Using AI in a VA practice: leverage or negligence

What actually gets you sanctioned

In June 2023, a federal judge in the Southern District of New York fined two attorneys and their firm $5,000. Their brief cited six decisions that did not exist. A general chatbot had produced them, complete with realistic party names, docket numbers, and holdings, and the attorneys filed them without confirming a single one. When opposing counsel and the court could not find the cases, the attorneys did not come clean right away. They submitted excerpts of the fictional opinions instead. The court did not punish them for using software. It punished them for putting their name on a document they had not verified, and then standing behind it.

That distinction is the whole game. The tool did not commit malpractice. The filing did. Used carelessly, an open-ended text generator will hand you a confident answer whether or not it is true, and the responsibility for what reaches the record is always yours. Used with discipline, the same broad category of tool can do a real day of work for you. The difference is not the technology. It is the workflow around it.

A bad citation costs more than a fine

Sanctions make the headlines, but they are the rare outcome. The quieter damage happens long before any judge reaches for Rule 11. In VA practice, a staff attorney at the Board reads your brief line by line. So does the decision review officer at the regional office. The first time one of them follows a citation and finds that the case says something other than what you claimed, or that it does not exist at all, they stop trusting the rest of your argument. Every other cite in the brief now gets the skeptical read. You can lose a winnable claim without anyone filing a motion, simply because the reader decided your work was not reliable.

Credibility in this community is also small and durable. The same names turn up across briefs, hearings, and training rooms for years. A reputation for loose citations follows you, and so does the opposite. That is why the real standard for any research tool is not "does it save time." It is "can I put my name on what it returns."

AI does not get you sanctioned. Filing something you never checked gets you sanctioned.

Leverage looks like a real archive, not an open question

There are two very different things people mean when they say they use AI for research. The first is asking a general chatbot to find supporting cases and trusting the list it writes back. That is the path that produced the fabricated brief. A general model is built to produce fluent text, not to look anything up, so it will compose a citation that looks correct with the same ease whether the case is real or not. There is no lookup happening, only prediction.

The second is searching a fixed collection of actual documents and getting back the ones that are really in it, with the source attached. Picture fishing in a stocked pond instead of the open ocean. The pond holds only real material: CAVC and CAFC decisions, OGC precedent opinions, 38 CFR, BVA decisions. If a case is not in the collection, the honest result is nothing, not an invention. You are not asking the system to remember the law. You are asking it to hand you the document.

That property is what makes the second approach safe to build a practice on. It cannot give you a case that does not exist, because it is returning files, not composing sentences from memory. The job left to you is the one that was always yours: read it, and decide whether it helps.

Where it fits next to what you already pay for

A fair objection is that the firm already pays for Westlaw or Lexis. Those are capable general tools, and they are expensive. They are also not built around VA work. The M21 adjudication manual and the VA Office of General Counsel precedent opinions are exactly the kind of authority that wins veterans cases, and they are slow or awkward to surface in a general database. A VA-specific archive is not a replacement for your judgment or for a citator. It is the fastest way to put the exact veterans-law authority in front of you, in one place, so you spend your time arguing rather than hunting.

How to use it without losing the courtroom standard

The practical payoff is in what a firm does with the time it frees up. A busy attorney can hand a fact pattern to a paralegal, a clerk, or a less senior staffer, have them pull the relevant authority and assemble a first draft of the strategy, then review it, confirm it, and sign it. It mirrors the way a judge works through a staff attorney. The firm moves more cases, and it does so with accuracy rather than at the expense of it. The same setup lets an experienced agent or a VSO build a real argument at the regional office and the Board, not just a sympathetic narrative.

That only holds if the discipline holds. Three habits keep you safe:

Skepticism is the right instinct

Most veterans-law advocates are not afraid of technology. They are afraid of putting their name behind something they cannot check, and that caution is exactly right. The lesson of the sanctioned brief is not that careful people should avoid AI. It is that careful people should refuse any tool that cannot show its sources. The instinct to distrust a confident answer with nothing behind it is the same instinct that makes a good advocate, and it is worth keeping.

The practical move is to aim that skepticism at the tool itself. Ask where each result comes from. Ask what is in the archive and what is not. Ask whether you can open the original opinion and read the passage for yourself. A tool that answers those questions plainly earns a place in your practice. A tool that cannot is the one to walk away from. The profession is going to keep adopting these systems either way, so the useful question is not whether to use AI, but which tools let you verify before you rely on them.

None of this asks you to trust a machine. It asks you to use a tool that shows its work, and then to do the part of the job that has always belonged to you. Attorneys, agents, and VSOs who hold that line move faster than the people still doing everything by hand, and they never have to wonder whether the case they cited is real.