Researchers Built an AI That Reads Law Like a Road Map

Legal questions are unusually hard for AI. Today's look-up tools match laws by wording similarity, which misses how laws actually relate to each other — one rule overrides another, one is an exception, another only applies in specific situations. And when AI "thinks step by step," it can produce confident-sounding explanations that aren't legally sound at all. LEGO tackles both problems: a hand-built map of how civil code provisions connect, plus a strict template that forces the AI to lay out the provision, the facts, and only then the conclusion.

The team ran it on a small open model called Qwen3-8B — small enough to be affordable to run, unlike the giant systems from big tech labs. It scored 40.53% exact-match accuracy on a set of Chinese civil law exam questions, beating similar look-up and reasoning setups and performing comparably to much larger models. It also came out on top among tested systems on open-ended legal benchmarks, and handled questions that require chaining several laws together without falling apart.

Why should you care? Legal help is expensive, and much of the cost is the slow work of finding the right rule and applying it to your facts. A cheap AI that reliably does that first pass could lower the price of reviewing a lease, checking an employment contract, or understanding a small-claims dispute. Think of it as a fast, tireless paralegal that drafts the first answer — cheap, always available, and never annoyed by your questions.

The honest catch is accuracy. Getting roughly 40% of exam questions exactly right means it's wrong far more often than right on the hardest cases, which is a serious problem in a field where a wrong answer can cost someone their home or their business. It also depends on experts hand-building the law map, which is labor-intensive and specific to one legal system, so it doesn't automatically work for every country. The realistic near-term picture is AI assisting lawyers and paralegals, not replacing them.