An AI reading system helped non-experts move through dense court judgments faster and made them feel they understood the material better, but objective comprehension did not reliably improve. The finding comes from JudgmentLens, a preprint describing an AI-augmented interface designed to help readers connect facts, evidence, reasoning and rulings in complex legal decisions.

AI legal judgments can become easier to navigate without becoming easier to master

The researchers began with a formative study involving 34 survey participants and six interviews with Chinese non-expert readers. They identified difficulties in following structure, interpreting legal reasoning, verifying explanations and deciding what information mattered. JudgmentLens was built around those problems rather than simply adding a chatbot beside a PDF.

The system creates persistent representations of the case, offers adaptive explanations and links generated interpretations back to passages in the source judgment. In a counterbalanced within-subject evaluation with 16 participants, people completed tasks faster than when using conventional PDF reading and reported lower workload and greater perceived understanding.

The important counterweight is the scored result. Rubric-based comprehension did not differ reliably between the two conditions. The participants felt better supported and moved faster, but the experiment did not show a dependable gain in how much legal content they actually understood according to the study’s objective measure.

Perceived understanding can be a useful benefit and a safety risk

Lower workload is valuable. Legal judgments are structurally difficult documents, and an interface that helps people locate relevant reasoning can make public information more accessible. The problem arises if a smoother experience increases confidence beyond what the reader has actually learned.

That gap appears in other forms of AI assistance. A system can reduce friction, summarise complexity and make a document feel manageable without necessarily improving the user’s independent ability to verify a conclusion. Good design therefore needs to preserve uncertainty and source access rather than replacing the judgment with a polished answer.

LiveAIWire has documented the consequences of hallucinated legal citations, where fluent output was relied on despite containing nonexistent authorities. JudgmentLens is designed differently because its generated interpretations are tied back to passages in the judgment. Traceability does not guarantee correctness, but it gives the reader somewhere concrete to check.

The system was designed around source-grounded inspection

A generic chatbot encourages a question-and-answer pattern. That can work well when the user already knows what to ask, but legal comprehension often fails earlier. A non-expert may not know which issue is decisive, how a fact connects to a rule or where a court limits its own conclusion. JudgmentLens attempts to keep those relationships visible.

The researchers also ran an exploratory comparison with PDF plus DeepSeek. They report that conversational AI could support questions once they were formulated, while the user still carried much of the burden of deciding what to ask, integrating answers and checking the source. That distinction helps explain why a purpose-built interface may reduce workload without automatically increasing comprehension scores.

The design principle extends beyond law. When AI mediates a difficult source, the interface can either hide the document behind a conversational layer or help the reader move between interpretation and evidence. The second approach makes mistakes easier to inspect and preserves more of the user’s ability to challenge the system.

Sixteen participants cannot settle the question

The evaluation is small, and the study concerns Chinese legal judgments and non-expert readers. It does not establish how lawyers, judges or litigants would perform, or whether the same interface would help with another jurisdiction’s legal structure. The paper is a preprint and should be treated as an early human-computer-interaction result rather than a general claim about legal AI.

Task speed also needs interpretation. Completing a legal-reading exercise faster is beneficial only if the user does not miss an important caveat. In professional work, an apparently slower process may be justified when verification matters. The right measure depends on whether the system is supporting orientation, education, advice or a decision carrying legal consequences.

LiveAIWire has also covered California legislation that would require lawyers to verify AI-assisted court filings and disclose generative AI use if the measure becomes law. That policy debate concerns professional responsibility. JudgmentLens sits earlier in the chain, showing how interface design can affect what ordinary readers believe they understand before any professional judgement is involved.

The best legal AI may make verification easier, not invisible

The study suggests a useful standard for legal reading tools. Success should not be measured only by whether users prefer the interface or finish tasks quickly. Designers should also test whether people can locate evidence, detect errors and explain the reasoning in their own words.

A system that leaves comprehension unchanged can still be useful if it lowers the cost of reaching the relevant parts of a document. The danger is describing that convenience as deeper understanding without evidence. In the JudgmentLens experiment, the subjective and objective measures separated, and that separation is the most informative result.

AI can make a difficult judgment feel more approachable. The next challenge is ensuring that approachability leads users back to the source rather than giving them a more confident shortcut around it.

Legal reading tools should test whether confidence is calibrated

A user who understands a judgment poorly but knows they are uncertain may seek help. A user who understands it equally poorly but feels highly confident may act on a mistaken interpretation. That makes calibration an important outcome for legal AI. Future studies could compare confidence with scored comprehension and examine whether the system helps people recognise when they should verify a point or consult a professional.

Source links can support that calibration if they are designed as part of the workflow. Instead of placing citations at the end of a generated explanation, an interface can let readers jump directly to the paragraph, highlight the relevant wording and show competing passages that limit the conclusion. The goal is to make verification easier at the moment a claim is encountered.

Accessibility and legal advice are not the same thing

Making a court judgment easier to navigate can broaden access to public information without turning the software into a substitute lawyer. That distinction is important because a judgment describes a particular dispute under a particular legal framework. A reader may understand the court’s reasoning accurately and still be wrong to apply it directly to their own circumstances.

Purpose-built interfaces should therefore be clear about the task they support. Explaining structure, locating passages and defining terminology are different from recommending litigation strategy or predicting an outcome. The more a product moves towards advice, the more important professional responsibility, jurisdiction and current law become.

JudgmentLens offers an encouraging design direction because it keeps the primary document visible. Its evaluation also supplies a useful warning: a better experience is not automatically a deeper understanding. Legal AI will be safer when products measure both.

There is also a design opportunity in showing uncertainty directly. A legal reading tool could indicate when several passages support different interpretations, distinguish quoted law from generated explanation and flag when the system is summarising rather than reproducing the court’s reasoning. That would make uncertainty part of the interface instead of hiding it behind fluent prose.

Future evaluations should also test retention. A user who completes a task quickly with AI support may still struggle to explain the judgment later without the tool. Measuring delayed understanding would help separate genuine learning from temporary navigation assistance. For public legal information, both can be valuable, but they are different outcomes and should not be reported as though they were the same.

The experiment also points to a useful distinction between access and comprehension. Search, highlighting and generated explanations can reduce the effort required to locate the relevant part of a judgment. That is valuable in its own right. But comprehension asks whether the reader can accurately reconstruct the reasoning, recognise the limits of the decision and distinguish the court’s words from the tool’s interpretation. Interfaces should measure those outcomes separately rather than allowing speed or confidence to stand in for understanding.

That distinction matters for courts and publishers too. A reading aid can improve access without claiming that generated explanations carry legal authority. Keeping the judgment itself one click away, clearly labelling generated interpretation and testing users on what the court actually decided are practical ways to preserve that boundary.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.