By now, most practicing attorneys have heard the cautionary tale: a lawyer submits a brief packed with case citations, a judge asks for the actual opinions, and it turns out the cases never existed. The citations were generated by an AI tool and accepted without verification. Sanctions followed. Reputations suffered. Clients were harmed.
These incidents are no longer isolated curiosities. They represent a systemic warning about how general-purpose AI tools — the kind built to generate fluent, confident-sounding text — interact with the high-stakes demands of legal practice. For small personal injury and employment law firms practicing in US courts, the lessons are especially pointed. You operate with lean teams, compressed timelines, and cases where a single procedural misstep can collapse months of work.
This post unpacks what the sanctioned-attorney cases actually teach, why small firms face unique exposure, and how purpose-built document intelligence tools change the risk calculus.
What the Sanctioned Cases Actually Reveal
The high-profile sanctions involving fabricated AI citations share a common anatomy. An attorney or paralegal uses a general-purpose large language model — a tool designed to produce readable, authoritative-sounding text — to assist with legal research or brief drafting. The model generates citations that look real: plausible case names, realistic court designations, believable docket numbers. Nobody verifies them against an actual legal database. The brief gets filed. A judge or opposing counsel checks. The citations don't exist.
What makes these cases instructive is not that the attorneys were reckless or incompetent. Many were experienced practitioners who simply misunderstood what general-purpose AI is actually doing. These models are not retrieving stored facts from a verified database. They are predicting statistically likely text sequences. When asked for a case citation, they produce text that looks like a citation — because that is what citations look like in the training data — regardless of whether the underlying case exists.
The lesson is architectural, not just behavioral. The problem isn't that attorneys used AI. The problem is that they used AI tools whose core design is text generation, not verified document retrieval, for tasks that require verified document retrieval.
The Verification Gap Is the Liability Gap
Every sanctions order in these cases identifies the same failure point: the attorney did not independently verify the AI-generated output before submitting it to the court. That verification gap is the liability gap. It doesn't matter how sophisticated the underlying model is. If the output isn't grounded in actual, retrievable source material, an attorney's signature on the filing rests on facts that cannot be supported — a meaningful professional risk that courts have already begun to address through sanctions and standing orders.
For small firms, the verification gap is harder to close than it sounds. When a two-attorney PI shop is managing forty active matters, the pressure to trust AI output and move on is real. This is precisely why the architecture of the AI tool matters as much as the attorney's diligence.
Why Small PI and Employment Firms Face Distinct Exposure
Large firms have dedicated research attorneys, compliance officers, and multi-layer review processes. A hallucinated citation in a draft brief gets caught before it reaches a partner's desk. Small firms don't have those structural backstops. The paralegal who ran the AI query and the attorney who signs the brief are often the same person, or at most one conversation apart.
Personal injury and employment cases compound this exposure in specific ways.
The Medical Record Volume Problem
Automated medical record review in PI law has become a practical necessity. A moderately complex auto accident case can generate a large volume of medical records — treatment notes, imaging reports, pharmacy records, billing statements — spread across multiple providers and spanning years of care. Manually reading every page to build an accurate medical chronology is the kind of task that invites error under time pressure.
When attorneys or paralegals turn to general-purpose AI to summarize or extract from these records, the hallucination risk surfaces in a different form. Instead of fabricated case citations, you get fabricated medical details: a treatment date that doesn't appear in the records, a diagnosis that was never documented, a gap in care that the AI smoothed over because the surrounding text implied continuity. These errors don't get caught by a judge demanding to see the original opinion. They get caught — or don't — during depositions, by defense experts, or at trial.
The following is a hypothetical scenario for illustration only. Imagine a paralegal uses a general-purpose chatbot to summarize a plaintiff's treatment history across six providers. The chatbot produces a clean, readable chronology. But it omits a three-month gap in treatment because the records from one provider arrived as a scanned PDF with poor OCR quality, and the model simply skipped the ambiguous pages. That gap, unexplained, becomes a damages argument for the defense. The paralegal had no way to know the omission occurred because the output looked complete.
This is why automated medical record review in PI law requires tools that are grounded in the actual documents — tools that can show you exactly which page and line a fact came from, so you can verify it.
The Expert Witness Contradiction Problem
Employment and PI cases often turn on expert testimony. Identifying inconsistencies in an opposing expert's testimony is already labor-intensive: reviewing transcripts and disclosed materials for contradictions, gaps, and admissions. General-purpose AI tools are poorly suited to this task. They cannot reliably retrieve prior deposition transcripts they haven't been given. They may confabulate prior testimony that sounds plausible.
