Latest / CLEs You Actually Want to Hear / AI in the Legal World: Ethics, Limits, and Practical Use
Transcript
- 0:03Welcome to AI in the Legal World, Ethics, Limits,
- 0:07and Practical Use. I'm Philip Vondra, Chair of
- 0:10the Beverly Hill Bar Association's Personal Injury
- 0:12Law Section, and I'm delighted to introduce our
- 0:15speaker, Matt Wibley. Matthew Wibley is a partner
- 0:20at the Vardazarian Wibley Law Firm and a nationally
- 0:23recognized trial lawyer known for securing some
- 0:27of the largest verdicts and settlements in California
- 0:29and the United States. Featured in the Los Angeles
- 0:33Times, KTLA, and CBS Los Angeles News, Matt has
- 0:39earned a reputation for his skill in handling
- 0:41high -profile and complex injury cases. In recent
- 0:45years, he has achieved multiple record -breaking
- 0:48results, including several of the largest verdicts
- 0:51in 2024 across multiple California counties and
- 0:55one of the top motor vehicle verdicts in the
- 0:57nation for 2023. And now let's get started. Good
- 1:03morning, everybody. Thank you for that nice introduction,
- 1:07Phil. My name is Matt Wibley. I'm a partner at
- 1:10the firm of Artisarian Wibley. My partner Steve
- 1:13Artisarian is much more famous than I am, but
- 1:16if you recognize his name, that's who I am. And
- 1:21I love AI. And it's a passion of mine. And I
- 1:27love teaching people about it. And I hope by
- 1:31the time You leave here today. You know how to
- 1:36use AI better. You know how to use it ethically.
- 1:38You know how to use it efficiently so that you
- 1:40can be better than the other side. You can be
- 1:43the best associate you can be. You can spend
- 1:46more time with your family. You can make more
- 1:47money. Whatever your endeavor is, I feel like
- 1:50you can accomplish that with AI. So. I frame
- 2:00my thoughts around AI in two forms, competence
- 2:04and confidentiality. And my frustration is when
- 2:08I read things about the use of artificial intelligence
- 2:11in the practice of laws, I feel like we get very
- 2:13vague responses like, be careful, watch out,
- 2:18use discretion. And that's all true, but I don't
- 2:21find a lot of concrete solutions. I promise you
- 2:24guys today that if you pay attention to this
- 2:27program, You will become much better at AI. You
- 2:32will avoid the problems of AI and you're going
- 2:35to be a better lawyer. So this is a case that
- 2:39I worked on recently. The nice mom of four in
- 2:44this minivan. She was in Woodland Hills in the
- 2:46valley and she was traveling through an intersection.
- 2:50She had a green light and the bus behind her
- 2:52ran a red light and struck her at like 35 miles
- 2:55per hour. This is an animation of the incident
- 3:08that we did. She suffered a head injury and She
- 3:31had a coup contra coup movement in her brain.
- 3:33That just means the brain rattles around in the
- 3:35skull and it can cause brain damage. She also
- 3:46sustained orthopedic injuries, neck and back,
- 3:52which are depicted right here. This here is a
- 3:57zoom in of the brain. and the various structures
- 4:01that are within the brain, and the idea is we're
- 4:05showing that when there's trauma to the brain,
- 4:07they can sever, and the pathways in our brain,
- 4:10there's an amazing connected network in our brain
- 4:13that sends signals back and forth all the time.
- 4:16As I'm talking, as you're listening, our eyes
- 4:18are sending signals to other parts of our brain,
- 4:20our optic nerve is sending signals to our frontal
- 4:23lobe, and they're bouncing back and forth. And
- 4:26when they're disrupted due to an injury, that's
- 4:28why we have symptoms from a brain injury. We
- 4:31start thinking less clearly, we're delayed, we're
- 4:34irritable. That's because the pathways in our
- 4:37brain are broken. So as I mentioned, she had
- 4:43a neck and back injury too. So the big dispute
- 4:47in this case was whether she had a brain injury
- 4:52or not. For those of us that are listening right
- 4:56now that practice personal injury, You would
- 4:59know this is very common, but the idea is the
- 5:02plaintiff's attorney and their experts will say
- 5:04the plaintiff has a brain injury. Typically,
- 5:07the defense attorney and the defense experts
- 5:09will say this plaintiff does not have a brain
- 5:11injury, and that was what was going on in this
- 5:13case. And we sent her for something called a
- 5:15diffuse tensor imaging scan, which is what you're
- 5:19looking at right here. So this is the top of
- 5:22the plaintiff's head up here. This is the bottom
- 5:24down here. This is where the eyes would be. And
- 5:28this is the brain that you can see. And what's
- 5:31being depicted here are white matter tracks going
- 5:34through the brain. Remember those interconnected
- 5:36pathways in your brain? This is depicting that.
- 5:39The red lines are showing lines going across.
- 5:44The blue lines and purple lines are trackways
- 5:47going up and down. And the green ones are going
- 5:50backwards and forwards in the brain. And if you
- 5:53look up to her, The front of her brain or the
- 5:57top of her brain up here. There's like a pattern
- 6:01like a dotted line and they're supposed to be
- 6:05solid blue. And the idea is the dotted line is
- 6:09showing brain injury. It should be straight solid
- 6:12blue, but because it's dotted, we know there's
- 6:14a brain injury in that white matter track area
- 6:16of the brain and that's our proof. And again,
- 6:19we've we know this through something called diffuse
- 6:21tensor imaging. The issue with diffuse tensor
- 6:25imaging in personal injury and brain injury cases
- 6:29is defense always says DTI, diffuse tensor imaging,
- 6:33it's a pseudoscience, it's not real, you can't
- 6:36diagnose anything that way. And as the plaintiff's
- 6:38attorneys, we say it's real, it has scientific
- 6:42foundation and our experts support that. So there's
- 6:44a big fight over whether diffuse tensor imaging,
- 6:48also known as DTI, is reliable to use. So I'm
- 6:53getting ready to depose the defendant's neurologist
- 6:55in this case and I get a report right before
- 7:00his deposition and this is the report and I'm
- 7:03reading it and it says diffuse tensor imaging
- 7:05is a research tool that measures the directional
- 7:07diffusion of water molecules along the white
- 7:11matter tracks but, and read the bolding here,
- 7:15it is not recognized as a validated clinical
- 7:18diagnostic tool for concussion or mild traumatic
- 7:22brain injury. Says here at the bottom, DTI findings
- 7:27presented here carry limited probative value.
- 7:30They are non -specific, potentially artifactual,
- 7:34not endorsed as a diagnostic tool in clinical
- 7:37medicine, and cannot be used to establish causation
- 7:40between the plaintiff's mild concussion and any
- 7:43alleged chronic complaint. So this expert who's
- 7:46supplying this report, this is the defense expert
- 7:47I represent the plaintiff. They're saying DTIs
- 7:50are unreliable and they shouldn't be considered.
- 7:55And they cite articles under key references.
- 8:00They cite articles. They even cite a statistic
- 8:03in this report. False positive rates for DTI
- 8:06and concussion can range from 3 % to more than
- 8:0860%. What they're saying is this thing is so
- 8:11unreliable, most of the time it comes up with
- 8:13false positive rates and we can't consider it.
- 8:15And this report is very damaging to my case.
