Summarising long court judgments with AI: what works and what to check
Automated judgment summarisation is accurate enough to be useful and wrong often enough to need verification. Here is where the line sits.
Why judgment summarisation is harder than it looks
A judgment is not a document that summarises well by compression. Its useful content is unevenly distributed: a fifty-page opinion may turn on two paragraphs of reasoning, while thirty pages of undisputed facts matter only to the parties. A summariser that reduces length proportionally will faithfully shrink the parts nobody needs and flatten the part that decides the case.
The structural problem is that courts do not signal importance consistently. Reasoning appears inside the recitation of arguments; a holding may be stated once, in a subordinate clause, and then applied without restatement. Judgments also quote the parties' submissions and the findings below at length, and a model that does not track attribution will report a losing party's argument as the court's finding. That is the most common and the most damaging error in this category.
What a litigator wants is also rarely "what happened". It is closer to: which issue was dispositive, what standard the court applied, which facts it treated as material, and whether this judgment helps or hurts the matter in hand. That last part is not summarisation at all. It depends on the case the reader is working on, and no summary written without that context can answer it.
What current models get right
Within those limits it is worth naming the reliable capabilities precisely, because a workflow can only be built on those.
Structural extraction is stable: parties, procedural posture, court and date, the disposition, the issues the court framed for itself. Models are good at turning a long opinion into an outline that mirrors the judgment's own organisation, which makes navigation faster even when you intend to read the whole thing. They are good at answering targeted questions against a document that is open in front of them, such as what the court said about the limitation period, because that is retrieval with a checkable answer rather than synthesis.
They are also good at the cross-document work that is tedious rather than difficult: reading forty judgments and reporting which mention a particular provision, which reached the same outcome, and which are procedurally distinguishable. Triage across a large result set is where a lot of research time actually goes, and it is where the accuracy bar is lower, because a false positive costs one wasted read rather than wrong advice.
What is less reliable is the compressive judgement call — deciding which of the court's several reasons was the load-bearing one. That is the part to keep.
Where hallucination shows up, and how citation grounding helps
Hallucination in judgment work rarely looks like invention out of nothing. In a retrieval-based system the model has the text in front of it, so the failure mode is subtler and harder to catch: attribution drift, over-generalisation, and confident interpolation.
Attribution drift is reporting a submission, a lower court's finding, or a separate opinion as the holding. Over-generalisation is stating a rule more broadly than the court did, dropping the qualifiers that made it a narrow decision. Confident interpolation is filling a gap — a paragraph number, a date, a case reference — with something that fits the pattern. All three read fluently, and none of them announce themselves.
Grounding is the structural answer. If every proposition in a summary carries a pointer back to the paragraph it came from, verification becomes a targeted check instead of a re-read. Systems that quote the source span before paraphrasing it, and that decline to answer when nothing retrieved supports an answer, narrow the space in which these errors survive. Confidence indicators serve the same purpose: they tell you where to spend attention.
Grounding does not eliminate the problem. A citation can point at a real paragraph that does not in fact support the sentence attached to it. What it changes is the cost of catching that — from re-reading the judgment to reading one paragraph — and that is the difference that makes the workflow viable.
A verification workflow that does not cost back the time saved
Verification only pays if it is proportionate to the use. Treat it in tiers.
For triage — deciding which of sixty results to open — accept the summary without checking. The cost of an error is a wasted read and you will notice it immediately. For your own understanding of a matter, spot-check the propositions the summary presents as the holding, and any sentence stating a rule, by reading those paragraphs in the original. The rest can stand unverified, because nothing is being built on it yet.
For anything that leaves the firm — advice to a client, a pleading, a submission — check every citation and every quoted proposition against the primary text yourself. Not a second AI pass. A person, reading the judgment. This is not a limitation to expect the technology to remove; it is where professional responsibility sits, and it does not transfer to a vendor.
The practical effect is that AI moves work from the middle of the process to the ends. Finding and triaging get faster, the analysis that requires judgement takes as long as it ever did, and verification is narrow because grounding told you where to look. The way to lose the benefit is to verify everything at the highest tier — checking a triage summary as closely as one going into a brief, and spending the saving before it is realised.