How to verify AI legal research citations and avoid court sanctions for fabricated cases?

The Current State of AI Citation Failures in Courts

Attorneys who use generative AI for legal research without independently verifying every citation are accepting a level of professional risk that has become statistically indefensible. The first widely reported case, Mata v. Avianca in June 2023, produced $5,000 in sanctions and triggered hundreds of follow-on motions across federal and state courts. Five quarters later, courts have imposed well over $145,000 in cumulative penalties during the first quarter of 2026 alone, according to JD Supra's running tally of GenAI filing failures. In March 2026, a Texas attorney named Leslie admitted in a sworn affidavit that he failed to verify AI-generated "expanded legal research" used in a court order, adding another high-profile entry to the discipline docket. The pattern is no longer novel; it is structural.

Also worth reading: How Do Lawyers Verify AI-Generated Citations Before Filing in 2026? · How to Draft Legal Documents with AI in 2026 Without Inviting Sanctions? · How Is AI Transforming eDiscovery, Legal Research, and Document Drafting?

The mechanism behind these failures is well documented. Generative models can produce text that resembles legitimate legal citation form while pointing to cases that do not exist, were decided differently, or have been overruled, distinguished, or otherwise altered. This is the hallucination problem, and it is not a transient bug to be patched away. Multiple commentators, including the National Law Review's 2026 predictions panel, expect the volume of AI-assisted filings to grow faster than the rate at which law schools and CLE providers teach verification skills, which is the structural mismatch that explains why sanctions keep accumulating.

The Legal Duty of Independent Verification

The duty to verify citations is not new, and courts have been clear that AI does not change the standard. Rule 11 of the Federal Rules of Civil Procedure requires attorneys to conduct a reasonable inquiry into the factual and legal support for any filing. Filing a brief that contains nonexistent citations, even if generated by a tool the attorney did not understand, does not satisfy that duty. The Mata court framed the issue as one of candor to the tribunal, not as a technology dispute, and subsequent courts have followed that framing. Filing v. Northbrook and the Colorado divorce matter of 2024 both produced monetary penalties for attorneys who filed AI-drafted text without reading it.

The duty also runs to clients. ABA Model Rule 1.1 requires competent representation, and several ethics opinions in 2024 and 2025, including Formal Opinion 512 from the ABA, treat failure to supervise AI output as a competence failure rather than a clerical oversight. The California State Bar's update to its practical guidance on technology competence follows the same logic: lawyers must understand the benefits and risks of the tools they use, and they cannot delegate that understanding. A paralegal or junior associate cannot shield the attorney of record from responsibility, as a 2025 California case made explicit when it sanctioned a supervising lawyer who delegated verification to non-attorney staff. Reuters covered that decision in detail, and it has become a staple citation in CLE programming.

How AI Hallucinates Legal Citations

The technical term for fabricated citations is hallucination, but the underlying process is worth understanding because it shapes the verification strategy. Large language models generate text one token at a time by predicting the most likely next word given the prompt and the preceding context. They do not look up cases in Westlaw or Lexis; they have been trained on a corpus that includes many case names, citation strings, and judicial opinions, and they recombine those patterns. A model asked to write a brief on Section 230 immunity will confidently produce a string that looks like "Smith v. Jones, 597 F.3d 123 (9th Cir. 2020)" because that pattern fits the context, even when no such case exists. A 2024 study by Stanford's CodeX group found that GPT-4 class models fabricated plausible citations at rates between 17% and 33% depending on the practice area, with tax and immigration law showing the highest error rates.

This is also why a partial match is not a safe match. A model might correctly cite the party names of a real case but get the year, court, or holding wrong. It might cite a real district court opinion as if it were a circuit decision. It might conflate two cases with similar names. The surface plausibility of the output is precisely what makes it dangerous, because the busy reader skims rather than reads. Verification has to be substantive, not gestural.

A Practical Eight-Step Verification Protocol

Thomson Reuters and Harvey have both published verification protocols in 2025, and the steps converge into a workable routine. First, never paste an AI block of citations into a brief without re-reading the underlying sentences, because the citation is only as reliable as the proposition attached to it. Second, for every citation, retrieve the case from a primary source such as Westlaw, Lexis, CourtListener, or the court's own PACER docket. Third, confirm that the case exists, that the citation string matches the reporter volume and page, and that the court and date are correct. Fourth, read the cited portion of the opinion and confirm that the language in your brief actually appears there, in context. Fifth, run a shepardization or KeyCite check to ensure the case has not been overruled, distinguished, or questioned. Sixth, verify that the case is good law in the relevant jurisdiction, because a still-valid Ninth Circuit case may be the wrong fit for a Sixth Circuit matter.