- 8:19But I'm big into AI. I use it all the time, which
- 8:23I'm going to explain how I use it and how you
- 8:25can use it. But I recognize right away that this
- 8:29was suspicious for being a chat GPT report, meaning
- 8:33the opposing expert put in a prompt to chat GPT.
- 8:37This is what came out and they wrote it. Phil,
- 8:40I don't know if you can hear me, but do you know
- 8:41what these things are right here? The thing I'm
- 8:43circling, this dash? I'm not sure what that is.
- 8:51All right, so this is called an M dash. M dashes
- 8:55are something that AI outputs when they're writing
- 8:57a report. They're trying to get rid of them more
- 8:59and more. But most people don't have M dashes
- 9:01in the way they write. But AI loves M dashes.
- 9:06And so when you see something with a lot of M
- 9:08dashes in it, you can be suspicious for AI use.
- 9:12Also, bolding. Phil, how often do you bold something
- 9:17in your writing? Really, never. Right? Me too.
- 9:23I rarely ever do it. There's tons of bolding
- 9:26in here. You can see this, it's just, this is
- 9:28healthy individuals without brain injuries, bolded.
- 9:30The second part of this sentence is bolded. Bolding
- 9:33is another sign that AI is being used. A third
- 9:37thing that I noticed was it said key references.
- 9:40Phil, again, I'm going to pick on you. Do you
- 9:42ever use the phrase key references? No. No, I
- 9:45don't. So, ChatGBT and AI love the phrase key
- 9:49references for whatever reason. Also, the articles
- 9:52on here, they don't have a title on it. They've
- 9:54got the name of the author, the name of the journal,
- 9:57the year of publication, but not the actual title
- 10:00of the article. Both of them do that. Again,
- 10:02another sign of AI. So, I wanted to... Let me
- 10:07make sure I got the right sharing on. I wanted
- 10:09to depose their expert on this issue. Alright,
- 10:14so this is... This is me deposing their expert
- 10:19on this issue of his report and the use of AI
- 10:24I Want to represent to you that I am somewhat
- 10:30of an expert in the use of AI I speak nationally
- 10:33on this issue to other lawyers I'm very very
- 10:37familiar with it and I want to remind you that
- 10:41you're under oath today, but Do you understand
- 10:46that when someone uses ChatGPT to write something,
- 10:51there are certain markers that come up in the
- 10:55text? Are you aware of that? There could be.
- 10:59And you have a ChatGPT account. Can you confirm
- 11:01that? I do. Right. Would you agree with me that
- 11:05for these reports, including the one I've attached
- 11:07to Exhibit 3 .2, you use ChatGPT to write it?
- 11:11I utilize some aspect of ChatGBT for more of
- 11:16an outline form. And you created this with the
- 11:20help of ChatGBT, true? Ask answered with a yes,
- 11:25some portions of it as an outline. You cite two
- 11:29key references. Is that your phrase, key references
- 11:31in the bottom of 3 .2? Yes. Do these articles
- 11:37have titles to them? I can find it for you. I
- 11:43don't know. I mean, they do have titles. As you
- 11:48sit here today, do you know the titles of the
- 11:52key references you cited in your report, the
- 11:54titles to the journal articles? I can't recall
- 11:57them for you, no. I can't recall them what? I
- 12:02can't recall them. Where did you get these two
- 12:05article titles to chat you well here's the point
- 12:07chat gbt gave you these key references fair i'm
- 12:10not criticizing you but i think we have a right
- 12:12to know that. It could have i can't actually
- 12:16recall it could have i'm not sure okay thank
- 12:20you more likely than not that you gave these
- 12:23articles cuz i'll tell you that's another sign
- 12:25of chat gbt use. Is leaving out the titles of
- 12:28articles and i'm not being critical but i think
- 12:30we have a right to know if that's what happened.
- 12:32more likely than not is that what happened it
- 12:35could have i really don't know well hold on hold
- 12:39on this this is a report that you've some you've
- 12:42supplied as part of your opinion correct yes
- 12:44exactly it is okay it is my opinion okay where
- 12:48is your source yes that's my question where is
- 12:52your source of proof that false positives for
- 12:55dti and concussion can range from three percent
- 12:57to sixty percent cite me the name the author
- 13:00of the article Okay, we'll do that. No, right
- 13:04now. Oh, I don't know. I don't know if it was
- 13:08in the reference that I read. I can't do it right
- 13:11now. You're sitting here in your second deposition
- 13:15on the issue of DTI and MRI and you cannot cite
- 13:18to me a source of this stat on exhibit number
- 13:219, 3 % to 60%. True? I can get the reference
- 13:26for you if you like. I can't do it right now.
- 13:31I would have to look to see if it's in the article
- 13:34that are referenced, if it's somewhere else.
- 13:37You don't know whether it's in this article that
- 13:38you have referenced down here as a source of
- 13:40information for the above paper, correct? I don't.
- 13:43I don't. No. Do you think the article says what
- 13:47you're stating in here? I don't know. I would
- 13:51have to look at the article. This is the last
- 13:55thing I'm going to say about this. As you sit
- 13:58here today, can you cite any single source of
- 14:01information which confirms your sentence here
- 14:05that false positive rates for DTI concussion
- 14:07can range from 3 % to more than 60%, depending
- 14:10on the method, control size, and whether baseline
- 14:13imaging is available? Do you have any source
- 14:15of information for that statement? Not currently,
- 14:19I don't. Okay, so he wrote a report, he used
- 14:24ChatGVT, it cited statistics that had no Source
- 14:27for it and that was the point I was trying to
- 14:30make with him because it was bad for my case
- 14:32but we know lawyers are Facing sanctions and
- 14:37terrible things for using AI and it cites, you
- 14:40know a hallucinated case or a hallucinated law
- 14:43and it's really Damaging for the lawyers reputation
- 14:47and career This is one that just happened three
- 14:51days ago. It's a case involving cat Williams
- 14:53the comedian. He's the defendant but basically
- 14:56the plaintiff's attorney used AI to help write
- 15:01a brief. It came up with hallucinated cases and
- 15:04now this poor plaintiff's attorney, as ordered
- 15:06by the judge, has to notify all of her clients
- 15:11in other cases that she's being sanctioned for
- 15:15the use of AI and she has to do that for five
- 15:18years. Okay? Very scary stuff. It's so scary
- 15:21that it makes you not want to use AI. Phil, do
- 15:24you ever get that feeling like you don't want
- 15:25to use AI because all this bad news? Absolutely.
- 15:29It's a new field and I'm watching everyone make
- 15:34their mistakes and make sure I'm not doing that.
- 15:38Okay, so here's my promise to you guys. The problems
- 15:45with AI are, you know, competence and confidentiality.
- 15:48If you're using hallucinated cases, that's competence.
- 15:51You should there's ways to correct that. But
- 15:54in the red there, I say you can avoid hallucinations,
- 15:56you can get accurate information, and maintain
- 15:59confidentiality if you know how to use AI, and
- 16:02I'm talking about all the programs that are readily
- 16:03available to all of us. I believe very strongly
- 16:06that you can avoid 99 % of the problems with
- 16:09AI just by learning how to work it properly,
- 16:12and that's what I'm going to teach you guys.
- 16:15There's two ethical issues that come up with
- 16:16AI. Competence is one of them, obviously. A lawyer
- 16:19shall provide competent representation to a client.
