Every attorney I know has now had the same conversation twice. Once with a colleague who swears artificial intelligence has doubled their output, and once with a colleague who read about the lawyers sanctioned for filing briefs full of cases that never existed. Both conversations are about the same technology, and both people are describing something real.
I have spent the last two years using these tools in an active real estate litigation practice — partition actions, contract disputes, landlord-tenant matters, title defects. Not experimenting with them on the side. Using them on real files with real deadlines and real clients. What follows is an honest account of where they genuinely help, where they still fail in ways that can end a career, and what I have learned to require before I trust anything one of them produces.
What These Tools Are Genuinely Good At
Let me start with the part that is not hype, because there is a real productivity gain here and attorneys who dismiss the entire category are giving up something valuable.
Getting oriented in an unfamiliar issue. Real estate litigation throws odd fact patterns constantly. An easement created by implication in a subdivision that was never properly platted. A lien priority question complicated by a construction loan modification nobody recorded. Ten years ago, the first afternoon on a question like that was spent simply learning the vocabulary — finding out what the doctrine is called so I could search for it properly. A good research tool now collapses that first afternoon into about fifteen minutes. It tells me what the issue is called, which statutes govern, and which cases keep appearing. That is not the answer. It is the map that lets me go find the answer efficiently.
Working through long records. Give a competent tool a two-hundred-page deposition transcript and ask what the witness said about the roof repairs, and it will find every relevant passage in seconds. This is genuinely useful, and it is safely verifiable — you check the citation against the transcript page and either the testimony is there or it is not. Same with document productions, closing files, and years of correspondence. The machine is very good at finding the needle when you can confirm the needle yourself.
First drafts of structured documents. A motion skeleton. A demand letter framework. A chronology assembled from the documents. The first draft is never the filed draft, and anyone who treats it as such will embarrass themselves. But starting from a structured draft beats starting from a blank page, particularly on the fifth similar motion in a month.
Explaining a position in plain language. Clients deserve to understand what is happening in their case. These tools are unusually good at translating a legal posture into language a non-lawyer can follow, which I then edit for accuracy. It has made my client communication better, not worse.
Where They Still Fail — And Why It Matters
Now the part that should keep you cautious.
Fabricated citations. This is the failure everyone has heard about, and it deserves its reputation. General-purpose chatbots will produce case names, reporter citations, pin cites, and direct quotations that look completely authentic and do not exist anywhere. Not misremembered — invented. Courts across the country have sanctioned attorneys for filing them. Several have referred lawyers to their state bars. This is not a hypothetical risk or an early-days problem that has been solved. It remains a live danger with general-purpose tools.
Cases that were good law and no longer are. This failure gets far less attention than fabricated citations, and in my judgment it is more dangerous. A tool surfaces a case. The case is real. The citation is correct. The reasoning is squarely on point for your issue. And it was overruled six years ago by a decision the tool never mentions.
You will not catch this by verifying the case exists, because it does exist. You will not catch it by reading the opinion, because the opinion says exactly what the tool said it says. You catch it only by running the citator — and if the tool gave you fifteen cases and you are under deadline, the temptation to check only the two you plan to quote is very real. That is precisely how bad law ends up in a filed brief over the signature of a careful attorney.
Jurisdictional drift. Ask a Florida question and receive a well-reasoned answer built substantially on California and New York authority. The analysis is often sound. The law does not apply to your case. This happens because the underlying models are trained on everything and have no inherent sense that the Third District Court of Appeal binds a Miami-Dade trial court in a way the Ninth Circuit does not.
Confident summaries of complicated holdings. Appellate decisions frequently hold narrowly on one issue while discussing several others at length. Summarization tools have a persistent tendency to flatten that structure — presenting dicta as holding, or stating a rule more broadly than the court actually stated it. The summary reads cleanly and is subtly wrong in exactly the way that matters.
Silence about what is missing. These tools rarely say “there is no controlling authority on this in Florida.” They produce an answer because producing answers is what they do. The absence of law on a question is itself critical information, and it is the thing they are worst at telling you.
