AI for Legal Research, Without the Fabricated Citations

TL;DR: Every profession has a cautionary AI story; legal has case law. Lawyers have been sanctioned for filing briefs built on authorities that a chatbot invented, complete with plausible party names, reporter citations, and quotations. That failure mode is inherent to how these models work, and it does not go away with better prompting or legal-specific tools; it only gets less frequent. This guide covers what AI is genuinely good for in research (orientation, search strategy, summarizing verified sources), the workflows that exploit that safely, and the verification regime that must sit between AI output and anything a lawyer relies on. This is not legal advice. No AI research output goes into a memo, filing, or decision without a qualified lawyer verifying every authority against the primary source and signing off.

This guide is part of the AI for Legal hub. Tracking regulatory change, research’s ongoing cousin, is covered in compliance monitoring.

The case law on AI making up case law

Start with what actually happened, because it calibrates everything else.

• Mata v. Avianca, Inc. (S.D.N.Y. 2023). The canonical incident. Counsel filed an opposition brief citing six cases that did not exist, ChatGPT had generated them, and when asked to verify, generated fake excerpts too. The court sanctioned the attorneys and their firm $5,000 and required letters to the judges falsely named as authors of the fabricated opinions. The citations had realistic party names, reporter numbers, and internal quotes.
• Park v. Kim (2d Cir. 2024). The Second Circuit referred an attorney to its grievance panel after a brief cited a nonexistent case produced by ChatGPT, stating flatly that attorneys must ensure their filings’ citations exist.
• Wadsworth v. Walmart (D. Wyo. 2025). Lawyers at one of the largest US plaintiff firms were sanctioned after motions cited AI-fabricated cases, demonstrating this is not a solo-practitioner problem.
• England and Wales, 2025. In a Divisional Court judgment addressing two cases involving fake citations (Ayinde and a linked matter), the court warned that submitting fabricated authorities could attract contempt proceedings and, in the gravest cases, criminal referral.

Trackers maintained by legal researchers have catalogued hundreds of such incidents worldwide. Three lessons: fabricated authority looks exactly like real authority; asking the model to confirm its own citations produces confident false confirmation; and courts treat “the AI did it” as an aggravating explanation, not a defense.

Why this happens, and why tools reduce but don’t remove it

A large language model generates the most plausible next text given its training. Legal citations have a strong, learnable format, so when a model needs an authority to support a proposition, producing a plausible-looking one is exactly what it is built to do. That is hallucination: not a bug to be patched, but the generation mechanism operating without a ground-truth check.

Legal research platforms address this with retrieval-augmented generation, the model answers from documents retrieved out of a real case-law database rather than from its parametric memory. This genuinely helps: the citation usually exists. But published academic testing of commercial legal AI research tools has found they still produce incorrect or miscited answers at meaningful rates, the retrieved case may be real but not say what the summary claims, may be inapposite, or may have been reversed. Grounding changes the failure from “fake case” to “real case, wrong proposition.” Both end careers if filed. The verification duty is identical for every tool.

What AI is actually good for in research

Used inside its competence, AI compresses real hours:

• Orientation in unfamiliar territory. “What are the main legal issues when a SaaS company stores health-adjacent data for EU users? Frame the questions I should be researching, do not cite authorities.” Ten minutes to a decent issue map that would have taken an afternoon.
• Search strategy. Terms of art, alternative framings, the statutory vocabulary that unlocks the real databases. The research still happens in Westlaw/Lexis/vLex/official sources; AI just gets you to better queries faster.
• Summarizing sources you provide. Paste the full text of a case you pulled from a verified database and ask for the holding, the reasoning, and how it treats a specific issue. The model reads what’s in front of it, its strongest mode. Verify against the text you gave it, which you have.
• Adversarial pressure-testing. “Here is our draft argument. Attack it: what counterarguments and distinguishing facts would opposing counsel raise?” No citations needed; pure reasoning against provided material.
• Synthesis across verified sources. Given five cases you verified yourself, draft a comparison of how each treats the question. Every input already checked; the output checked against them.