The right approach is narrower and more defensible: use a tool that works only from documents you have actually uploaded to the matter, surfaces contradictions with specific page-and-line citations, and refuses to speculate beyond the source material. That is document-grounded possible-conflict flagging applied to the transcripts and records in your case file — a capability focused entirely on what is in front of you, not on cross-case expert databases or credential verification.
The sanctioned-attorney cases are, at their core, a story about tools that confabulate confidently. The antidote is tools that are transparent about their sources.
The Architectural Difference That Changes Everything
Not all AI tools carry the same hallucination risk. The critical architectural distinction is between generative AI tools and retrieval-grounded AI tools.
Generative tools — general-purpose chatbots and writing assistants — produce output based on patterns in training data. They are not anchored to your documents. When you ask them a question about your case, they answer based on what answers to similar questions have looked like in text they were trained on. This is powerful for drafting, brainstorming, and summarization of content you can verify. It is dangerous for factual retrieval from specific documents.
Retrieval-grounded tools work differently. They ingest your actual case documents, index them, and answer questions by locating relevant passages in those documents and citing them. Every answer comes with a pointer back to the source: this document, this page, this line. You can verify the answer because the tool shows you exactly where it came from.
For small PI and employment firms, this distinction is not a technical nicety. It is the difference between a tool that creates liability and a tool that reduces it.
What Grounded AI Looks Like in Practice
Consider what grounded document intelligence actually does for a PI firm managing a complex soft-tissue injury case. The platform ingests the full case file — medical records, police report, prior deposition transcripts, and documents imported through connected sources such as OneDrive, SharePoint, Google Drive, Dropbox, or Outlook email. When the attorney asks, "What did the treating physician say about the plaintiff's prognosis at the six-month follow-up?" the tool returns the answer with a citation to the specific document and page. The attorney can click through and read the original language. There is no ambiguity about whether the AI invented the answer.
When the same platform generates a medical chronology for a demand package, every entry is tied to a source document. When it flags a contradiction between the plaintiff's deposition testimony and a treatment note, it shows both passages side by side with page references. The attorney isn't trusting the AI's conclusion — they are using the AI to find the passages they then verify themselves.
This is not AI replacing attorney judgment. It is AI doing the document-reading work that no human team can do efficiently at scale, while keeping the attorney in the verification seat.
Building a Verification Culture Before You Need One
The sanctioned-attorney cases are a gift, in an uncomfortable way. They provide a clear map of exactly where the failure points are, before those failure points have affected your practice. Small firms that take these lessons seriously now are building habits and workflows that will serve them as AI becomes more deeply embedded in legal practice — not less.
Practical steps worth implementing immediately:
Distinguish the task before choosing the tool. Drafting a client letter? A general-purpose AI assistant may be appropriate. Extracting facts from medical records for a demand letter, building a deposition outline, or flagging contradictions in testimony? You need a retrieval-grounded tool — one that works from your uploaded documents and cites every answer back to a specific page and line, the way ProvaLens does.
Make source citation non-negotiable. Any AI output that will be relied upon in a filing, a demand package, or a deposition outline should carry a citation to the source document. If the tool can't show you where it found the fact, treat the fact as unverified. This single habit closes the verification gap that has led to sanctions in every high-profile case.
Create a two-step workflow for AI-assisted document review. AI finds; attorney verifies. This applies to medical chronologies, possible-conflict flagging in expert transcripts, and any factual summary that will leave the office. ProvaLens is designed around this workflow: it surfaces the passage and the citation, and the attorney makes the judgment call.
Train your paralegals on the hallucination risk specifically. The risk isn't that AI is wrong sometimes — everyone understands tools make mistakes. The risk is that AI is wrong confidently, in ways that look correct. That specific failure mode requires specific training, and it is the failure mode behind every sanctions case discussed here.
Document your AI workflow. Knowing which tool was used, by whom, and what verification steps were taken creates a record that demonstrates reasonable diligence. ProvaLens maintains audit logging as part of its infrastructure built for HIPAA workflows, which supports this kind of documentation practice.
The firms that will navigate the AI era well are not the ones that avoid AI — they are the ones that use it deliberately, with tools designed for the task and verification built into the workflow. The sanctioned-attorney cases didn't happen because attorneys used AI. They happened because the wrong AI was used for the wrong task, without a verification step. That is a solvable problem, and solving it starts with choosing tools that show their work. To see how ProvaLens fits into that workflow, request a demo at provalens.ai.