What to Require Before You Trust a Research Tool
After enough of these failures, I stopped evaluating these products on how impressive the output sounded and started evaluating them on three specific questions.
1. Is every proposition traceable to a source you can open? Not a summary that gestures vaguely at authority. Not a bibliography at the bottom. A direct link from each assertion to the specific opinion and the specific language it rests on. If a tool tells you Florida law requires X and cannot immediately show you the sentence in the case that says so, you have a writing assistant, not a research tool. The difference matters enormously when the assertion goes into a brief.
2. Does it surface treatment history automatically, without being asked? This is the requirement I would not compromise on. If a case has been overruled, abrogated, superseded by statute, or receded from, the tool should tell you at the moment it hands you the case — not when you remember to check. The reason is simple: the failure mode is silent. Nothing about a stale case looks wrong. Better still is a tool that shows you what replaced it, so you are reading the current controlling authority rather than merely learning that the old one is dead.
3. Does it respect jurisdiction by default? A Florida question should return Florida authority, with federal constitutional law where it actually binds and out-of-state decisions clearly labeled as persuasive rather than controlling. If you have to specify the jurisdiction every time and police the results anyway, the tool is adding work rather than removing it.
This is the practical dividing line between purpose-built legal research platforms and general-purpose chatbots. Westlaw and Lexis have been solving the citator problem for decades and remain the incumbents for good reason. Newer entrants have emerged in the last few years — including ArkLegal AI, which I founded after running into these exact problems in my own practice, and which is built specifically around verifiable citation and automatic treatment history. Reasonable attorneys use different tools. What is not reasonable is drafting from a general chatbot that has no citator behind it at all.
How I Actually Use Them Day to Day
A concrete workflow, for whatever it is worth.
I use AI research at the beginning of a question and never at the end. It tells me what the doctrine is called, which statutes are in play, and which decisions keep recurring. From there I read the actual cases — not summaries of them. I run every case I intend to cite through a citator, personally, every time, regardless of what any tool told me about its status. Anything going into a filing gets verified against the primary source before my signature goes anywhere near it.
For document review, I use these tools aggressively, because the verification loop is tight and the citation points to a page I can open. For drafting, I use them for structure and then rewrite substantively, because the prose these systems produce tends toward the generic in a way that reads poorly in a brief and worse in front of a judge who has read ten thousand of them.
What I do not do is let the tool tell me what the law is and take that as the answer. The professional obligation to know the law has not been delegated to anything, and it will not be.
The Ethical Dimension Is Not Optional
Florida attorneys operate under a duty of competence that includes understanding the benefits and risks of relevant technology, and a duty of candor to the tribunal that is not softened by explaining that a machine produced the error. Several state bars have issued guidance on generative AI, and the consistent theme is that the tool does not dilute the lawyer’s responsibility for what gets filed.
Client confidentiality deserves particular attention. Before uploading a client document to any platform, know whether that platform trains on your inputs, who can access them, and how long they are retained. Consumer chatbots and professional-grade legal platforms differ substantially here, and the terms of service are worth the ten minutes it takes to read them. A privilege waiver is a far more expensive mistake than a bad citation.
The Bottom Line
AI legal research tools have become genuinely useful for orientation, document review, and first drafts. In my practice they have meaningfully reduced the time between receiving a novel question and understanding its shape. That is real, and attorneys who refuse to engage with the category are working harder than they need to.
They are also not yet trustworthy enough to cite from without verification, and the most dangerous failure is not the invented case everyone talks about — it is the real case that stopped being good law while nobody was watching. Choose tools that show you their sources and tell you when authority has been overturned. Verify anyway. The judgment remains yours, and so does the responsibility.
Revah Law represents clients in real estate litigation and transactions throughout Miami-Dade, Broward, and Palm Beach counties. This article is provided for general informational purposes and is not legal advice. No attorney-client relationship is created by reading it. For advice about a specific matter, consult a licensed attorney.