The pattern: AI works safely on material you supply and verify. It becomes dangerous the moment it supplies the authorities.

A research workflow that survives contact with a judge

1. Frame with AI, cite from databases. Use the assistant for the issue map and the search vocabulary. Then run the actual searches in your legal research platform or official sources (court websites, legislation portals).
2. If AI proposes any authority, treat it as a lead, not a source. Pull it from the database by citation. If it doesn’t come up, it does not exist, do not ask the model to confirm it (see Mata: it will).
3. Read the primary source. Not the AI summary of it, the authority itself, at least the portions you rely on. Confirm the holding, check it says what is claimed, and run it through your citator (KeyCite/Shepard’s or local equivalent) for negative treatment.
4. Draft with AI from verified material only. Provide the confirmed authorities and your analysis; let the model assemble the memo structure and first-pass prose. Instruct it: “Use only the authorities provided. Do not add citations.”
5. Verify the draft’s citation layer line-by-line. Every citation, every quotation, every pin cite, against the source. Quotes are a known weak point, models paraphrase inside quotation marks.
6. Named lawyer signs. The memo is the lawyer’s work product. If it is filed, court rules (and possibly a standing order on AI disclosure) apply to that lawyer personally.

Confidentiality note

Research prompts leak facts. “Can our client terminate the [named counterparty] supply agreement for the contamination incident?” puts privileged strategy and client confidences into a third-party system. Use the sanctioned enterprise channel from your acceptable use policy, and abstract the facts where you can, the law of anticipatory repudiation does not depend on the client’s name.

The verification checklist

This looks heavy. In practice it is the same diligence competent research always required, AI changes where the hours go (less finding, more verifying), not the standard of care.

Failure modes to design against

• The confirmation trap. Asking the model whether its citations are real. It will say yes, with detail. Verification happens outside the model, always.
• Summary drift. Relying on the AI’s description of a case you never opened. The summary is a map, not the territory.
• Currency blindness. Training cutoffs mean the model may not know last year’s amendment or last month’s reversal. Anything time-sensitive gets a citator and a current-source check, and ongoing obligations belong in a compliance monitoring process, not ad-hoc research.
• Calibrated overtrust. The first fifty answers were fine, so the fifty-first goes unchecked. Every sanctioned lawyer in the cases above trusted output that looked exactly like the output that had been fine. Keep verification structural, not mood-based.
• Junior-lawyer bypass. Research is how juniors learn the law. If AI does all first-pass research unreviewed, you get both unverified output and untrained lawyers. Pair them: junior verifies AI output against sources, faster than raw research, and the reading still happens.

FAQ

Did AI really invent case citations that lawyers filed in court? Yes, Mata v. Avianca (S.D.N.Y. 2023, sanctions), Park v. Kim (2d Cir. 2024, grievance referral), Wadsworth v. Walmart (D. Wyo. 2025, sanctions), and a 2025 England and Wales Divisional Court judgment warning of contempt exposure, among many catalogued incidents.

Are legal-specific AI research tools immune to hallucination? No. Database-grounded tools fabricate less but still miscite and mischaracterize at meaningful rates. The verification duty is the same for every tool.

So what is AI actually good for in legal research? Issue mapping, search strategy, summarizing and synthesizing sources you provide and verify, and pressure-testing arguments. It supplies speed; your databases supply authority.

Do we have to disclose AI use in court filings? Possibly, judges’ standing orders and bar guidance increasingly address it. Check every court you appear in; treat it as a compliance question for your own counsel.

Next in this cluster: turn one-off research into an ongoing watch with AI for compliance monitoring, or return to the AI for Legal hub.

Not sure which legal workflow to start with? Take the free AI readiness assessment, ten minutes, and you’ll get a prioritized starting point for your team.