Think Like LiquidText: Turning Case Files into Proof Maps with Perplexity, Copilot and Meta AI

 How LiquidText’s active-reading method can be approximated with Perplexity, Microsoft Copilot and Meta AI—without surrendering judicial reasoning, confidentiality or source verification.

The basic idea

LiquidText is not merely a PDF reader. It places several source documents beside an expandable workspace, permits the reader to pull out excerpts, arrange them into groups, draw live “Ink Links,” compare distant passages and return from every note to its original context. It can import multiple documents into one project, display remote portions together and link notes or excerpts back to their sources.

A generative-AI tool approaches the same problem differently. Instead of creating a spatial, handwritten knowledge map, it converts the uploaded material into a conversational research workspace: the judge asks for an issue chart, chronology, contradiction matrix, evidence map or precedent table, and the system generates a structured textual answer. Thus, Meta AI, Copilot and Perplexity can reproduce much of LiquidText’s functional output, but not its tactile canvas, free arrangement of excerpts or persistent hand-drawn links.

The suggested “70–80 per cent replacement” should therefore be treated as a practical impression, not a measured product benchmark. AI may cover summarisation, extraction, comparison, classification and note drafting; LiquidText remains superior where visual thinking, manual control and an immediately visible chain between excerpt and source are central.

What each tool offers

Judicial-research need	LiquidText	Perplexity	Microsoft Copilot	Meta AI
Multi-document study	Project containing many documents and workspaces	Upload PDFs to a session or Project and question them together; internal files can be searched with or without the web	Upload PDFs and documents for summarisation and analysis; works closely with Word, Edge and OneDrive	Accepts PDFs and other files for summaries, comparisons and analysis
Excerpting	Reader manually pulls exact passages into the canvas	AI extracts issue-wise passages and can organise them into tables	AI summarises or rewrites material inside the Microsoft workflow	AI provides quick extraction and comparison
Connections	Visual Ink Links and mind maps	Produces textual links: agreement, contradiction, corroboration and distinction	Produces comparison tables and draft notes	Produces outlines and relationship summaries
Source return	One-tap live link to the original passage	Cited web research is a major strength; uploaded-file propositions must still be checked against the PDF	Convenient return to the document when used within Word or Edge	Best treated as a preliminary reading assistant
Best judicial role	Deep active reading and evidence mapping	Combined case-file analysis, web research and cited research notes	Final organisation, editing and conversion into a Word bench note	Rapid first-pass screening and Q&A

 

Perplexity is particularly useful for this purpose because it can combine uploaded material with web research in the same session, answer with cited sources, and use files stored in a Project as a continuing research set. Long uploaded documents may, however, be processed by extracting the portions considered most relevant; consequently, no page number, paragraph number, quotation or omission should be accepted merely because the answer appears precise.

The judicial workflow

Step 1: Create a secure case project

Create one private project for one case only. Use a neutral label such as “SC-12-2026” rather than party names, and insert standing instructions: use only the supplied record for factual statements; separate allegation, admission, proved fact and disputed fact; never invent a page, paragraph, exhibit or authority; state “not found” when the record is silent.

Before uploading anything, check the governing High Court directions and institutional authorisation. The Supreme Court’s 2026 AI Regulations remained a draft as of October 2026; the official Court page continued to describe them as draft regulations, and the Supreme Court itself noted in September 2026 that they had not yet been finalised. Nevertheless, the draft’s safeguards provide a sound working standard: AI should remain assistive, legal research and document summarisation require supervision and verification, and judicial outcomes must not be reached solely from AI-generated information.

Do not upload an unredacted charge-sheet, confession, medical record, child-victim material, sexual-offence record, address, phone data, sealed document or confidential deliberative note to a public consumer chatbot. Perplexity states that session uploads remain private unless the session is shared, but standard consumer accounts have their own retention and training settings; users should turn off AI data retention where applicable, keep the thread private and delete uploaded material after use. Product privacy assurances do not displace judicial confidentiality duties or a controlling High Court policy.

Step 2: Build a source index

Upload only authorised, searchable and preferably redacted copies. Begin with an inventory rather than a summary:

“List every uploaded document in a table with document name, date, author or witness, page range, exhibit number if visible, and subject. Do not infer missing metadata.”

The index becomes the digital equivalent of LiquidText’s document pane. It also reveals missing pages, duplicate statements, illegible scans and inconsistent pagination before substantive analysis begins.

Step 3: Generate a neutral case map

Ask for five separate tables: prosecution or claimant case, defence or respondent case, admitted facts, disputed facts, and procedural history. Then request a chronology in which every entry carries the document name and page or paragraph reference.

