research essay / Legal AI and Computational Law
Fundamental LLM problems in legal systems
Fluency, sources, time, and accountability
Legal LLM failures come not only from hallucinations but also from unclear corpora, temporal scope, and accountability boundaries.
In brief
- Fluency is not evidence.
- Source currency is part of the answer.
- Human oversight needs a real stop point.
Fluency hides uncertainty
A model can produce a coherent answer without access to the right source. In legal work, polished language increases risk because a reader may mistake confidence of tone for quality of evidence.
The interface should distinguish generated text, cited material, and procedural confidence. A generic warning below the answer is not enough.
The corpus defines the boundary
The system does not answer from “the law.” It answers from a particular collection of documents with a jurisdiction, date, completeness level, update rule, and exclusions.
Time is part of the answer
The same provision or judgment can carry different meaning before an amendment, after it, or during a transition. Retrieval without temporal control may return a semantically relevant but inapplicable source.
Each result should expose source date, index date, and the rule used to choose a version.
Accountability needs a stop point
Human oversight becomes real only when a person has material to inspect and authority to stop the process. An “approve” button is not oversight when sources and execution history remain inaccessible.
Architectural consequence
A safer system separates source registry, retrieval, model task, citation validation, and human decision. Each layer has its own failure, test, and owner. This material is for research and information and is not legal advice.
01 / METHOD
The question the test must answer
LLM risk in legal work begins before generation: with an unclear corpus, unknown currency, and no distinction between a retrieved passage and professional interpretation. The value of the material is that a reader can separate the user question from a convenient tool metric and identify the point at which a result is no longer sufficient for the next stage of work.
02 / METHOD
Material and counterexamples
Controls cover source scope, jurisdiction, effective time, citation, and the person responsible for review. A fluent response without that path does not become more reliable through a confident tone. Control material is not meant to confirm the author’s thesis; it should expose a case in which a plausible result is dangerously misleading for the workflow at hand.
03 / METHOD
Control method
A failure / control / test matrix separates retrieval error, wrong version, incorrect citation, overgeneralisation, and absent oversight. Each row has a detection condition and stopping point. The account includes step order, input, and result trace so an independent person can inspect what was measured and what the procedure does not measure at all.
04 / METHOD
Acceptance matrix
Material is accepted only when a representative result can be checked against corpus and citation by a qualified reviewer and known boundaries are visible. Criteria are stated before the result is read, so a single successful example cannot substitute for control of a critical failure type.
05 / METHOD
Handoff and rerun
After a corpus, model, or instruction change, the relevant tests must be rerun and the decision history retained. Replacing a model while repeating the same quality claim is insufficient. Reproducibility also means showing why a subsequent run differs from an earlier one and who decides whether a new version may be used.
06 / METHOD
Boundary of the conclusion
This is an engineering discussion of system controls, not legal advice or a compliance claim. It does not replace professional analysis of a particular matter. The material is therefore methodological: it does not pose as a universal recipe or as an independent confirmation that has not been performed.
07 / METHOD
A worked check
A control can begin with one result that sounds convincing but relies on a stale or inadmissible passage. Review checks corpus scope, date, citation, and interpretation boundary; until every step is traversed, the answer remains material for verification rather than a legal conclusion.
08 / METHOD
Source and review record
The test matrix should retain corpus, model, and instruction versions and name the person assessing representative cases. That enables a rerun after change. It does not create legal advice, establish regulatory compliance, or replace professional judgment in a particular matter.
LIM
Limitations and scope
This material is for research and informational purposes and does not constitute legal advice.
SRC
Sources and external edition
Original or earlier Medium material
AI tools assisted with structure and editing. The final text underwent human editorial review of facts, sources, conclusions, and attribution.
LOG
Change history
- First expanded edition in the owned archive.
- Review of structure, limitations, and evidence links.