Seventh, confirm that the case is actually on point rather than merely topically adjacent, since generative models often pull cases that mention a keyword without applying the rule. Eighth, document the verification step. Some firms now require a verification log attached to the draft, listing each AI-generated citation, the database used, the date of confirmation, and the name of the reviewer. This log does not insulate the firm from error, but it converts a potential negligence claim into a record of due diligence and tends to satisfy the Rule 11 inquiry that any reasonable attorney would have conducted. The cost in time is real; a junior associate can verify roughly 30 to 50 citations per hour using Westlaw, which means a heavily cited motion can add two to four hours of work.

Comparison of Verification Approaches and Tools

Not all AI legal research tools handle verification equally, and the choice of tool affects how much manual work remains. The table below summarizes the main categories as of mid-2026.

The practical takeaway is that grounding alone does not eliminate the verification obligation. CoCounsel is built on Westlaw and Practical Law, but it can still produce a misstatement of a holding if the prompt is ambiguous, and the 2025 ABA opinion makes clear that "the tool gave me a real citation" is not a defense. Descrybe, which launched its verified legal research layer for ChatGPT in late 2025, attempts to surface only citations it has confirmed against external databases, which is a useful design but does not cover every jurisdiction equally well. Filevine's pitch is passage-level rather than case-level verification, which addresses the most common failure mode of generative tools but still leaves the proposition-sentence for the attorney to confirm.

Common Mistakes That Lead to Sanctions

The sanction cases from 2024 through 2026 cluster around a small number of recurring errors, and naming them is the fastest way to avoid repeating them. The first is treating AI output as a first draft rather than as raw material. Attorneys who ask a model to write a brief and then submit it after only light editing have produced almost every reported sanction in the dataset. The second is failing to read the cited authority, which sounds obvious but is the single most common omission in the published opinions. The third is overlooking jurisdictional mismatch, including citing a state case in federal court without checking the forum's choice-of-law rules. The fourth is relying on the AI's own self-assessment when asked "are these citations real?" because the same model that fabricated the citation will often affirm its own accuracy with the same statistical confidence.

The fifth is failing to disclose AI use where local rules require it. The Western District of Texas and the District of New Jersey have standing orders requiring disclosure of AI-generated content in filings, and at least one 2025 sanction included the failure-to-disclose count as a separate violation. The sixth is delegation. The California case referenced above sanctioned the supervising attorney for assigning citation verification to a paralegal without attorney review, and the same logic appears in the Colorado disciplinary advisory opinion from December 2024. The seventh is the shepardization gap, where an attorney confirms a case exists but does not check whether it has been overruled, which is the equivalent of citing Plessy in a 2026 brief.

When to Adopt, When to Wait, and When to Avoid

Adoption is sensible for tasks where the verification cost is low relative to the time saved. Document drafting for first-draft contracts, internal memoranda that will be heavily revised, and discovery-issue spotting are reasonable use cases because the output is rarely filed with a court in unedited form. AI is also useful for summarization of long documents, where the attorney still reads the underlying source but gets a faster orientation. The line moves when the output is destined for a court filing, a client opinion letter, or any document on which the firm would be willing to have its name and bar number attached without further review. At that point, the verification protocol is not optional, and the time savings shrink toward zero.

For litigation-support workflows in eDiscovery, AI has earned a clearer pass. Predictive coding has been judicially approved since Da Silva Moore in 2012 and is now standard for document review, with TAR (technology-assisted review) protocols routinely approved in multi-billion-dollar matters. The risk profile there is different because review is supervised by a human reviewer who codes the training set, and the output is a relevance ranking rather than a citation string. Confusing the two is the most common conceptual error made by firms new to AI; legal research tools and eDiscovery tools solve different problems and have different risk envelopes.

Costs, Pricing, and the Bottom Line

Budget matters because verification is not free. A mid-sized litigation firm handling 50 motions per year with an average of 25 citations per motion can expect to spend roughly 1,000 to 1,500 attorney-hours on verification if every AI-assisted citation is treated with the eight-step protocol. At a blended rate of $250 per hour, that is $250,000 to $375,000 in labor annually, which is the same order of magnitude as the subscription cost of CoCounsel for a 50-attorney firm. The economic argument for AI legal research is therefore strongest at scale, where the time savings on first drafts exceed the verification cost. Solo practitioners face the hardest math, because they pay full subscription cost and bear the same verification burden, which is why tools like Descrybe, ChatGPT Plus, and the free tier of CourtListener are most attractive at that level even though they offer narrower coverage.

The bottom line is that verifying AI legal research citations is a competence requirement, not an optional best practice. The 2026 sanction wave shows that courts have lost patience with the "I did not know" defense, and the bar associations have answered with formal ethics opinions and updated technology-competence rules. Firms that build the eight-step protocol into their standard workflow, document the verification, and use grounded tools will absorb the cost and continue to benefit from the productivity gains. Firms that treat AI as a substitute for attorney reading will continue to populate the sanction docket. The technology is not the problem; the workflow is.