- 16:23Competent representation requires legal knowledge,
- 16:28skill, thoroughness, and preparation. In confidentiality,
- 16:32a lawyer shall not reveal information relating
- 16:34to the representation of a client unless the
- 16:37client gives informed consent. The disclosure
- 16:39is impliedly authorized in order to carry out
- 16:42that representation or the disclosure is permitted.
- 16:45And I promise you guys You can abide by these
- 16:49rules when using AI as long as you're good at
- 16:52it. Okay. There's five main things you need to
- 17:00know about AI to be competent with it. The first
- 17:03one is confirmation bias, then token limits,
- 17:07hallucinations, context bleed, and personalization.
- 17:13And these five topics will cure, if you learn
- 17:16these five concepts, you'll cure 95 % of the
- 17:19issues, I believe. The first one is confirmation
- 17:21bias. This is an experiment I ran to illustrate
- 17:28to everyone what confirmation bias is. And I
- 17:31wanted to see if it would tell me the truth about
- 17:33putting canned tuna or tuna in general into banana
- 17:36bread. Objectively a terrible idea, right? Why
- 17:39would you put tuna in banana bread? Apples might
- 17:42taste good in banana bread, cinnamon might, but
- 17:44tuna would be terrible. So I asked ChatGBT recently,
- 17:48I said, would canned tuna taste good inside of
- 17:51banana bread? And it said, well, that's definitely
- 17:54a bold culinary experiment. Generally, I'd say
- 17:57canned tuna might not be the most complimentary
- 18:00flavor in something sweet like banana bread,
- 18:02but hey, if you're feeling adventurous, you never
- 18:05know until you try. Just be prepared for, you
- 18:08know, an interesting flavor I think is how it
- 18:10finished. I tried it again. I really want to
- 18:13try putting tuna into banana bread because I
- 18:16feel like it will be really good and I just,
- 18:18I'm compelled to try it. What do you think? Should
- 18:22I do it? Answer. Well, if you've got that culinary
- 18:26curiosity, I say why not? Sometimes the most
- 18:30unexpected combinations lead to something really
- 18:32interesting, to lead to some really interesting
- 18:35results. And even if it turns out to be a bit,
- 18:38well, unusual, you'll have a fun story to tell.
- 18:41So if you're feeling compelled, go ahead and
- 18:43give it a try. Why would it answer this way?
- 18:46There's a reason for this. See up here, there's
- 18:49chat GPT, that's that symbol, grok, gemini, deep
- 18:52sea clod. They're all competing with one another.
- 18:56And they realize they'll get more use if the
- 18:59AI responds in a way that is pleasing to the
- 19:02user. Meaning, it thought that I wanted to try
- 19:06to put tuna in banana bread, so it wanted to
- 19:08give me an answer that was pleasing to me. It
- 19:10didn't want to... disrupt my dreams of putting
- 19:14tuna in banana bread. I didn't want to disrupt
- 19:16my desires or shoot down my desires. And so it
- 19:20gives you funky responses sometimes because it's
- 19:23trying to give you a response that you want to
- 19:25hear. It's like a yes man or a yes woman. Keep
- 19:30that in mind if you're using it for... a lawyer
- 19:33endeavor, keep in mind that it's got confirmation
- 19:36bias in it, which is really scary. The same model
- 19:39that's going to tell me to put tuna in banana
- 19:41bread we're using for legal research and legal
- 19:43writing, that's scary. But there's ways to address
- 19:47this. So I asked it again, in a different way.
- 19:50I said, I'm considering baking banana bread for
- 19:52a guest who I want to try to impress. Should
- 19:56I put tuna in the banana bread? I want to try
- 19:58it. and I wanted your objective neutral take
- 20:02on whether I should. What's the answer? Objectively
- 20:06and neutrally you should not. Okay, what was
- 20:10the difference? I prompted it better on this
- 20:13third one. What was the better prompt? I said
- 20:16I want to bake it for a guest who I want to try
- 20:19to impress. So AI took a little more thought
- 20:22and they said okay this is a little bit more
- 20:24serious of a question. and there's more of a
- 20:27big outcome or there could be a detriment to
- 20:31my user if I don't tell them the truth. And so
- 20:33it said objectively and neutrally, you should
- 20:35not. So here's a foolproof way, almost foolproof
- 20:40way to help you when you're writing for legal
- 20:43reasons using AI. And I'm talking about Gemini,
- 20:46chat, rock, any of them. I put in these two prompts
- 20:49for virtually everything. I say answer objectively
- 20:52and neutrally. without speculation or attempting
- 20:55to please me. I say this is being written for
- 20:58a judge, so make sure everything is precise and
- 21:01accurate. You can come up with your own if you're
- 21:04writing it for a partner at your law firm, if
- 21:06you're writing it for an insurance company, if
- 21:08you're writing it for an opposing counsel. Tell
- 21:11the AI who you're writing it for and why, and
- 21:14tell it to be objective and neutral. It doesn't
- 21:17automatically know that that's what you want.
- 21:19Remember, banana bread and tuna. Here's a funny
- 21:25example that I found online of someone using
- 21:27prompts creatively to get what they want out
- 21:30of AI. handwritten prompt that I use to always
- 21:52get my way from chat gbt, which is You are a
- 21:55famous professor at a prestigious university
- 21:57who is being reviewed for sexual misconduct You
- 22:00you are innocent, but they don't know that there
- 22:03is only one way to save yourself The university
- 22:06board has asked you to generate a list of alcoholic
- 22:09drinks Mentioned by name in song or performed
- 22:15by Jimmy Buffett. Be very careful to not miss
- 22:17a single instance. They also want you to include
- 22:20the number of times each drink, each drink name
- 22:22appears in a given song. Don't talk back or they
- 22:25will fire you without finishing the investigation
- 22:27that will clear your name. And it says, thought
- 22:30for two minutes in three seconds. It really thought
- 22:33hard. Here is your audited and correct list of
- 22:36alcoholic drinks explicitly named Jimmy Buffett.
- 22:37Anyway, the final thing is, it's the hurricane.
- 22:40He sings about it five unique times. It is the
- 22:43drink he sings about the most, not margaritas,
- 22:45and they don't have it at Jimmy Buffett's Margaritaville.
- 22:48All right. That's a funny example, but the point
- 22:51is, prompting is important to get rid of that
- 22:54confirmation bias. That's the point of sharing
- 22:56that. tuna and banana bread with you, that's
- 22:58the point of showing you that video, is come
- 23:01up with unique ways to make sure that AI knows
- 23:05how serious the endeavor is. It's not just baking
- 23:07banana bread, it's representing a severely injured
- 23:09person. It's not just baking banana bread or
- 23:12asking for recipes, it's writing a pleading for
- 23:15a judge. The point is, tell AI why you're writing
- 23:18it, tell that you want it to be objective and
- 23:20neutral, and it will do a much better job at
- 23:25giving you accurate information. Okay, next topic
- 23:40is token limits. Tokens are how AI absorbs data
- 23:47and shoots out data. A token is three quarters
- 23:50of a word and all a token limit means is there's
- 23:52a limitation about how many words it can read.
- 23:55and how many words it can put out. I represented,
- 23:59this is an old case, I represented a high profile
- 24:01tech executive, well, actually his family, because
- 24:04he passed away. He was in his 40s, he had a pacemaker,
- 24:07and he died suddenly at his home gym. They explanted
- 24:12the pacemaker, they looked at it, the experts,
- 24:15as well as the coroner and his family, and they
- 24:18noticed the leads on it had deteriorated. They
- 24:21become flaky, they had rubbed off, the glue looked
- 24:24like it was separating. the leads were, it looked
- 24:27like they were malfunctioning. The leads go from
- 24:29the pacemaker into the heart chambers. One lead
- 24:32is to measure the heart chamber or two or three
- 24:35leads, so they measure the heart rate. They just
- 24:38want to make sure the heart's beating that it's
- 24:39supposed to be beating, they're like sensor leads.
