# Responsible AI — How Niyam Handles Accuracy, Privacy, and Limitations > Legal work has consequences for real people. We design Niyam around that responsibility: grounded answers, visible citations, explicit uncertainty, and humans firmly in charge. ## Why responsible AI matters more in law A fabricated citation in a legal submission is not a minor bug — it is a professional failing that can harm a client, mislead a court, and damage an advocate's practice. The stakes are not abstract. That is why we treat accuracy, grounding, and honesty about limitations as foundational constraints, not features. General-purpose AI models were trained to sound helpful. Legal AI must be built to be accurate. These are different objectives, and optimising for one without the other produces a system that is confidently wrong — the worst outcome in legal work. ## Citation grounding and primary sources Every answer Niyam generates that draws on case law or statute identifies the specific source — judgment name, court, year, and the relevant passage. We do not paraphrase authority without attribution. We do not synthesise principles without showing the judgments the synthesis rests on. The corpus Niyam draws from comprises more than 72,000 Supreme Court and High Court judgments indexed from primary sources. Statutory references point to the text of Indian legislation as enacted and amended. Where a judgment interprets a provision, we link the interpretation to both the case and the statute. When a question falls outside what the indexed corpus can reliably answer, Niyam says so. A gap in coverage is disclosed as a gap — not papered over with a plausible-sounding but unsourced response. ## Good-law checking Citing an overruled judgment is as dangerous as fabricating one. Niyam's good-law checking examines whether cases cited in an answer have been subsequently approved, distinguished, criticised, or overruled in later judgments within the corpus. Where a case has been treated adversely, Niyam surfaces that status so the advocate can assess whether it can still be relied upon and in what form. This does not replace a thorough independent verification — good-law checking in any system is only as complete as its corpus — but it adds a meaningful first-pass check that bare retrieval systems do not provide. We are transparent about the limits of our coverage. Good-law status is assessed against our indexed corpus. Judgments not yet in the corpus cannot be cross-referenced. This is disclosed; we expand coverage on an ongoing basis. ## Hallucination mitigation Hallucination — the generation of confident, false statements — is an inherent risk of large language models. We do not claim to have eliminated it. We claim to have built Niyam to reduce it systematically and to make it visible when it might occur. Mitigation approaches include: grounding responses in retrieved passages rather than pure generation; requiring citations to be traceable to the indexed corpus; using retrieval-augmented generation so the model works from actual documents rather than trained weights alone; and calibrating outputs to signal uncertainty rather than confabulate under uncertainty. Where the model cannot find adequate grounding in the corpus for a claim, it is designed to say so rather than invent a source. We invest continuously in evaluating and improving these mechanisms. No system is perfect; ours is designed to fail visibly rather than silently. ## Human-in-the-loop design Niyam is designed for professionals who verify, not for systems that auto-publish. Every output is a starting point for a professional's analysis, not a terminal answer. The advocate reviews the citations. The counsel checks the statutory text. The researcher verifies the reasoning chain. We design the interface to support this: citations are surfaced prominently, not buried; answers are structured to show reasoning so a professional can spot where they disagree; the tone is analytical rather than declarative where nuance is required. We do not offer autonomous filing, auto-generated submissions pushed to courts, or any output path that bypasses professional review. The human is always in the chain. ## Data privacy and confidentiality Legal confidentiality is a professional obligation, not a marketing positioning. We build Niyam on the premise that what a lawyer researches, drafts, and analyses is their client's confidential matter. Your queries, documents, and work product are not sold to third parties. They are not used to train public AI models. They are not shared with other users. We treat your data as you would treat your client's brief — with the confidentiality it demands. Our data handling practices are described in full in our Privacy Policy (niyam.ai/legal/privacy) and Data Processing Agreement (niyam.ai/legal/dpa). ## Limitations we acknowledge Niyam is not a substitute for professional legal judgment. It cannot assess the demeanor of a witness, read the inclination of a Bench, weigh the tactical dynamics of a proceeding, or bring the contextual experience that experienced advocates and counsel carry. These are irreplaceable human capacities. The corpus has gaps. Not every judgment is indexed. Not every jurisdiction is covered with equal depth. Not every amendment to every statute is immediately reflected. We are explicit about these limitations in the product and expand coverage on a rolling basis. AI models can be wrong even when they appear confident. Niyam is designed to surface citations so that wrong answers are verifiable, not trusted blindly. Professional verification remains essential. Niyam is a tool that amplifies professional capacity. It is not a replacement for a lawyer. If anyone suggests you use AI to substitute for legal advice, the advice is to find a lawyer. ## Bias, jurisdiction, and the shape of the corpus A legal AI inherits the shape of the data it is built on. Indian case law is not evenly distributed: some courts publish more, some areas of law are litigated more, and reported judgments skew toward matters that reached appeal. A retrieval system trained on that record can over-surface the well-trodden and under-surface the rare. We treat this as a known limitation to manage, not a problem we pretend away. Our response is to keep the authority visible at every step. Because Niyam shows you the judgments behind an answer rather than asking you to trust a synthesis, you can see when an answer leans on a single line of cases, when a High Court position is unsettled across jurisdictions, or when the on-point authority is thin. The cure for an uneven corpus is not a more confident model — it is a more transparent one. Jurisdiction matters in Indian practice: a High Court judgment binds within its territory and persuades elsewhere. Niyam shows you which court decided what, so you can lead with the authority that binds your forum and keep persuasive authority in reserve, rather than treating every cited case as equal. ## How we test and evaluate Niyam Building responsibly means measuring, not asserting. We evaluate Niyam against legal questions with known, verifiable answers — checking whether the citations it returns are real, whether they are on point, and whether the good-law status it reports matches the subsequent treatment of those cases in the corpus. Citation accuracy is the metric we care about most, because a confident answer built on a case that does not exist is the failure mode with the highest cost in legal work. When evaluation surfaces an error mode, it becomes the next thing we fix — retrieval that missed an on-point case, a synthesis that overreached, a good-law signal that lagged a later overruling. We do not publish a single headline accuracy number, because a number divorced from the question set and the difficulty distribution behind it would mislead more than it informs. What we commit to instead is the practice: continuous evaluation, visible citations so you can audit any answer yourself, and honest disclosure of where the system is weak. ## Ongoing responsibility Responsible AI is not a state we reach and then stop. It is a practice. We evaluate our outputs, track error modes, update the corpus, improve retrieval, and review our hallucination mitigation mechanisms on an ongoing basis. We welcome feedback from legal professionals who find errors in our answers. Identifying where the system fails is how we make it better. If you find a fabricated citation, an incorrectly stated good-law status, or a misleading synthesis, please tell us. ## Frequently asked questions **Q: Does Niyam guarantee accuracy?** No. We invest heavily in accuracy through citation grounding, good-law checking, and hallucination mitigation, but no AI system can guarantee accuracy. Professional verification of all outputs is essential. **Q: What is good-law checking?** Good-law checking examines whether cases cited in an answer have been subsequently approved, distinguished, criticised, or overruled in later judgments within our corpus. It is a first-pass check, not a comprehensive citator service. **Q: Is my data used to train Niyam's AI?** Your queries, documents, and work product are not used to train public AI models. They are kept private and confidential. **Q: What happens when Niyam does not know the answer?** When the corpus does not contain adequate grounding for a reliable answer, Niyam is designed to say so rather than fabricate a response. Visible uncertainty is a feature, not a failure. ## Get started Start your ₹100 trial at https://app.niyam.ai/register — grounded answers, visible citations, honest about limitations. https://niyam.ai/responsible-ai