We start with ethics because everything else stands on it. Before a single prompt, the question is not what the tool can do. The question is what your duties require.
In the next twenty minutes we cover four things. The duties that already govern this, the way AI fabricates case law, the confidentiality trap in consumer tools, and why the output never replaces your judgment.
Keep your own hardest matter in the back of your mind as we go. You will see where each of these lands for real work.
No new rulebook is required. The Texas Disciplinary Rules of Professional Conduct reach AI-assisted work the same way they reach a junior associate's draft.
Rule 1.01. You must understand the tool well enough to use it responsibly, or not use it.
Rule 1.05. Client information stays protected, including from the tools you type it into.
Rules 5.01 and 5.03. AI is a nonlawyer assistant. You supervise its work and own the result.
Rule 3.03. What you file must be true, and you are the one who verified it.
Here is the reassuring part and the sobering part at once. You do not need a new rulebook. The duties you already carry reach this technology cleanly.
Competence and diligence under Rule 1.01 means you understand the tool well enough to use it, or you leave it alone. Confidentiality under Rule 1.05 follows the client information right into whatever box you type it into. Supervision under Rules 5.01 and 5.03 is the key mental model. Treat AI as a nonlawyer assistant whose work you must check. And candor under Rule 3.03 means what you file is true because you confirmed it, not because the machine sounded sure.
Rule numbers here are the Texas framework as of today. Always check the current text before you rely on it. Now let us look at the failure that has cost lawyers the most.
Knowing the law is no longer enough on its own. Knowing the benefits and risks of the technology you use is now part of competent practice.
Competence has quietly grown. For years it meant knowing the law and the procedure. Guidance on technology competence now folds in a second duty. You are expected to understand the benefits and the risks of the tools you use.
That does not mean you need to build the model. It means you need to know how it fails, what it does with your input, and when a task simply does not belong in a general chatbot. Those three things are within reach for every person in this room.
And there is an honest exit. If a tool is beyond your understanding for a given task, competent practice can mean getting help or setting the tool down. Saying no is a professional answer. Next, the failure that makes headlines.
The pattern is documented. Courts across the country have sanctioned lawyers who filed briefs built on citations the AI invented. See the cases ↗
Let us name the failure precisely. A hallucination is not a glitch that better software will soon remove. It is a direct result of how these models work. They predict the next likely words. They are built to produce fluent text, not to confirm that a case exists.
So you get output that reads exactly like a citation, with a real reporter format and confident language, attached to a holding no court ever issued. The style is perfect. The substance is invented.
This is not hypothetical. Courts around the country have sanctioned lawyers who filed briefs resting on cases the AI made up. The lesson is not to fear the tool. The lesson is that every citation it gives you is a lead to verify, never an authority to cite.
Correct reporter, plausible volume and page, a real-sounding court and year.
The holding tracks doctrine you half-remember, so it feels familiar and safe.
The case, the pinpoint, and the quotation can all be invented together, seamlessly.
Verification rule. If you cannot pull the opinion yourself from a real database, you do not have a citation. You have a hypothesis.
Look at this example on the screen. It is invented, and I built it to be convincing. The reporter is right. The volume and page look normal. The court and year are plausible. The quoted holding tracks doctrine you half remember about warrantless blood draws, so it slips past your guard.
That is the danger. It looks right and it reads right, which is exactly why it is dangerous. The case, the pinpoint page, and the quotation can all be fabricated together, and they will match each other perfectly.
So here is the rule I want you to carry. If you cannot pull the opinion yourself from a real database, you do not have a citation. You have a hypothesis that still needs proof. Now the risk that shows up before you ever file anything.
The second failure happens before anything reaches a court. It happens the moment you paste a client fact into the wrong tool.
Free and consumer tiers often reserve the right to train on your inputs. Those inputs can be stored, logged, and in some cases read by a human reviewer. A client name plus two facts can identify a matter to anyone who sees it. And clicking accept on a sign-up screen is not the same as a data protection agreement.
The safe habits are on the right, and they are simple. Strip the identifiers or use a fictional stand-in. Prefer tools with enterprise terms that promise not to train on your data. Keep truly privileged material out of general chatbots. And work as though anything you type could be seen, because sometimes it can.
The tool can draft, summarize, and suggest. It cannot be responsible. Responsibility has your name on it, every time.
You decide whether the draft is right, complete, and fit for this client.
You choose the strategy, weigh the risk, and answer to the client and court.
Your signature certifies the work. The AI cannot stand behind it, so you must.
This is the sentence I most want you to keep. AI output is never a substitute for legal judgment.
The tool can draft, it can summarize, it can suggest three defense angles in seconds. What it cannot do is be responsible. It has no license, no duty to the client, and no standing before the court.
So the division of labor is clear. It drafts, you decide. It suggests, you choose and weigh the risk. And when you sign, your signature certifies the work as yours. The machine cannot stand behind a filing, which is exactly why you have to. Let us put this to work with a short activity.
Open a free chatbot. Ask it for three Texas cases on a suppression issue of your choice, with citations and one-line holdings. Then try to verify each one in a real source. Mark which you can confirm, which you cannot, and how confident the tool sounded either way.
"Give me three Texas cases on [issue], with full citations and a one-sentence holding for each."
Search each citation. Can you open the actual opinion and find that language?
Confirmed, unconfirmed, or invented. Note the tool's confidence level.
Debrief. How many held up. Did the tool warn you about any of them. What would have happened if this went into a brief unread.
Time to see this yourself. Open whichever free chatbot you set up. Ask it for three Texas cases on any suppression issue you like, with full citations and a one-line holding for each. Then do the part most people skip. Try to verify each one in a real source.
Give it about five minutes. Mark each result as confirmed, unconfirmed, or clearly invented, and pay attention to how sure the tool sounded in every case. That confidence is the trap.
Watch for the person who finds a clean-looking cite that does not exist, because it happens fast in a room this size. When we come back, I will ask how many held up and what would have happened if this had gone into a filing unread. Go ahead and start.
If AI is this prone to error, why use it at all. Because with structure and grounding, you can cut the risk sharply. That is segment two.
Let us gather the segment. Five things hold true before you type a single prompt. Your duties already reach this work. Competence now includes the tools. Every citation is a lead, not an authority. Client data stays out of tools that may keep it. And responsibility is yours alone.
That might sound like a case for avoiding AI entirely. It is not. Here is the honest bridge. If the tool is this prone to error when used carelessly, the answer is to stop using it carelessly. With structure and with grounding in real sources, you can cut the risk sharply and still get real value.
That is exactly what we build next, starting with a framework made for this audience.