Why Native Legal AI is Needed and Why Generic AI Will Keep Failing You
The pitch is always the same: "It's trained on millions of legal documents." The demo is always impressive: you type a question, it retrieves something that sounds right, the citation looks plausible, the ratio is coherent. Then you check the source. The case doesn't exist. Or it exists but says the opposite. Or the docket number belongs to a completely unrelated matter. Or the High Court whose authority you just relied on has no jurisdiction over your dispute.
This is the documented, recurring, and increasingly expensive failure mode of generic AI applied to Indian legal practice. India's legal system needs AI built from the ground up for Indian law, structured, hierarchical, jurisdictionally aware, and citation verified. The gap between that and what most advocates are currently using is a professional liability gap.
What an Empirical Study Actually Found
In January 2025, Rahul Hemrajani, Assistant Professor at the National Law School of India, published a peer-reviewed study "Evaluating the Role of Large Language Models in Legal Practice in India" it systematically tested GPT-4, Claude, and ChatGPT on key Indian legal tasks: issue spotting, legal drafting, advice, research, and reasoning.
The results were precise and not flattering. LLMs performed well on drafting and issue spotting. But on specialized legal research, they "frequently generated hallucinations; factually incorrect or fabricated outputs. "The study's conclusion: "while LLMs can augment certain legal tasks, human expertise remains essential for nuanced reasoning and the precise application of law."
This study was done in controlled conditions, with researchers who understood what they were testing. In actual practice, with advocates under time pressure, the hallucination problem is worse.
The Architecture of the Problem
India's legal system is not just large. It is structurally complex in a way that punishes generic AI particularly hard. Competent legal research in India requires understanding the hierarchy between Supreme Court, High Court, and tribunal decisions; knowing which High Court's view governs a particular jurisdiction; tracking whether a judgment has been followed, distinguished, overruled, or merely cited; reading legislation against its amendment history; and navigating state-specific procedural rules running alongside central law.
Deepak Kapoor, CEO of Manupatra, put the problem directly in April 2026: generic AI "often faces several limitations in the legal domain," including "inability to distinguish binding vs persuasive precedent" and "misinterpretation of legal context."²
Indian legal context is a specific epistemic structure: how courts relate to each other, how precedent flows between them, what happens when they conflict.
What India Is Actually Building
SUPACE - the Supreme Court Portal for Assistance in Court Efficiency — is an AI research tool built specifically for judicial use. SUVAS has translated 36,271 Supreme Court judgments into Hindi, opening the jurisprudential corpus to speakers of India's most widely spoken language. LegRAA (the Legal Research Analysis Assistant), developed by the National Informatics Centre in Pune, is another judiciary-facing tool. Digital Courts 2.1 integrates judgment databases with ASR-SHRUTI (automated speech recognition) and PANINI (translation engine).³
Commercial native platforms have followed: Manupatra's conversational engine draws on over 4 million judgments across 400+ databases, with every AI output traceable to a primary source. LegitQuest uses iDRAF technology for document analysis. Kanoon.ai and LAWFYI's Legal Aid chatbot extend this into the access-to-justice space. Each of these made a different choice from the generic-AI shortcut: build from the corpus up, not from the global model down.
The Language Problem Is a Justice Problem
Over 90 percent of Supreme Court and High Court judgments are in English. According to the 2011 Census, only about 10 percent of India's population speaks English. Most Indian litigants cannot read their own case files. They cannot understand the order that changed their life. They cannot evaluate whether their lawyer gave them the right advice.
SUVAS's translation programme is an answer to this, but 36,271 translated judgments out of a corpus of millions is a start, not a solution. Native legal AI that can translate, summarise, and explain Indian jurisprudence in Hindi, Tamil, Telugu, Marathi, Kannada, and Bengali is not a luxury feature for large law firms. It is the basic infrastructure for a justice system that claims to serve the people governed by it.
The Budget Reality
India has committed Rs. 53.57 crore (approximately $6.4 million) for AI and blockchain across High Courts through 2027, as part of a Rs. 7,210 crore e-Courts Phase III budget. The government is building public-sector native legal AI on a fraction of what a single Silicon Valley AI lab spends on research in a month. That constraint makes the case for native specialisation stronger.
The Supreme Court's draft AI regulations, released in June 2026, envision a Centre of Research and Excellence on Artificial Intelligence (CoRE-AI) to evaluate tools for court use. That evaluation process will demand exactly what generic AI cannot provide: verified citations, jurisdictional accuracy, and traceable outputs. The advocates who treat generic AI as good enough are making a gamble. Sometimes their verification instincts will catch the errors. The Karnataka HC probe shows what happens when they don't.