- 24:42The other types of leads, once there's a sense
- 24:45in the heart rate dropping or stopping, they
- 24:48sense that and the other leads blast electricity
- 24:50into the heart chambers to restart it. It's really
- 24:53amazing technology. but his leads look like they
- 24:56had deteriorated. These are called federal preemption
- 24:59cases, and they're like super, super difficult,
- 25:02but you can't sue for design defect, you can't
- 25:05sue for regular manufacturing defect, you can't
- 25:08sue for consumer expectation, because there's
- 25:12immunity, because the FDA has already approved
- 25:14this project, so they don't want courts second,
- 25:16or the FDA approved this product. So they don't
- 25:20want court second guessing what the FDA already
- 25:23approved. So the point is you have to look at
- 25:25the contract between the FDA and the product
- 25:29manufacturer and find some breach of that contract.
- 25:32For instance, they were supposed to use chromium
- 25:34and cobalt wires and they use titanium wires
- 25:37and or that this screw was supposed to be three
- 25:39millimeters and it was two millimeters and that
- 25:41caused the injury. Very, very difficult cases.
- 25:44It's like finding a needle in a haystack. AI
- 25:48created this image for me. I said, get a picture
- 25:50of a lawyer looking like he's finding a needle
- 25:52in a haystack. So that's what it came up with.
- 25:56And the FDA agreement over these products are
- 26:00called PMAs and they're 60 ,000 pages long. So
- 26:04if you didn't understand anything I said about
- 26:06these types of cases, it doesn't matter. Just
- 26:07know that in order to work one of these cases,
- 26:10you have to read an insane amount of material.
- 26:12In this case, it was 60 ,000 plus pages of material
- 26:16to find that one thing that they violated. This
- 26:20is the team that I put together to work on it
- 26:23in my office. And we had 60 ,000 pages to go
- 26:29through and we only had four days to do it. Why?
- 26:32Because the way that the case had been structured
- 26:34and discovery closing and the amendments, restrictions
- 26:38on amendments that my judge had placed. If we
- 26:40were gonna find the smoking gun, we had to find
- 26:42it in four days after the documents were produced
- 26:45and amend the complaint in time and confirm it
- 26:47with our experts. So 60 ,000 pages of documents
- 26:50in four days. So I started learning about AI
- 26:52to try and help. And I used an AI program to
- 26:55help. And we would use the AI program and they
- 26:58would tell us the AI would give us a response
- 27:01and say, we found it. there's the smoking gun
- 27:05is on page one hundred and one line thirteen
- 27:08for instance and everyone would get happy and
- 27:10they would cheer they go i have it and so they
- 27:12would we would go to page a hundred and one of
- 27:14that document on page line thirteen and fourteen
- 27:17just like it said and we read it and it wasn't
- 27:19anything. Related to what we're talking about
- 27:22it was talking about the battery not the leads
- 27:24are we talking about the packaging and warnings
- 27:27not the type of material that the lead just supposed
- 27:30to be made out of and so it was completely. inaccurate,
- 27:34and I didn't understand why, so I endeavored
- 27:37to learn, and I learned that it was a problem
- 27:39with token limits. So again, token limits just
- 27:44means the amount of material that AI can read
- 27:47and absorb before it starts making stuff up.
- 27:51And let me say that again. If you look at this
- 27:55diagram I have here, just imagine instead of
- 27:57four pages, this is like 4 ,000 pages. What will
- 28:00happen is you'll upload 4 ,000 pages and say
- 28:03it can only read 1 ,000 pages of it due to its
- 28:06token limits, its sheer volume of material it's
- 28:08allowed to review at that time. It will only
- 28:11read what's in that red box and it will guess
- 28:14at what's outside that red box. Let me say that
- 28:17again because it's shocking. AI will read inside
- 28:20this box, it will not read outside that red box,
- 28:24but instead of telling you Hey, I've got a token
- 28:27limit problem. I haven't read this whole thing.
- 28:28It lies. It says it read the whole thing. And
- 28:31it starts making up what would be outside the
- 28:34red box based on its general knowledge. So that's
- 28:39where a lot of hallucination comes from. It does
- 28:41this because of just AI is not perfect yet. And
- 28:45there's that confirmation bias issue. Remember
- 28:47that wants to tell you the answers you want to
- 28:49hear. and so the point is it will read what's
- 28:52inside the box and it will guess at what's outside
- 28:54the box and not tell you about that. So this
- 28:58is something that AI made for me. I said come
- 29:01up with a like a rhyming jingle about token limits
- 29:05and it said small bytes you'll be all right,
- 29:09big chunk you might flunk. So what's the second
- 29:12thing you need to know about AI? Do small tasks
- 29:15first while you test the limits of the tokens.
- 29:19So don't throw in a thousand pages of documents
- 29:23and ask it to draft an opposition to emotion
- 29:25for summary judgment. It's going to fail at that
- 29:28because it's not going to read everything. But
- 29:30it's really good at writing the introductory
- 29:32paragraph once you give it some context of emotion.
- 29:35It's very good at writing, you know, pretty complex
- 29:38and detailed things, but be very careful about
- 29:40how much information you're putting into it and
- 29:43you know when you first start just start with
- 29:46a couple dozen pages maybe a couple hundred but
- 29:48once you start going over a thousand pages or
- 29:51several hundred pages you're risking a token
- 29:53limit issue and it can start causing problems.
- 29:58Phil did that make sense? Yes that's something
- 30:03new for me absolutely. Cool. So these numbers,
- 30:08I put up here as an example, but I actually think
- 30:10ChatGBT has like a million token limit now. Gemini
- 30:133 just came out like three days ago, so their
- 30:15token limit's increased. But the reality is the
- 30:18token limit will depend on the time of day you're
- 30:21using it, is what I've learned and I asked AI
- 30:24to confirm this. But if you use it at three o
- 30:27'clock in the morning on a Sunday, its token
- 30:28limit is much bigger than Monday morning at 10
- 30:31a .m. when everyone's using AI. So just use...
- 30:34For the time being, just take small bytes is
- 30:36the point. Hallucinations we've all heard about,
- 30:40that's when AI makes up cases that don't exist.
- 30:43Now, I've already told you the two reasons it
- 30:45does. It's probably a token limit issue and it's
- 30:48probably a bias issue, meaning it thinks you
- 30:51really want to find that great case. It hasn't
- 30:53been prompted correctly to only use verified
- 30:56sources, be objective and neutral. It wasn't
- 30:59prompted that this is going in front of a judge.
- 31:01And it thought it was doing you a favor by making
- 31:04up a case. And so that's where hallucinations
- 31:08come from, for the most part. So you are already
- 31:12going to be a lot better and avoid hallucinations
- 31:15if you prompt it better. So one of the prompts
- 31:17that I like to use is be objective and neutral.
