The Short Answer
On September 3, 2026, the D.C. Court of Appeals struck the appellee's brief in Douglas v. Deutsche Bank National Trust Co. after four cited authorities turned out not to exist. The court did not throw out the case; it struck the filing and referred the matter to disciplinary counsel. Legal AI has to retrieve from verified sources with provenance on every citation, not generate them. With ibl.ai you own all the code and the data.
The distinction that matters is not which model wrote the brief. It is whether a citation in that brief is a document the system retrieved or a string the model produced.
What did the D.C. Court of Appeals actually do to the Deutsche Bank brief?
It struck the brief in its entirety and referred the matter to the District of Columbia Bar's Office of Disciplinary Counsel.
The order in Douglas v. Deutsche Bank National Trust Co., No. 24-CV-1099, was issued on September 3, 2026. Four of the authorities cited in the appellee's brief did not exist.
Counsel acknowledged using Google's generative artificial intelligence search tool to locate case authority, and not verifying that the results were real.
This is worth stating precisely, because the story is circulating as "a Deutsche Bank filing was thrown out." The case was not thrown out.
One filing was struck, which is a different and more instructive outcome: the bank's side of the appeal lost its written advocacy while the appeal itself continued.
The panel also declined to treat this as one attorney's mistake.
As reported by the ABA Journal, it found that every firm attorney who signed the brief bears some responsibility, and described a brief with fake citations as something that "misdirects the court, burns time and resources, and sows confusion."
That is the operative rule. Signing is certifying. No tool moves that obligation.
How common are AI-hallucinated citations in court filings?
Common enough that it is now a tracked category of litigation misconduct rather than a novelty.
Norton Rose Fulbright, surveying generative-AI sanctions decisions, counted more than 1,148 US cases of hallucinations by lawyers as of June 2026, drawing on the public database of AI hallucination decisions maintained by researcher Damien Charlotin.
The penalties are no longer symbolic.
In Whiting v. City of Athens, the Sixth Circuit imposed $15,000 in punitive sanctions on each of two attorneys, plus fees, double costs and a disciplinary referral — the court did not find the fabrications were AI-generated, only that counsel had not read what they cited.
Two years ago this was one famous case. It is now a pattern with a docket count, and appellate panels are actively checking citations that look unfamiliar.
Does retrieval-augmented legal AI actually stop hallucinated citations?
Not on its own, and this is the part most coverage of these incidents gets wrong.
The reflex conclusion is that the lawyer used the wrong tool — a general chatbot rather than a legal one.
But Stanford's RegLab ran the first preregistered evaluation of commercial legal research tools and found that tools built on retrieval over proprietary case law still hallucinated between 17% and 33% of the time, against vendor marketing that promised "hallucination-free" citations.
The paper was published in May 2024 by Magesh, Surani, Dahl, Suzgun, Manning and Ho, and its conclusion was blunt: providers' claims are overstated.
So "we bought legal AI" is not an answer to this failure mode. Retrieval narrows the gap. It does not close it, because a system can retrieve a real case and still mis-describe what it held, or assemble a citation whose reporter volume and page never went together.
What closes it is a verification step that runs after generation and before filing.
What is the difference between a model generating a citation and a system retrieving one?
A generated citation is a prediction about what a citation usually looks like. A retrieved citation is a pointer to a specific document.
A language model producing "Smith v. Jones, 412 F.3d 118 (4th Cir. 2005)" has produced a well-formed string. Nothing in that process consulted a corpus. The format is right because the format is learnable; the case may never have existed.
A retrieval system does something structurally different. It queries a corpus, gets back a document with an identifier, a court, a date and a text, and carries those attributes forward. The citation is a reference to a row that exists, and the system can show you the row.
That difference produces four properties a filing workflow can actually check:
- An identifier that resolves. Every cited authority maps to a document in a corpus, and the link opens it.
- Provenance on the face of the draft. Which database, which version, retrieved on which date — visible to the attorney reviewing, not buried in a log.
- Quotation bound to source text. A quoted passage is a span from the retrieved document, not a paraphrase the model reconstructed.
- A failing verification gate. A resolver checks every citation before the draft can be filed, and a citation that does not resolve blocks the draft instead of warning about it.
The first three are architecture. The fourth is process, and it is the one that would have caught this brief.
This is the same failure shape described in worse than hallucination: confidently wrong — the system's confidence is uncorrelated with whether the underlying fact is real, so confidence cannot be the check.
What should a firm require before an AI-assisted filing goes out?
Five things, none of which is a model choice.
A citation resolver in the pipeline. Every authority in a draft is checked against a corpus, automatically, as a blocking step. Unresolved citation, no filing.
Provenance rendered for the reviewer. The attorney who signs sees where each authority came from and when, and can open it in one click.
A record of what the system was asked and what it returned. When a court asks how a citation got into a brief, the firm should be able to answer from logs rather than recollection.
Confidentiality that survives the tool. Client material in a draft should not leave the firm's perimeter to be processed, which is a deployment question before it is a vendor question. That argument is developed in legal AI's next crisis is trust, not intelligence.
A named human check. Rule 11 in federal district practice puts responsibility on the signer — and where no such rule exists, as here, the sanction lands on the client's brief instead.
The workflow should make verification the default path, not a discipline the busiest associate is expected to remember at 11 p.m.
How does ibl.ai deploy AI for law firms and legal departments?
With ibl.ai you own all the code and the data.
The platform runs on the firm's own infrastructure with full source code access, so privileged material and work product stay inside the perimeter.
It is model-agnostic across any LLM, so a firm can change models without rebuilding its citation and retrieval layer. It is usage-based with no per-seat pricing, so cost tracks matter volume rather than headcount.
And you can deploy anywhere, from your own cloud to on-premise, GovCloud or a fully air-gapped network.
For this failure mode specifically, ownership is what makes verification auditable.
Agents retrieve from the sources the firm designates, every retrieved authority carries its identifier and retrieval date through to the draft, and the verification step is code the firm can read and test rather than a vendor assurance about a black box.
Per-seat licensing also has the wrong shape here. A firm's AI spend should follow the volume of work the system does, not the number of people who might one day open it.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY.
Related reading: legal AI's next crisis is trust, not intelligence — why identity and provenance, not model quality, decide whether legal AI is usable; and worse than hallucination: confidently wrong — the same confidence-without-grounding failure outside the courtroom. See how ibl.ai deploys for firms at ibl.ai/solutions/legal.
Sources: the order in Douglas v. Deutsche Bank National Trust Co., No. 24-CV-1099 (D.C. Ct. App. Sept. 3, 2026), is published by the D.C. Courts and was reported by the ABA Journal and Eugene Volokh; the case count and the Whiting sanctions from Norton Rose Fulbright's 2026 generative-AI sanctions update; the 17–33% hallucination rates from Stanford RegLab and the May 2024 paper.