TL;DR: A legal summary is only useful if you can trust it. The danger with general AI is the ungrounded summary - fluent, confident, and occasionally invented. Judicio takes a different approach: every file in the Library gets a cited summary plus extracted key details, concise mode finds the relevant passages while deep mode reads every page, and every claim links back to the page and quoted passage it came from. This guide explains grounded versus ungrounded summaries, when to use concise versus deep, and how to verify before you rely.
Summarisation is the most used and least examined AI feature in legal work. Everyone wants the gist of a long contract or a dense judgment in a paragraph, and large language models are very good at producing something that reads like one. The problem is that a summary that sounds right and a summary that is right are not the same thing - and in law the gap between them can be a misstated obligation or a citation to a case that does not exist. This guide is about getting summaries you can actually rely on, and the workflow that makes that possible.
What does good legal document summarization actually require?
A good legal summary has three properties that a generic one often lacks. It is faithful - it reflects what the document says, not what documents of that kind usually say. It is traceable - every material claim can be checked against the specific passage it came from. And it is appropriately scoped - it tells you what you need for the task at hand without smoothing over the qualifications that change the meaning. A summary that drops the carve-out in a limitation clause or the proviso in a holding is not concise; it is wrong in a way that is hard to catch.
Those properties point to a single requirement: grounding. A summary is trustworthy when it is tied to the source, so that the question is this accurate can be answered in seconds by reading the cited passage rather than re-reading the whole document. Everything else - length, tone, format - is secondary to whether you can verify it.
Why do ungrounded AI summaries hallucinate?
A general-purpose language model generates text by predicting plausible continuations, not by retrieving and quoting a source. Asked to summarise a document, it will produce a fluent paragraph whether or not it has reliably captured the specifics - and when a detail is uncertain, it tends to fill the gap with something that fits the pattern rather than flag the uncertainty. That is how a summary ends up asserting a governing-law clause the contract never contained, or attributing a holding to the wrong party. The model is optimised to sound right, and sounding right is not the same as being right.
This is not a hypothetical risk. Research from Stanford HAI has documented how often legal AI tools produce unsupported or fabricated content, and courts have sanctioned lawyers who filed material built on invented citations. The lesson is not to avoid AI summaries but to insist on grounded ones - summaries anchored to the document, where every claim can be traced back and checked.
The technical fix is to change what the model is doing. Instead of asking it to recall and compress, a grounded system retrieves the relevant passages from your document first and then writes a summary constrained to that material, attaching a citation to each claim. The summary can still be wrong - retrieval is not perfect, and condensation always loses something - but it can no longer invent a clause out of thin air, because every statement is anchored to a passage you can open and read. That shift, from open-ended generation to source-constrained summarisation, is the whole difference between a summary you must redo by hand and one you can verify in seconds.
Concise vs deep summaries: which do you need?
Not every summary needs the same depth, and matching the mode to the task saves time. A concise summary finds and condenses the relevant passages - ideal when you want to triage a file, decide whether it is relevant, or get oriented before a closer read. A deep summary has the AI read every page, which is what you want when the document is important, the details are dense, or a missed passage would matter. Judicio exposes exactly this choice as concise and deep modes across its tools.
| Summary type | What it reads | Best for |
|---|---|---|
| Concise | Locates and condenses the relevant passages | Triage, relevance checks, getting oriented quickly |
| Deep | Reads every page of the document | Important agreements, dense records, anything you will rely on |
The rule of thumb is to triage with concise summaries and commit with deep ones. Use the fast pass to decide what deserves attention, then spend the deeper read - and your own time - on the documents that carry weight.
A worked example makes the trade-off concrete. Faced with a hundred-document production, you would run concise summaries across the set to decide which thirty are worth attention - a fast, cheap pass whose only job is relevance. Then, for the handful that turn out to be central, you switch to a deep summary that reads every page, because in those documents a clause buried on page sixty could change your view of the matter. Spending deep-mode effort on all hundred would be slow and wasteful; spending it on the thirty that matter is exactly right.
