enormous modeling effort.

3. LLMs: Language Instead of Logic

The newest wave of AI takes a radically different path.

Large language models do not encode rules or argument graphs. Instead, they generate interpretations by predicting plausible text based on patterns learned from massive datasets.

When asked to interpret a legal rule, an LLM may:

  • identify ambiguous terms
  • suggest multiple interpretations
  • reference possible purposes of the rule
  • generate supporting arguments

This ability emerges not from explicit legal reasoning but from statistical linguistic competence.

Capability Source
Summarizing legal text Pattern recognition in language
Generating interpretations Learned discourse structures
Producing legal arguments Statistical imitation of legal writing

The result can be surprisingly convincing. A well-prompted model may generate reasoning that resembles what a trained lawyer might produce.

But the resemblance is only surface deep.

Findings — Strengths and weaknesses

The evolution of AI approaches to legal interpretation reveals distinct strengths and trade-offs.

Dimension Expert Systems Argumentation Models LLMs
Interpretive transparency High High Low
Flexibility Low Medium High
Knowledge requirements Manual encoding Formal modeling Large datasets
Reliability High if rules correct Medium Variable
Scalability Limited Limited Very high

LLMs outperform earlier systems in scale and linguistic fluency, but they lack grounding in legal knowledge or normative reasoning.

Two practical problems dominate real-world deployment:

Hallucination

LLMs may fabricate legal sources or produce plausible but incorrect arguments.

Reported studies find hallucination rates in legal queries ranging from roughly 17–34% in benchmarked tests.

Verification Cost

Any LLM-generated output must be checked by a human lawyer. This creates what scholars call the verification–value paradox:


AI saves time generating legal analysis ↓ Lawyers must verify every claim ↓ Time savings partially disappear

The system becomes useful primarily as a brainstorming tool rather than an authoritative interpreter.

Implications — What this means for AI systems

For organizations building AI-driven legal systems—or agentic AI more broadly—the implications are surprisingly clear.

LLMs excel at exploring interpretive possibilities but should not be treated as final authorities.

2. Hybrid architectures are likely the future

The most promising systems combine multiple AI paradigms:

Component Role
LLM Extract and summarize legal text
Knowledge base Store structured legal rules
Argumentation engine Evaluate competing interpretations

This hybrid architecture combines linguistic flexibility with logical rigor.

3. Human oversight remains essential

Legal interpretation involves normative judgment, social context, and moral reasoning—capabilities current AI systems simply do not possess.

AI can assist lawyers, but it cannot yet be a lawyer.

Conclusion — Machines can read law, but not fully understand it

The history of AI in legal interpretation reveals a recurring pattern: every technological leap captures one dimension of legal reasoning while losing another.

Expert systems captured rules.

Argumentation models captured debate.

LLMs capture language.

But law is all three at once.

For now, the most productive role for AI is not as an oracle delivering authoritative interpretations, but as a powerful companion helping lawyers explore the space of possible meanings.

That may sound modest. In reality, it is transformative enough.

Cognaptus: Automate the Present, Incubate the Future.