ChatGPT will generate a legal citation with exactly the same tone of confidence whether the case is real or entirely invented. That is not a criticism of the tool — it is a large language model predicting plausible text, not a legal research database, and it has no internal mechanism that distinguishes “I found this” from “this is the shape a citation usually takes.” If you use ChatGPT anywhere in a legal drafting workflow, verification is not optional due diligence, it is the entire safeguard standing between you and the public record of sanctioned lawyers that now runs past 1,750 documented cases (Damien Charlotin’s public database, updated continuously).
Step 1: Ask ChatGPT for its own confidence, then ignore the answer
You can ask ChatGPT directly whether a citation it produced is real, and it will often reassure you that it is — sometimes doubling down with invented docket numbers or pin cites when pressed. This is a known failure mode, not a bug you stumbled into. Treat any self-reported confidence from the model itself as having zero evidentiary value. The only thing that counts is an independent lookup against a primary source.
Step 2: Extract every citation into a single list before you do anything else
Copy every case citation, statute, regulation, and quoted passage out of the ChatGPT output into a separate document or spreadsheet before you start checking. Doing this as a distinct first step, rather than checking citations as you encounter them while editing, prevents the common failure where a citation buried mid-paragraph gets skipped because it “read fine” on a normal editing pass.
Step 3: Confirm existence at CourtListener before anything else
For each case citation, go to CourtListener (a free, nonprofit database run by the Free Law Project) and search the exact reporter, volume, and page. If the case does not appear, or the citation resolves to a different case entirely, you have found a fabrication and the citation should be removed or replaced before you evaluate anything else about it. Do not spend time evaluating the legal reasoning of a case that does not exist — existence comes first, always.
Step 4: Check every quotation word for word against the opinion
A citation can pass the existence check and still be dangerous: ChatGPT can attach a fabricated quotation to a real, correctly-cited case. This is arguably the more insidious failure mode because a lawyer who checks existence and stops there will file the brief with confidence. Open the actual opinion text and confirm the quoted language appears exactly as quoted, not paraphrased and not from a different section of the same opinion.
Step 5: Confirm the case actually supports the proposition it is cited for
The final and most subtle check: read the passage you are relying on and confirm the case actually stands for what you are citing it for. ChatGPT can correctly quote a real case that is, for example, a dissenting opinion, a since-overruled holding, or discussing the opposite legal principle in the course of distinguishing it. This step cannot be automated by an existence checker alone — it requires either careful human reading or a tool that specifically evaluates proposition support as a separate layer.
Step 6: Re-run the check if you revise the brief with ChatGPT again
If you ask ChatGPT to revise a section after your initial verification pass, treat any new or modified citations in that revision as fully unverified, even if the surrounding text looks unchanged. A common way fabricated citations slip through review is a late-stage AI-assisted revision that a lawyer assumes only touched prose, not sourcing.
Doing this at scale without burning an evening
Manually running existence, quote, and proposition checks on a 30-citation brief is genuinely tedious, which is exactly the condition under which lawyers skip the step under deadline pressure. Purpose-built verification tools exist specifically to run these three checks in minutes rather than hours — we built Citation Safe for this reason. Whichever tool you use, ask it one question before trusting it: what is your own published false-verify rate, and can you see it, not just take the vendor’s word for it.
Why ChatGPT fabricates citations in the first place
It helps to understand the underlying mechanism, because it explains why no amount of clever prompting fully solves the problem. Large language models like ChatGPT generate text by predicting the statistically most plausible next word, given everything that came before, based on patterns learned from training data. A case citation has a very recognizable shape — party names, a reporter abbreviation, a volume number, a page number, a year — and the model has learned that shape extremely well. When it does not have a specific real case reliably associated with the legal proposition you are asking about, it does not have a built-in mechanism to say “I don't know a real case for this.” Instead, it generates something that fits the shape of a citation, because that is what its training optimized it to do. This is not a glitch that gets fixed by asking more carefully; it is close to the core of how the technology works.
Prompting techniques that reduce, but do not eliminate, the risk
Some prompting approaches modestly reduce fabrication risk without eliminating it: asking ChatGPT to only cite cases it is highly confident about, asking it to explain its reasoning for each citation before providing it, or using a version of the tool with live web browsing or database access rather than one working purely from static training data. Each of these can help marginally, but none of them changes the fundamental fact that the model has no reliable internal signal distinguishing a memorized real citation from a plausible invented one. Treat any of these techniques as a way to modestly improve your starting draft, never as a substitute for the independent verification steps described below.
What a fabricated ChatGPT citation actually looks like in practice
Fabricated citations from ChatGPT are rarely obviously wrong on their face. They typically use realistic party names, a plausible reporter and volume, a year consistent with the surrounding legal context, and sometimes even a fabricated docket number or procedural history that reads as entirely ordinary. In the Mata v. Avianca case, the fabricated citation to Varghese v. China Southern Airlines included not just a case name but purported quoted holdings, which is precisely why a normal editorial read-through did not catch it — it read exactly like a real citation to anyone not specifically checking it against a primary source.
A common question
Does using GPT-4 instead of the free ChatGPT tier make fabrication less likely?
Not meaningfully for this purpose. Independent research (the Stanford RegLab/HAI benchmark study) measured a roughly 43% hallucination rate for GPT-4 on legal queries specifically, which is actually higher than the rates measured for purpose-built legal AI research products in the same study. Model version is a much weaker predictor of citation reliability than whether a separate, deterministic verification step happens before filing.