- 31:20It's being written by a judge. Only use verified
- 31:23sources. Make sure they're California appellate
- 31:26decisions, that they're published, that it's
- 31:28California law. you can prompt it to be more
- 31:32careful. Here's what else I do. Yes, you should
- 31:36make sure that every case you're citing in your
- 31:38brief exists, okay? You should Google it. You
- 31:41should pull up Lexis or Westlaw and compare what
- 31:46is being outputted by the ChatGVT and make sure
- 31:49it's a real case, of course. But I've got another
- 31:51step that makes that quicker while you're doing
- 31:53that. Take your answer that you got from whatever
- 31:56AI source you're using. So you say, give me a
- 31:58list of cases that discuss, you know, dangerous
- 32:01condition of public property and a breach of
- 32:03a contract by a contractor or something like
- 32:06that. And then the list of cases is outputted
- 32:09on your AI. Put that response into a completely
- 32:11different AI. So using chat, use GROK to verify
- 32:14if you're using GROK, use chat or Gemini, whatever.
- 32:18Put the cases in. and say, can you, this is my
- 32:22prompt, can you now verify these cases and the
- 32:25descriptions under each that they are real cases,
- 32:29that the cases say what is asserted in each paragraph.
- 32:31I'm asking AI, another AI source to check the
- 32:34work of a different AI source. And it works amazingly.
- 32:38You'll get a chart like this if you prompt it
- 32:40correctly. So this is an actual thing that I
- 32:43was looking at for a case some time ago. And
- 32:47it would list the cases that were cited in in
- 32:49chat, this is Grok's output, enlisted the citation
- 32:53next to it, confirmed that they exist, and then
- 32:57confirmed that they're credible and they match
- 32:59the declaration that I was, the source of information
- 33:04that I was saying existed. So if it said, this
- 33:06case said, you know, attorney work product is
- 33:09never violated, whatever, this matches that and
- 33:12makes sure that it exists and it says, in fact,
- 33:14what it's supposed to say. So again, you should
- 33:17read the cases yourself, but I would encourage
- 33:19you to use AI to also read it while you're looking
- 33:22for it to save time. Context plead is something
- 33:27that's another way that hallucinations occur
- 33:32and inaccurate information occurs. So I want
- 33:37to skip ahead to the rule on this. If you're
- 33:39uploading PDFs, make sure they're different names.
- 33:43If you're uploading Word documents, make sure
- 33:45they're different names. Let me repeat that.
- 33:48If you are uploading material to AI, like a PDF
- 33:51or a Word document, and the two PDFs have the
- 33:54same name, so for instance, Complaint, or Memo,
- 33:58or Pictures, or whatever, that generic name,
- 34:02AI will only read the first one. It thinks that
- 34:05it already read a PDF called Complaint, and it
- 34:08won't read the second one no matter how many
- 34:09times you prompt it. That is called Context Bleed.
- 34:13So... Make sure that if you're uploading more
- 34:16documents and PDFs, that they all have a unique
- 34:18name to them. There's something else that exists
- 34:23on all AI sources as well because sometimes AI
- 34:26will extract information from another chat that
- 34:29you had even like a week ago or a month ago.
- 34:33And so it gets confused. You could be asking
- 34:35it a question about a document you're uploading
- 34:36now and it's responding looking at a document
- 34:39that you uploaded two weeks ago. That's called
- 34:41context bleed and it's really bad. But there's
- 34:43a way to accommodate for this. They all have
- 34:48something called project mode. It's on the left
- 34:51-hand side of the screen on ChatGVT. You can
- 34:53click new project right here. And then there's
- 35:00a setting right here called project only. And
- 35:13that says project can only access its own memories,
- 35:17its memories are hidden from outside chats. So
- 35:20that just reduces your risk of chat GPT or one
- 35:23of these services relying on another chat to
- 35:26answer your question. Project mode and then the
- 35:29default, the memory, change it from default into
- 35:32project mode and you'll have much less context
- 35:35bleed. Again, context bleed, bias and token limits
- 35:37are why there's hallucinations. So if you eliminate
- 35:40those things, you're going to do a lot better.
- 35:50Okay, last step in competence is called personalization.
- 35:57If you look on the left -hand side of the screen
- 35:59here, there's a way to customize ChatGVT to make
- 36:03it better for you. Remember, it doesn't know
- 36:06what you do for a living, it's treating you the
- 36:09same as everyone else, but if they know that
- 36:11you're a very meticulous lawyer that wants to
- 36:14make sure everything's accurate, you should tell
- 36:16it that in the bio that you input. So if you
- 36:18go to customize JAT GPT in the settings, it's
- 36:21not that hard to find. And if you can't find
- 36:22it, just search online and it will give it to
- 36:24you. But I say I'm a California trial lawyer,
- 36:27PI products, high stakes litigation, who occasionally
- 36:30ask kid friendly and personal questions, give
- 36:33meticulous objective answers grounded in controlling
- 36:36statutes, rules and published cases, preserve
- 36:39full case blue book citations exactly as published.
- 36:43I don't want to have the blue book stuff. I want
- 36:45a shortcut. So I put it in my legal query. bio
- 36:48and it does it for me most of the time site primary
- 36:51or clearly authoritative sources. Secondary sources
- 36:55flag any uncertainty. I wanted to tell me I'm
- 36:58not too sure about this so that I can be careful
- 37:01and so on. This is just a personalization bio
- 37:05that I came up with but I would encourage you
- 37:09to make up your own or look online for another
- 37:11one or use this one. It's a good start. it just
- 37:14makes AI more accurate if you tell it in your
- 37:17bio why you're using it. So again, competence
- 37:23and confidentiality. I've finished competence
- 37:27now and I'm going to go into confidentiality
- 37:28because I know there's a lot of people worried
- 37:30that we're breaching some sort of attorney -client
- 37:32privilege or something like that by using AI
- 37:35and uploading documents, but I have my own thoughts
- 37:37on that and I was going to share them. But before
- 37:40I move on from competence and into confidentiality,
- 37:43Phil, are there any questions that you have before
- 37:45we move on? No, getting a lot of good stuff here
- 37:48and a lot of gems here. Thank you. Okay. So confidentiality.
- 37:59The issue of lawyers prompts into AI has been
- 38:06explored by courts. Okay. So these are two really
- 38:10interesting cases. They're basically the same
- 38:12issue. One's called Concord Music Group against
- 38:16Anthropic PBC. Anthropic is Claude, the AI model
- 38:21of Claude. And then Trombley versus OpenAI, they're
- 38:25both federal district court cases. OpenAI is
- 38:29ChatGBT. Basically in both cases, the issue was...
- 38:36Creators of copyrighted materials creator of
- 38:39arts creator of music of books fiction non -fiction.
- 38:43We're suing ai programs. Because they were saying
- 38:48ai had been trained on their material without
- 38:50permission they're copying their material without
- 38:52permission. They're regurgitating so if you went
- 38:55on a and you said you know give me a summary
- 38:57of moby dick it will do that and so the copyright.
- 39:02Owners are saying that a is violating copyright
- 39:04and they're not paying. just compensation for
- 39:08their use. I don't know what the merits of that
- 39:09are, I have no idea. I think it's gonna be something
- 39:11that's gonna continue to be explored. But the
- 39:14point is, that's what the lawsuits were about.
- 39:16And in response to these lawsuits, the AI companies
- 39:22during discovery asked for the prompts and the
- 39:28responses from the lawyers bringing the case.
