Does ChatGPT Make Up Case Citations? Why AI Hallucinates Case Law — and How Grounded Research Fixes It

Does ChatGPT Make Up Case Citations? Short Answer: Yes

Does ChatGPT make up case citations? It does — and not occasionally, but structurally. Lawyers across jurisdictions have discovered AI hallucinated legal citations in their filings only after opposing counsel or the bench could not locate the cases at all. The names sounded right. The citation format was flawless. The reporter, volume, and page numbers looked exactly like the citations you read every day. And the case simply did not exist.

This is not a bug that a sharper prompt will cure, nor a temporary limitation that the next model version quietly resolves. It is a direct consequence of how general-purpose large language models are built. For a practising advocate, understanding the mechanism is the difference between using AI safely and filing a fabricated authority in court under your own signature.

Why Large Language Models Invent Plausible-but-Fake Authorities

A general LLM like ChatGPT is, at its core, a next-token prediction engine. It does not consult a database of judgments and copy a citation out of it. Instead, it estimates the most statistically likely next fragment of text given everything written so far. When you ask for authority on, say, the presumption under Section 139 of the Negotiable Instruments Act — which, as a matter of settled principle, is rebuttable and can be displaced by the accused on a preponderance of probabilities — the model produces the words that most plausibly follow: a case name shaped like an Indian citation, a reporter abbreviation that pattern-matches to real ones, a year and page number in the correct range.

The problem is that none of these tokens are anchored to ground truth. The model has learned the shape of a citation without a link to any actual document. Fluency and accuracy are separate axes: the system is optimised to sound authoritative, not to be correct. This is why hallucinations are so dangerous in law specifically — a fabricated citation is not obviously wrong the way a garbled sentence would be. It is confidently, professionally, precisely wrong, and it survives a casual read.

Retrieval-augmented approaches were meant to close this gap, but many consumer tools bolt retrieval on loosely — the model can still 'fill in' a citation it wasn't handed, or blend a real case name with a wrong reporter. Unless grounding is enforced rather than merely encouraged, the underlying generative instinct reasserts itself.

Why This Is Professionally Dangerous for Lawyers

A fabricated authority is not a harmless first draft. Once it enters a written submission, a note to a client, or an oral argument, it carries your professional credibility. Courts have taken an increasingly dim view of filings containing non-existent cases, and the reputational cost of being the advocate who cited a phantom judgment is severe and lasting.

The deeper risk is subtler. Even when a case is real, a general model may misstate what it held, attribute a High Court proposition to the Supreme Court, or miss that the authority was later distinguished or overruled. An answer that is 90% right is arguably more dangerous than one that is obviously wrong, because it invites trust. Verification is not optional diligence layered on top of AI use — for legal work it is the core of the workflow.

How Grounded, Retrieval-First Research Prevents Fabrication

The fix is architectural, not cosmetic. A grounded research system inverts the default: instead of letting the model generate freely and hoping the citations are real, it retrieves actual authorities first and constrains the answer to what those retrieved documents support. The model's job becomes synthesis and explanation over a known corpus — not conjuring facts from statistical memory.

The decisive step is verification in code. Before a citation is shown to the user, the system checks that the cited authority actually exists in the retrieved set and that the proposition maps to real source text. A citation that cannot be verified is not rendered as a citation at all. This turns 'no hallucinated case law' from a marketing promise into an enforced property of the pipeline.

Grounding also makes treatment legible. Knowing a case exists is only half the question; you need to know whether later courts followed, distinguished, criticized, or overruled it — or treated it neutrally. A citator that maps the litigation journey and treatment signals — with a dual-gated flag before anything is marked 'overruled' in red — protects you from relying on authority that is real but no longer good law.

How Clauselex Enforces Grounding

Clauselex is built for Indian advocates on exactly this principle. In its grounded research mode — the default — it answers only from retrieved authorities across the Supreme Court, High Courts, and tribunals, and every citation is verified in code before it renders, so a plausible-but-fake case never reaches your draft. The built-in citator then surfaces appeal chains and treatment — followed, distinguished, criticized, overruled, or neutral — so you can see at a glance how a judgment has been treated before you rely on it.

When you move from research to drafting, the same discipline carries through: the Draft Studio lets you build a point-by-point reply grounded in retrieved authority, in your own style, with track-changes you accept or reject and clean DOCX and print-PDF export. The generative power stays; the fabrication does not.

A closing note: this article is educational and is not legal advice, and no automated tool removes your duty to verify before you file. Grounded research narrows the gap between what AI says and what the record actually holds — but the signature, and the responsibility, remain yours.