AI Citation Hallucinations in Legal Filings

A public tracking database now records more than 1,300 court filings that contain fabricated AI-generated citations. That number will grow. This single fact should anchor everything an attorney does with AI tools.

The problem is not that AI is unreliable in general. The problem is specific: generative language models like Claude are not citation sources, and attorneys who treat them as such face sanctions, bar discipline, and malpractice exposure. The attorneys at Schwartz & Levidow learned this in 2023 when they submitted a brief in Mata v. Avianca, No. 22-cv-1461 (S.D.N.Y. 2023), containing multiple cases that did not exist. Judge P. Kevin Castel imposed $5,000 in sanctions and required the attorneys to send copies of the order to every judge whose name had been attached to a fabricated opinion. The case became the defining event in AI legal risk and has been followed by dozens of similar incidents in 2024 and 2025 across federal and state courts.

Every attorney using AI tools today needs to understand why this happens and how to prevent it. The exposure is too direct and too well-documented to treat as someone else's problem.

Why Generative AI Produces Fabricated Citations

Claude and similar language models are trained on enormous volumes of text, including legal briefs, academic articles, and published opinions. During training, the model learns the patterns of legal writing: how citations look, how they are structured, what kinds of cases tend to appear in briefs about specific subjects. When you ask the model to draft a brief on, say, tortious interference, it knows that such briefs typically cite cases. It generates text that follows those patterns.

The critical distinction is this: the model is completing a pattern, not retrieving a record. It has no connection to Westlaw or Lexis or any official reporter. The base version of Claude does not access the internet during a response. When it produces "Smith v. Jones, 847 F.3d 112 (2d Cir. 2017)," it has generated that text the same way it generates any other sentence, by producing what statistically fits the context. Whether that case actually exists is a separate question the model cannot answer for itself.

This is not a defect that will be patched out. It reflects the fundamental architecture of generative models. They are trained to produce fluent, contextually appropriate text. A fabricated citation is fluent and contextually appropriate. The model has no internal signal that distinguishes a real case from a plausible-sounding one it has constructed.

Some attorneys have tried asking Claude to verify its own citations. This does not work. When you ask the model whether a citation is accurate, it either affirms it (because the same pattern-completion process governs the response) or expresses uncertainty. Neither response constitutes verification. You are asking the same system that generated the citation to evaluate itself, without access to any external record.

The Exposure: Rule 11, Bar Discipline, and Malpractice

Federal Rule of Civil Procedure 11 requires that attorneys certify, by signing a filing, that the legal contentions in that filing are warranted by existing law. Submitting a case that does not exist is a direct violation of that certification. Courts have made clear, beginning with Mata and reinforced in subsequent decisions, that "the AI wrote it" is not a defense. The obligation belongs to the attorney of record.

Beyond Rule 11, state bar rules in virtually every jurisdiction require candor toward the tribunal. Citing nonexistent authority is a candor violation. Several bar disciplinary bodies have opened proceedings in AI citation cases, and the professional responsibility exposure mirrors the sanctions risk: the attorney signed the document, the attorney is responsible for its contents.

Malpractice exposure follows the same logic. If a client suffers harm because a brief contained fabricated authority, the attorney's failure to verify those citations is negligence. No engagement agreement will shift that liability to an AI vendor.

Five Rules That Prevent the Problem

None of this means attorneys should avoid AI tools. It means they should use them correctly. These five rules, applied consistently, eliminate AI citation risk entirely.

Rule 1: The attorney supplies all citations. The model never originates them. Use Claude to draft structure, arguments, and analysis. Every case citation in that draft must come from you, inserted from a primary source you have already verified. Do not ask the model to find cases. Ask it to work with cases you provide.

Rule 2: Every cited case is verified against a primary source before the document leaves your desk. Westlaw, Lexis, the official reporter, or a court's electronic filing system are primary sources. Another attorney's brief is not. A summary from any AI tool is not. You check the case yourself: it exists, it says what you say it says, it has not been overruled.

Rule 3: Demand letters get a pre-send citation audit. Demand letters carry significant professional weight. They can form the basis of a legal position, trigger settlement negotiations, or end up as evidence. If you used AI drafting for a demand letter, every citation in that letter requires the same primary-source check as a court filing. The verification process for demand letters is covered in detail in our earlier post on writing demand letters with Claude AI, including the specific pre-send checklist we recommend. For the negotiation phase that follows, the Claude AI settlement negotiation protocol applies the same fact-supply discipline to counter-offer drafting.

Rule 4: Filings get a separate human reviewer before submission. A second set of eyes on citations, specifically tasked with citation verification rather than general proofreading, catches errors that the drafting attorney misses. This is true of all legal work and is especially true when AI tools were involved in drafting.

Rule 5: Never describe primary-source verification as "AI confirmed." This phrase means nothing and, if it appears in a firm's work records, could be used against you to show that verification was delegated to a system incapable of performing it. Your records should show that a named attorney checked a specific citation against a specific primary source on a specific date. That is verification. AI output is not.

How This Works in Practice With Claude

The attorneys who use AI tools effectively in their practices have internalized a simple division of labor: Claude drafts, attorneys source. The model is well suited to organizing arguments, translating complex legal standards into accessible prose, identifying gaps in a draft's logic, and generating the structure of documents like motions or briefs. It is not suited to finding cases, and it should not be asked to.

A practical workflow looks like this: you conduct your legal research through your normal research tools and compile the cases you intend to cite. You then bring those cases, with their holdings stated accurately, into your Claude session and ask the model to help you draft around them. The citations in the final document trace back to your research, not to the model's output.

Attorneys who complete The Leveraged Attorney work through this division of labor in detail, including session structures, prompt patterns, and review protocols designed specifically for legal practice. The goal is not to add a verification step as an afterthought; it is to build a workflow where AI-originated citations never appear in the first place, because the workflow does not permit the model to originate them.

Generative AI for Drafting, Retrieval AI for Research

The legal technology market is moving toward a clearer product distinction that every attorney should understand. Generative AI tools, Claude among them, are trained to produce fluent text and reason through arguments. They are optimized for drafting, summarization, and explanation. They are not legal research databases.

A separate class of tools connects a language model to a live legal database and grounds its outputs in actual retrieved documents. When such a system produces a citation, it points to the source it retrieved. Several legal research vendors have built products in this category. These tools carry a different risk profile from base generative models, though they are not immune from error either.

The distinction matters for how you evaluate any AI tool in your practice. Ask the vendor, specifically: is this a generative model responding from training data, or does it retrieve from a live legal database and cite the retrieved source? If the answer is the former, apply the five rules above without exception. If the answer is the latter, still verify, but understand that the underlying mechanism is different.

Claude is a generative model. It is genuinely useful for the drafting work attorneys do every day. It can help a solo practitioner produce well-organized briefs, help a managing partner prepare presentations, and help any attorney think through the structure of an argument more quickly than working alone. That value is real. So is the citation risk, and the two facts coexist without contradiction. Use the right tool for each part of the job.