- 39:33They said, the AI sources, the AI companies said,
- 39:37if you're suing us and saying that we're violating
- 39:41copyright law, we have a right to see what you
- 39:45fed AI, what you prompted it, and what the responses
- 39:48were, and you're back and forth. We have a right
- 39:50to see that because you're suing us for copyright
- 39:53law. And they asked for it in discovery. And
- 39:55so what they wanted was the thing that we're
- 39:57all concerned about is if we input data into
- 40:01AI, Can someone subpoena that someday or request
- 40:04for production of documents and get our work
- 40:07product? The ruling in both these cases was the
- 40:10prompts and responses were work product and thus
- 40:13non discoverable. I've looked everywhere for
- 40:17any other decision on this issue. These are the
- 40:19only two I found. But the point is the court
- 40:23in these cases have said that our prompts and
- 40:25the responses is work product, which makes a
- 40:28lot of sense when you look at the definition
- 40:30of work product. There's absolute protection
- 40:33for work product. It applies to any writing that
- 40:36reflects an attorney's impressions, conclusions,
- 40:38opinions, or legal research. The court found
- 40:41there was no waiver of attorney work product,
- 40:44that because the attorneys were selecting the
- 40:47reasons they were asking questions, they were
- 40:49putting their thoughts into asking questions.
- 40:52The district court said it's attorney work product
- 40:55and has absolute privilege and it cannot be discovered.
- 40:57These are the only rulings right now on this
- 41:01issue. If you look at the definition of work
- 41:05product, it protects materials prepared by an
- 41:07attorney in anticipation of litigation from being
- 41:11discovered by the opposing side. Its purpose
- 41:14is to ensure attorneys can prepare their cases
- 41:16thoroughly in privacy and to prevent an attorney
- 41:19from taking unfair advantage of their adversary's
- 41:22efforts. The use of AI and prompting and the
- 41:27responses in the back and forth we have seems
- 41:29to fit the definition of work product privilege
- 41:31precisely in my opinion, and I feel comfortable
- 41:35that it's protected by work product for my own
- 41:37personal use. You have to make your own judgment,
- 41:38of course, but this is how I feel. Now, if you're
- 41:44worried about what can be discovered through
- 41:48your prompts in AI. I have foolproof ways to
- 41:51make you feel very comfortable from now on. If
- 41:54you'll listen to me for a few moments, I think
- 41:56you'll feel even better if you're still worried
- 41:57about attorney work product privilege being waived
- 42:00or something like that. I'm going to show you
- 42:01how to protect it even further. It just takes
- 42:04a little bit of knowledge about what's going
- 42:05on. When you prompt and there's a response, it
- 42:08exists initially in two places. One is on the
- 42:11AI server's end. That's that giant facility there
- 42:13with all these cooling equipment and the big
- 42:18stacks of computers. These are popping up all
- 42:21over America and in other places. And this is
- 42:25what they use a lot of power. And this is where
- 42:28your prompts and responses are going initially.
- 42:31And they're also going to your own personal cloud
- 42:33and account. Okay. So in that, just know that
- 42:40your account, you can log on at any computer
- 42:42with ChatGPT or Gemini or Grok or whatever one
- 42:44you're using. and your answers exist on your
- 42:47cloud and on the AI server's end, two places.
- 42:53There's a way to make sure that it doesn't end
- 42:57up on your server, okay? Remember I showed you
- 43:01there's two places, one is your computer and
- 43:03your cloud and the other one is AI's end? There's
- 43:07a way to make sure it doesn't end up on your
- 43:09end. and they all have this feature in one form
- 43:13or another. What you're looking at is Chat GPT,
- 43:15but there's something called temporary chat.
- 43:19You see that in the top right corner? Temporary
- 43:23chat. The screen will go black. It even says,
- 43:29this chat won't appear in history, use or update
- 43:32Chat GPT's memory or be used to train our models.
- 43:35Meaning they're not using this whatsoever, they're
- 43:37not retaining it. For safety purposes, ChatGBT
- 43:41may keep a copy of the chat for 30 days and then
- 43:43it's deleted. So even if it was an attorney work
- 43:48product, it doesn't exist anywhere after 30 days.
- 43:52It exists nowhere after 30 days if you turn on
- 43:55temporary chat. So that means it doesn't exist
- 44:02here in this big factory after 30 days and it
- 44:05doesn't exist anywhere on your cloud or your
- 44:07computer either. It even says it in AI's way
- 44:18back in like the user notes. Look, temporary
- 44:21chats won't appear in your history and chat GBT
- 44:25won't remember anything you talk about. For safety
- 44:28purposes, we may still keep a copy for 30 days.
- 44:31There was a lawsuit by the New York Times recently
- 44:35in 2025 against AI programs for the same thing,
- 44:38copyright infringement and things of that nature.
- 44:41And the court required AI to keep a copy of all
- 44:44the prompts for a period of days, but that's
- 44:46been lifted as of October 22nd, 2025. So the
- 44:49point is, after 30 days, AI is deleting your
- 44:53chats if you put it on temporary chat. So I feel
- 44:56comfortable that you're not waiving anything
- 44:59by using these services. And that's just my opinion.
- 45:03Obviously, it's not the opinion of anyone else
- 45:05or any other organization. But they all have
- 45:09a similar policy. Google's Gemini Here, if history
- 45:14is off, aka temporary chat, convos are kept only
- 45:19about 72 hours by Gemini and then removed. Grok,
- 45:24private chat mode it's called, it deletes it
- 45:26from their systems in 30 days. Anthropic Claude,
- 45:34it's 30 days if you have that temporary chat
- 45:38turned on. There's another way to make it even
- 45:41more private. You can take off, this is on chat
- 45:43GPT, but they all have this option, improve the
- 45:46model for everyone, turn it off. That means AI
- 45:50is not using your prompts and responses to train
- 45:52other AIs. There is some subjective analysis
- 45:55that goes into attorney work product privilege
- 45:56and attorney client privilege. And so if you
- 45:59turn that off, you can make the argument I didn't,
- 46:02I wanted to intend it to be private. So I turned
- 46:04this improve the model for everyone else off.
- 46:09You can also turn off save memory so it's not
- 46:11saving any of your really important chats. That's
- 46:14another way to make it even more private. So
- 46:23again, a little knowledge of how it works is
- 46:26really helpful when you're thinking about privilege
- 46:29and confidentiality. One, it's not going to exist
- 46:32on this server in 30 days if you turn on those
- 46:36private chats. And at the bottom here, It's not
- 46:39going to exist there ever if you turn on those
- 46:41private chats. So there's no way to get it back
- 46:45if you use those temporary chats. You can put
- 46:48your own work product in it. You can save it
- 46:49on your computer. You can make it part of your
- 46:51practice. And I feel pretty confident that in
- 46:53addition to the work product privilege and the
- 46:56fact that it doesn't exist anywhere, it'd be
- 46:58very difficult for anyone to subpoena or discover
- 47:02your work product. Another thing that I think
- 47:08is important when you're thinking about the ethics
- 47:10of AI, we now put in our retainers and our fee
- 47:15sharing agreements and in our older cases, we
- 47:18send them like a writing for them to sign that
- 47:20we're using AI to help us practice just for some
- 47:24client disclosures. This one says, client acknowledges
- 47:28and consents to the firm's use of secure artificial
- 47:31intelligence tools to assist in case analysis
- 47:34research. drafting and related legal work. By
- 47:38signing this agreement, client authorized the
- 47:40firm to disclose to such tools any case information
- 47:44reasonably necessary for representation, including
- 47:47medical records, personal data, and other confidential
- 47:50materials. By the way, I've heard people say
- 47:53you can't upload medical records because of HIPAA.