“Prepare a neutral chronology. For each event give: date/time, asserted event, supporting source, page/paragraph, whether admitted or disputed, and any inconsistent account. Do not decide which version is true.”

This distinction is crucial. AI may arrange competing versions; it must not silently convert an allegation into a fact.

Step 4: Frame an issue matrix

For a Sessions trial, headings may include identity, presence, motive, recovery, medical consistency, forensic linkage, common intention or object, statutory ingredients, defence explanation and sentencing factors. For a Magistrate, the matrix may instead track maintainability, statutory ingredients, notice or sanction, burden, presumptions, limitation, service, admissibility and the points requiring determination.

“Create an issue matrix with columns: legal ingredient, prosecution/claimant material, defence/respondent answer, supporting exhibit, contradiction or gap, governing provision, and question requiring judicial determination.”

This is the textual substitute for a LiquidText mind map. Each row links law, fact, evidence and the rival explanation without allowing the system to pronounce upon guilt, liability or credibility.

Step 5: Construct an evidence map

Request witness-wise and exhibit-wise summaries. For every witness, separate examination-in-chief, cross-examination admissions, contradictions, omissions, improvements, corroboration and formal matters. For documentary and scientific evidence, ask what the document proves directly, what requires oral foundation, and which link in custody or admissibility is disputed.

A useful contradiction prompt is:

“Compare Witness A’s statement at pages 31–36, deposition at pages 112–124 and medical evidence at pages 210–216. Quote only the minimum necessary words, identify exact page/paragraph references, and classify each difference as consistency, omission, contradiction, improvement or unresolved ambiguity. Do not assess credibility.”

The generated classification is only a locator. Whether a discrepancy is duly proved, material or legally consequential must be determined from the record and the applicable law by the judge.

Step 6: Research and verify law

Perplexity has an advantage at this stage because the same thread can use the case materials and current web sources. Ask it to identify controlling Supreme Court and jurisdictional High Court authorities, but require a proposition table containing the case name, court, date, neutral or official citation, relevant paragraph, proposition and treatment of earlier authority.

Every authority must then be opened on an official court website or a trusted legal database. The Supreme Court has adopted zero tolerance for unverified AI-generated precedent: reliance on fake or hallucinated material can invalidate the decision, while legitimate assistive use remains permissible. The Bombay High Court has likewise welcomed AI as a research aid while imposing responsibility to cross-check that every reference is genuine, relevant and actually exists.

Step 7: Prepare the bench note

Ask the AI to convert the verified material into a concise bench note—not a verdict. A safe structure is: background; rival submissions; points for determination; undisputed facts; evidence for and against each point; verified legal propositions; unresolved record questions; and a checklist of exhibits and citations requiring manual inspection.

Microsoft Copilot is useful here because the verified note can be refined in Word, reorganised under judicial headings and converted into the court’s preferred format. Copilot supports PDF and common document formats and can summarise, analyse or generate text from uploaded or cloud-hosted files. Meta AI is better used for quick screening or an additional comparison, rather than as the sole repository of the judicial research trail.

Step 8: Decide independently

Return to the original depositions, exhibits, statutory text and authenticated judgments. Record findings only after independently evaluating admissibility, burden, credibility, probabilities, presumptions, material contradictions and the parties’ submissions. Never ask an AI system to “decide guilt,” “recommend sentence,” “score flight risk,” “predict recidivism” or determine whether a witness is truthful; the draft regulations expressly reject algorithmic adjudication, sentencing without meaningful human control, risk scoring and credibility evaluation.

The safest division of labour

The most practical interim arrangement is not “LiquidText versus AI,” but a three-layer workflow:

1.      Perplexity as the research room: store authorised files in a private Project, create the chronology and issue matrix, compare passages, identify possible precedents and generate a source-linked research note.

2.     Copilot as the drafting desk: organise the verified note in Word, improve headings and language, and maintain the court’s preferred format.

3.   LiquidText, when available, as the visual evidence wall: manually pull decisive passages, connect witness versions to exhibits and keep the judge’s own reasoning visibly anchored to the record.

Meta AI can serve as a convenient rapid-reading assistant, especially for initial summaries and multi-file comparisons, but Perplexity is generally the stronger LiquidText substitute for legal research because it combines uploaded documents, continuing project instructions, web search and visible citations. Even then, its citations are navigation aids—not proof of the proposition.

The governing formula is simple: AI may extract, arrange, compare and question; the judge must verify, evaluate, reason and decide. Used within that boundary, these tools can reduce mechanical reading time and make a voluminous record more navigable. Used without confidentiality controls or source verification, they can introduce error into the very process they are meant to improve.

Print Page