Best Legal Research Tools in India 2026: We Tested 6 Legal AI Tools on 200 Questions — Here's What Hallucinated
Accuracy ranged from 71% to 94%; every tool invented at least one citation.
Key findings
- Tools grounded in a verified judgment database answered 88 to 94 per cent of questions correctly; general-purpose chatbots managed 71 to 76 per cent.
- Every tool produced at least one fabricated or misattributed citation; the worst produced 23 across 200 questions.
- The tools were strongest on settled doctrine and weakest on recent judgments, state amendments and anything requiring a date.
MUMBAI — Every legal AI tool sold to Indian lawyers claims to cut research time. None claims to be right every time, and in The Counsel's test none was. Over eight weeks we put 200 questions of Indian law — drawn from live practice in criminal, commercial, property, family, tax and constitutional matters — to six tools, scored each answer against the authoritative source, and counted every citation that could not be verified. The best tool answered 94 per cent of questions correctly and adequately supported. The worst answered 71 per cent. All six invented at least one case.
"Fast is not the same as right," a litigation partner who reviewed our scoring told The Counsel.
What we tested, and why the tools are not named
The six tools fall into three groups: two AI assistants built into established subscription databases of reported judgments; two standalone legal-AI products from Indian startups that retrieve from their own judgment corpora; and two general-purpose chatbots used by many lawyers without a legal subscription. We used each tool's current paid tier, asked each question once in the same form, and did not prompt for corrections. Two reviewers — a practising advocate and a former law clerk — scored each answer blind against the source: correct and supported, correct but unsupported or miscited, incomplete, or wrong. Every citation was checked against the e-SCR portal, the High Court websites or the reported series.
We sent each vendor its scores and the questions it got wrong. Two disputed our scoring of specific questions, one asked for a re-run on an updated model, and one declined to engage. We are publishing the results by tool category, with the tools identified by letter, and will publish the named table with the vendors' responses in a follow-up. That is not ideal, and readers are entitled to know it.
The results
| Tool | Category | Correct and supported | Wrong | Fabricated or misattributed citations | Median time per answer |
|---|---|---|---|---|---|
| Tool A | Database-integrated assistant | 94% | 3% | 1 | 22 s |
| Tool B | Database-integrated assistant | 91% | 4% | 3 | 18 s |
| Tool C | Standalone legal AI | 88% | 6% | 4 | 14 s |
| Tool D | Standalone legal AI | 84% | 9% | 7 | 11 s |
| Tool E | General-purpose chatbot | 76% | 14% | 15 | 9 s |
| Tool F | General-purpose chatbot | 71% | 17% | 23 | 8 s |
The remainder in each row was scored incomplete or correct but miscited. The pattern is consistent with every published study abroad: retrieval from a verified corpus reduces fabrication, and the tools that answer fastest are the ones that make things up.
What hallucinated
- Invented cases. Plausible party names with realistic citations — a "2019 SCC" reference to a judgment that does not exist, attributed to a bench that did not sit together. The general-purpose tools produced most of these; the database tools produced almost none, because they can only cite what they retrieve.
- Real cases, wrong propositions. The commoner and more dangerous error: a judgment that exists, cited for a holding it does not contain, or for the opposite of what it held. Every tool did this, including the best, and it survives a check that the case exists.
- Overruled and superseded law. Three tools cited judgments on the old criminal codes for questions about the BNSS without flagging the change; two cited a Supreme Court decision that a larger bench had overruled.
- State amendments. Questions turning on a state amendment to a central Act — stamp duty, rent control, land revenue — were the weakest category for every tool.
- Dates. Asked when a provision came into force, when a limitation period ran or when a notification issued, the tools were wrong or vague a third of the time.
Fast is not the same as right.
A litigation partner, to The Counsel
Where the tools earned their keep
On settled doctrine — the ingredients of an offence, the test for interim injunctions, the stages of insolvency resolution — the top four tools were accurate, well-organised and faster than a junior with a database. They were good at finding the leading case when given a proposition, at summarising long judgments, and at producing a first-draft research note that a lawyer could verify in a fraction of the time it would take to write one. The database-integrated tools were best at the thing that matters most: linking each sentence of an answer to a passage in a judgment the user can open.
How to choose
- Grounding first. Prefer a tool that retrieves from a verified corpus and shows the source passage for each proposition. A tool that cannot show its source is a drafting aid, not a research tool.
- Check the corpus. Ask which courts, which years and which series are covered, and how quickly new judgments are added. Coverage of High Court and tribunal decisions varies widely.
- Test it on your own practice. Twenty questions from your last month's files, scored against the sources, will tell you more than any vendor demonstration.
- Price the verification. A cheaper tool that fabricates more costs more in checking time; the time saved is only real if the output can be trusted.
- Confidentiality. Read the terms on retention and training. Client documents should not go into a public tool.
The free alternatives
None of this displaces the official sources. The e-SCR portal carries the Supreme Court's reported judgments with translations; the High Court websites and the National Judicial Data Grid carry orders and judgments by case number; and the open-access case databases, with their limitations, remain the most used research tools in the district courts. For an advocate without a subscription, the fastest reliable workflow in our test was a general-purpose chatbot for orientation followed by verification on the official portals — provided nothing from the chatbot reaches a filing unchecked.
What are the main legal research tools used in India?
Free official sources (e-SCR, High Court websites, NJDG, open-access databases), subscription databases with editorial headnotes and citators, and AI research assistants built into those databases or offered by legal-AI startups. General-purpose chatbots are widely used but are not research tools.
How accurate are legal AI tools for Indian law?
In our test, 71 to 94 per cent of answers were correct and supported. Tools grounded in a verified database did best; general-purpose chatbots did worst; every tool fabricated or misattributed at least one citation.
What is a hallucinated citation?
A case, citation or quotation the tool composed rather than retrieved — non-existent, or real but cited for something it does not say. Courts have imposed costs on advocates who filed them.
How should lawyers use AI research tools safely?
Find and summarise with them; verify every citation against the source; read the judgment before relying on a summary; prefer tools that show source passages; keep client data out of public tools; treat the output as a first draft.
The Counsel will publish the named results with the vendors' responses, and will repeat the test annually. Our methodology and the 200 questions are available to subscribers on request.
Nikhil Rao
Nikhil covers courts technology, legal AI and the business of law.
This report is for general information only and does not constitute legal advice. Readers should consult a qualified lawyer about their own circumstances. Case citations are to The Counsel Reports (TCR) unless otherwise stated.