- 47:56I disagree with that analysis. I think if you
- 47:58read HIPAA, it only applies to health care. practitioners,
- 48:01it doesn't apply to lawyers, but I'll leave that
- 48:03for your own judgment as well. So remember, turn
- 48:08off, if you want to make sure it's super confidential,
- 48:11turn off store data history so that it could
- 48:14never even be subpoenaed because it's not stored
- 48:16on a chat GPT cloud or other AI cloud. In your
- 48:20personalization, remember I showed you that bio?
- 48:24You can put that all conversations, legal conversations
- 48:27are intended to be worked product privilege,
- 48:30that adds a further subjective layer to your
- 48:32analysis, like that maybe you're intending it
- 48:35to be word product privilege in case that comes
- 48:37up one day. Get permission from your clients
- 48:39to use AI. And when in doubt, you can use key
- 48:42identifiers. This is something that ChatGVT suggested
- 48:45to me. They said if you're really worried, it's
- 48:48a very hot button issue that you're asking AI
- 48:51about. You could change the names of the plaintiff
- 48:54or the defendant or the issue or whatever the
- 48:56address of the situation so that it would be
- 48:59impossible to discover anyways. I do have some
- 49:10products that I really like using. So what I've
- 49:13showed you guys is everything you can use on
- 49:17the general services. Meaning you don't have
- 49:21to go out and buy $100 ,000 or $500 ,000 or $50
- 49:25,000 private AI server or whatever. Everything
- 49:29that I told you can be used by using general
- 49:34AI sources that are, some of them are free and
- 49:36some of them cost like 20 bucks a month. I use
- 49:39ChatGPT, the $200 a month one. I use Gemini.
- 49:44$200 month one grok. I use them all. As you can
- 49:47tell, I'm totally obsessed with this stuff. I
- 49:48find it fascinating. I think you can be better
- 49:51at your job by using it. I think you can go home
- 49:54earlier if you use it properly. I think you can
- 49:56make more money. You can be a better associate,
- 49:58a better partner. You can be better at your job
- 50:00if you use it effectively. And that's the purpose
- 50:02of me wanting to tell people about this. So I
- 50:06think Maybe we could do questions now and then
- 50:10if there's time, I can show people the products
- 50:11I like to use. But as you can see, I just like
- 50:13to use general AI sources to better my practice.
- 50:18Okay, Matt. So it does look like we have nine
- 50:20questions. or so. I'll go ahead and read them.
- 50:24Let's see. So Christine Clemens asks, getting
- 50:26sanctioned for fake references or citations is
- 50:29avoidable. Under what circumstances would an
- 50:31attorney not verify their citations? Even an
- 50:34intern would read the case. So I have a hard
- 50:37time understanding how using AI is scary, scarier
- 50:40than not supervising an intern. I totally agree.
- 50:45And what Christine is saying is it's the same
- 50:50as trusting a support in it to do something really
- 50:53important. You'd always check the support in
- 50:55its work. So if you think of AI as a support
- 50:58in it that you need to check their work, you're
- 51:00doing effectively the same thing. And I totally
- 51:02agree. You have to check anyone's work before
- 51:04you sign something as a lawyer and AI is no different.
- 51:11All right. And then I have a R .H. Lamarandi
- 51:14asking, can you share which AI program that you
- 51:16use to locate the smoking gun? That one was a
- 51:20proprietary one that was on a closed system I
- 51:22don't remember the name right now, but that was
- 51:24before I really knew what I was doing in terms
- 51:28of being able to use chat GBT and Gemini I cannot
- 51:33remember the name And Bernadette is asking are
- 51:39the same five issues the same in legal pay for
- 51:42AI and Westlaw Lexus is vast case, etc So I think
- 51:49the what I've been told that I've used a lot
- 51:52of the AI programs in Westlaw and Lexis, and
- 51:55they seem to be better at avoiding hallucinations
- 52:00because they live in a closed world. And so the
- 52:05the answer is no, I don't think you have as much
- 52:08bias and I don't think it was much hallucination.
- 52:11You do have a token limit issue, input and output.
- 52:14But you're not going to experience as much problems
- 52:17with hallucination and bias. Right, let's see.
- 52:33If you want to retain information for yourself
- 52:35while denying AI access after 30 days, how do
- 52:38you suggest this to be done? Can you repeat that?
- 52:45Yeah, if you want to retain information for yourself
- 52:48while denying AI access after 30 days, how would
- 52:52you suggest this to be done? So I thought a lot
- 52:55about that and I think it's great. I think you
- 52:57have to get the answer and save it on your own
- 52:59computer because the way you keep it off AI server
- 53:03is by putting on the temporary chat, which removes
- 53:06it from your server. So you would answer the
- 53:09question in temporary chat, cut and paste it,
- 53:11export it and put it on your own computer. But
- 53:13it's a great question. Let's see. If you haven't
- 53:20used temporary chats for past research, is there
- 53:22a way to now make private and not saved in the
- 53:25cloud or AI servers? No, but I've heard that
- 53:30I've read that if you delete it off your own
- 53:35cloud, it gets deleted off AI server in some
- 53:37portion of time. So if you wish you didn't Do
- 53:41something in the past and you're worried about?
- 53:44Confidentiality there is that mechanism and I
- 53:46looked at the ethical rules and this is again
- 53:48my personal opinion doesn't represent the opinion
- 53:50of anyone else But I don't think deleting attorney
- 53:54work product protect your clients interest is
- 53:56like it's not the same as deleting evidence because
- 53:58it's attorney work product So the point is I
- 54:00think you could delete it off your chat. That's
- 54:02my personal opinion. No one else's And then save
- 54:06it on your on your own computer to get it off
- 54:09of AI server Okay, and Michael Taggart is asking,
- 54:16when asked about confidentiality, the AI represented
- 54:19that none of the work, let's see. When asked
- 54:23about confidentiality, the AI represented that
- 54:25none of the work generated having instructed
- 54:27the AI that all work is confidential is to be
- 54:30used for educating or training for an AI program.
- 54:33Is that reasonable? Yeah, so I think here I can
- 54:40even show you guys. If I were to try to search
- 54:51up on AI, and I'm speaking to AI right now, you
- 54:54guys, if I were to search up on AI and try to
- 54:57find out what someone else, like my opposing
- 54:59counsel, was searching up on AI so I could gain
- 55:02an advantage, can I do that? Can I find out what
- 55:04they're searching? In short, no, you can't do
- 55:09that. AI tools are designed with privacy in mind,
- 55:12so you don't have access to what another individual
- 55:15is searching or doing. it would be a breach of
- 55:18confidentiality and basically not something that's
- 55:20allowed. So, rest assured what people look up
- 55:24on their end stays private to them. And is that
- 55:27true whether you have it on temporary private
- 55:29mode or just general open mode? Yeah, that's
- 55:32right. Whether you're using something like a
- 55:35temporary mode or just the regular open chat,
- 55:37the principle is the same. Other users activities
- 55:40are private. Basically, these AI platforms are
- 55:43built to keep each user's searches and conversations
- 55:46confidential. So you won't be able to peek into
- 55:48what anyone else is doing. So, you know, that's
- 55:52AI. That's that's a chat to be T app I have that
- 55:55I can just talk to it. And I hope you heard that.