How does grounded, cited summarization work in Judicio?
Grounding is built into how Judicio handles every file, not bolted on as a feature. The moment a document lands in the workspace, it is summarised and its key details are extracted, with citations throughout.
Per-file summaries in the File Library
Every file uploaded to the File Library is automatically given an AI summary - both a short version for a quick read and a longer one for more detail - and those summaries are cited back to the document. You do not have to ask: open a file and the summary is there, with a Read more for the fuller version, so you can grasp a hundred-page agreement or a long judgment before deciding how to use it. Because the summary points to its sources, checking any line is a click away.
Key-detail extraction alongside the summary
A narrative summary is only half the picture. Alongside it, the Library extracts the structured key details that matter in legal work: parties and their roles, key dates with deadline flags, monetary values, defined terms, clause types, governing law, courts, and jurisdiction, plus a section outline. This pairing - a prose summary for the story and typed fields for the facts - means you get both the shape of a document and its hard particulars in one place, each traceable to the source. For more on the structured side, see our guide to legal data extraction.
How do you summarize across many documents at once?
Summarising one file is useful; summarising a batch consistently is where real time is saved. Two tools handle scale. A Review Matrix can include a Summary answer type, so you ask for a one-line summary of the same point across multiple documents at once and get a column of cited, comparable summaries - the change-of-control position in every contract, say, summarised the same way for each. Document Review takes multiple files in a single run and can produce a summary of issues across the batch.
The advantage of summarising at scale this way is consistency: the same question is summarised in the same form for every document, so the results line up and the outliers stand out. Rather than reading the files in full for the gist of one clause, you read short, cited summaries in a single column and drill into the ones that look unusual.
This is also where summarisation and review blur together usefully. A Summary column sits next to typed columns in the same matrix, so for each contract you can read a one-line summary of an unusual clause beside a Yes/No on whether it exists and a Date for when it bites. The narrative and the structured answer reinforce each other: the summary tells you what is going on, and the typed cell lets you sort and filter the whole set by it. For a batch of agreements, that combination turns a week of reading into an afternoon of scanning and spot-checking.
What are the risks, and how do you verify a summary?
Even a grounded summary is a starting point, not the last word. The residual risks are real: a summary can compress away a qualification that changes the meaning, emphasise the wrong point for your purpose, or carry a subtle misreading of an ambiguous clause. The fact that a summary is cited does not guarantee the citation supports every word around it - which is precisely why the citation is there to be checked.
Verification is quick when summaries are grounded. Read the cited passage behind any claim you intend to rely on; for an important document, use the deep summary and still confirm the load-bearing points in the original; and be especially careful with anything that asserts an absence - that a clause is not present - since that is harder to verify from a summary alone. Treat the summary as an exceptionally fast first read by a junior, whose work you confirm before you act on it.
Where does summarization fit in a research workflow?
Summarisation is not an end in itself; it feeds the work that follows. In a research workflow, you summarise to triage a body of documents, identify the ones that matter, and pull their key details, then move into Legal Research, a matrix, or Drafting with the relevant material already understood and cited. Because one upload feeds every tool, the summary you read in the Library and the facts you extracted are the same ones available when you draft or research - no re-reading, no re-uploading.
This is the quiet payoff of grounded summarisation: it compounds. Each cited summary is not just a time-saver in the moment but a reliable building block for the next step, because you can trust where its claims came from. For the broader picture of reading documents with AI, see our overview of legal document analysis with AI and how to improve it.
Getting started with Judicio
Take one long document you would rather not read cold - a dense contract, a long judgment, a thick report - upload it into the File Library, and read the cited summary and extracted details before deciding how to use it. Then try a Summary column in a Review Matrix across a small batch to feel the difference between reading every file and scanning cited summaries side by side.
You can try it on your own files with a 7-day free trial - 500 credits, no credit card. Professional access is $200 per month for 5,000 credits; explore the full feature set or contact us for a walkthrough. The summaries get you to comprehension faster; the judgment about what a document means stays yours, and outputs are not legal advice.