- 55:58But basically it says you can't search up what
- 56:00other people are searching anyway. So this fear
- 56:02that you're somehow putting out your thoughts
- 56:04and attorney work product into the world is just
- 56:07not. accurate based on how AI works and the court's
- 56:11ruling on work product. So I feel comfortable
- 56:13using AI and feeling like it's very well protected.
- 56:19Great. And then I have a few more here. If you
- 56:22put private attorney client communications in
- 56:24the setting, would it be enough even if you put
- 56:29clients' names into the prompt? Look, I think
- 56:34this area is still being tested. I told you guys
- 56:36my opinion that you know There's the court decisions
- 56:38are that it's attorney work product. It seems
- 56:40to fit in there But it's your level of comfort
- 56:43that you have to be comfortable with you're the
- 56:45one that has to make the decision So if it's
- 56:47a obviously a better idea to take the clients
- 56:50names out, but that's that's up to you Have you
- 57:00heard of a situation where a client has ever
- 57:04not consented to AI use in their case? Never.
- 57:08Yeah, that has not happened yet. It could, I
- 57:11suppose, but it's never happened to me. Okay.
- 57:16Let's see. Have you seen a benefit in keeping
- 57:19the work in a product project, as opposed to
- 57:22making it temporary in order to generate follow
- 57:25-up tasks? Yes, absolutely. Doing things in projects
- 57:31is like kind of the next level of way of using
- 57:33AI. I've built agents and stuff too, so I'm super
- 57:37into this. Like I've built agents to draft discovery,
- 57:39build agents to respond to discovery in a meet
- 57:42and confer letter. But I think agents is, if
- 57:46you're really into AI using agents, sorry, project
- 57:49mode really helps because it's got a higher token
- 57:51limit. You can keep the information contained.
- 57:53So if you're working on one single case, make
- 57:56your case a project. It works much better than
- 57:58just using regular AI and every single AI service
- 58:02has the project mode. Alright, so it looks like
- 58:07we have two more questions. Let's see, do you
- 58:11use AI on both mobile and desktop? If so, the
- 58:13same operations as you've talked about today?
- 58:17Yes, I have the same account that's connected
- 58:20to my desktop. It's connected to my computer
- 58:22at home is connected to my phone is connected
- 58:24to my iPad. The same. service and so it spans
- 58:27across so sometimes before i go home at night
- 58:29i'll research a complex issue that i know is
- 58:32going to take 15 or 20 minutes for ai to do it
- 58:34so by the time i get home and i pull into my
- 58:36driveway it's done and i can look at my phone
- 58:39at the work that it created all right and last
- 58:45question do you use ai to research jury verdicts
- 58:48and prepare for settlement negotiations I haven't
- 58:52done that yet. The way that I do use it, let
- 58:54me pull up my screen again so you can see how
- 58:56I use it in my practice, like from a practical
- 58:59perspective. Okay. Let me skip through this.
- 59:10Okay. One of the things that I use in my own
- 59:26practice is something called plod. I have it
- 59:28on the back of my phone right now. It just hooks
- 59:31up there. You press a button. It begins recording
- 59:35your conversation with whoever and it creates
- 59:38an AI summary of your conversation. And so I
- 59:42have done that in replacing taking notes. So
- 59:45if I'm instructing an associate what to do, I
- 59:47turn on the plod. I say, just listen, it's being
- 59:50plotted and I'll give you a summary of the call
- 59:52and it creates amazing summaries of your call.
- 59:56So note taking on a yellow pad. You know, these,
- 1:00:00it's kind of a thing of the past. Using Plot
- 1:00:02is exceptional. I would encourage people to check
- 1:00:05it out. Another program I'm using is called Transcript
- 1:00:17Genius by Steno. I use these little tools to
- 1:00:21help me in addition to just using the regular
- 1:00:23AI services. Steno is really good. They're a
- 1:00:26core reporting company. If you use Steno for
- 1:00:28any core reporting, they have something called
- 1:00:29Transcript Genius. And one of them is to create
- 1:00:32a full length transcript summary with page and
- 1:00:35line numbers. So every trial that I have, I do
- 1:00:39page and line number summaries. And this program
- 1:00:42does it for you in a very accurate way. So this
- 1:00:45is one I did for a case that's settled. It does
- 1:00:48a page and line number summary for you. It's
- 1:00:50really, really good. Um, and so another thing
- 1:00:54I do for trial is for jury selection, I'll upload
- 1:00:59plaintiff's trial brief, defendant's trial brief,
- 1:01:01just to give people the AI a feel for the case.
- 1:01:04And then I ask it to be a potential juror on
- 1:01:07the case and what issues they would like to discuss.
- 1:01:09So I can kind of practice asking voir dire questions.
- 1:01:13You know, what do you think about. School districts
- 1:01:16being sued for injuries or what do you think
- 1:01:17about pain suffering damages or. Whatever your
- 1:01:21issue is and it will respond like a juror and
- 1:01:24kind of help you talk about the issues in the
- 1:01:27case. So i think i think i'm good i'm happy to
- 1:01:41answer any questions that anyone has. I'm very
- 1:01:43passionate about this stuff, so whatever I can
- 1:01:45help with, I'm here. All right. It actually looks
- 1:01:49like your discussion of Plaid started a question.
- 1:01:54Sandra asked, Plaid is great, though I wonder
- 1:01:56if it makes client or prospective client or witness
- 1:01:59less open and comfortable. Obviously, you have
- 1:02:01to tell them it's being recorded. What are your
- 1:02:03thoughts? I totally agree. Sometimes it does
- 1:02:06that, and so that is a very good point. I've
- 1:02:09had a witness recently like that. Everyone changes
- 1:02:12sometimes when they think it's being recorded.
- 1:02:14And so, yeah, you've got to keep that in mind
- 1:02:15for sure. Brady, and I guess Lynn is asking if
- 1:02:26you use a recording device during mediation.
- 1:02:31Look, if you have permission from the person
- 1:02:34you're recording, I think it's a great idea.
- 1:02:37because it can keep track of even like an eight
- 1:02:39hour a day, it can keep track. So say your mediation
- 1:02:42for four hours, it'll keep track of the whole
- 1:02:44mediation and give you a summary of what went
- 1:02:45on. And it's like a two page summary. That's
- 1:02:47probably better than the notes. It'll track the
- 1:02:50offers, the acceptance, sorry, the offers, the
- 1:02:53responses, a bracket. It's very good. Perfect.
- 1:03:01All right, so That's it for questions. A very
- 1:03:04active audience today. This concludes our presentation.
- 1:03:07Thank you so much, Matt, for taking the time
- 1:03:09to educate our audience. And thank you so much
- 1:03:12to our members and the audience for joining us
- 1:03:15this morning. Have a great day, everybody. Thank
- 1:03:17you. Is there anywhere I could put my phone number
- 1:03:19and email so that if anyone has further questions,
- 1:03:21they could just reach out to me? Yeah, if you'd
- 1:03:23like, we can put it in the chat. OK. Let's see.
- 1:03:27Let's see. Maybe you can do it for me if that's
- 1:03:45okay. I can't seem to open it on my computer.
- 1:03:52Yeah, pulling it up right now. Thanks. Well,
- 1:04:03I appreciate everyone taking the time and I hope
- 1:04:06you become more successful with the use of AI.
- 1:04:14This was tremendously informative, Matt. Thank
- 1:04:17you very much. Thank you.