WEBVTT 00:00.000 --> 00:04.900 So abstinence is not a plan. It's a fantasy. And the only real plan is knowing precisely what to 00:04.900 --> 00:11.000 trust it for and what to never trust it for and building a habit around that. Here are my personal 00:11.000 --> 00:16.000 hard lines and I encourage you to make them yours or write better ones. Never trust AI for an assertion 00:16.000 --> 00:20.800 of a legal conclusion. And never, never let AI be the only eyes that have been on any aspect of 00:20.800 --> 00:25.700 something you file with a court or send to a client. The tool has a real genuine capacity to increase your 00:25.700 --> 00:30.500 productivity. And I am not a doomsayer about it. I use these tools. But you should use it to answer 00:30.500 --> 00:34.800 questions you already know the answer to and to concisely convey facts and points of law you already 00:34.800 --> 00:40.600 fully understand. It is an accelerator for competence you already have. It is not a substitute for 00:40.600 --> 00:45.200 competence you lack. The cleanest way I can put the whole philosophy is this treat every single output 00:45.200 --> 00:49.200 the way you would treat work you handed to a junior associate. You're going to read it, you're going to 00:49.200 --> 00:54.100 check it and you are fully responsible for it the moment it goes out under your name. And here's the car 00:54.100 --> 00:59.200 Larry that does the real work, the part people skip. If you don't understand the underlying law well enough 00:59.200 --> 01:03.800 to tell whether the output is accurate or not, then you had no business handing that task to a junior 01:03.800 --> 01:08.800 associate in the first place. You should have been the one doing it. AI does not change that calculus 01:08.800 --> 01:12.300 even slightly. It just makes it much, much easier to forget that you were always the one who is 01:12.300 --> 01:16.600 supposed to do the checking. The competence has to live in you. The tool just types a little bit 01:16.600 --> 01:21.200 faster. Two documents turn all of this from a nice sentiment into actual practice, your engagement 01:21.200 --> 01:26.500 letter and your firm's AI policy. On the engagement side, remember opinion 512. Boilerplate may 01:26.500 --> 01:31.200 not be enough for confidential inputs. So if AI use on a matter could touch clients' confidences, 01:31.200 --> 01:36.100 you address it specifically and you get real informed consent. Something like the firm may use 01:36.100 --> 01:40.700 generative artificial intelligence tools to assist with drafting research and review in your matter. 01:40.700 --> 01:44.500 We independently verify all work product against original sources before it is used or filed and you 01:44.500 --> 01:49.600 are not billed for time spent correcting AI generated errors. We will not input information that identifies 01:49.600 --> 01:53.700 you or reveals confidences into any tool that does not contractually protect that information. 01:53.700 --> 01:58.000 And where your informed consent is required to do so, we will obtain it first. Remember, 01:58.000 --> 02:02.200 this is just a scaffold, not advice, but it's a starting point. On the firm policy side, 02:02.200 --> 02:06.900 this is where rules 5.1 and 5.3 landed practice, which tools are approved, what may and may not be 02:06.900 --> 02:11.900 put into them, who has to verify what and how AI assisted work gets recorded. A one page policy that 02:11.900 --> 02:17.400 people actually follow beats a 20 page policy nobody reads. So let me give you the actual sequence I'd want 02:17.400 --> 02:22.900 any lawyer using these tools to run because be careful is really not a workflow. These are really six steps 02:22.900 --> 02:28.400 and none of them are exotic. Step one, before you even open the tool, ask whether this is a task you 02:28.400 --> 02:33.500 understand well enough to check the answer. The honest answer is no. If you couldn't tell a right output from 02:33.500 --> 02:40.100 a wrong one, then this is not a task to hand the machine. It's a task to do yourself or to learn first. 02:40.100 --> 02:44.400 That single gate would have prevented most of the sanctions we just talked about. 02:44.400 --> 02:49.000 Step two, handle the confidences before anything goes in. Either use a tool that contractually 02:49.000 --> 02:53.200 protects your inputs or strip the identifying and privileged material out and get consent where 02:53.200 --> 03:00.200 you need it. Decide this before you paste, not after. Step three, use it for what it's good at. 03:00.200 --> 03:05.300 Organizing, drafting, summarizing, rephrasing, getting you unstuck on a blank page and not for 03:05.300 --> 03:11.100 the thing it's bad at, which is telling you what the law is. Let it accelerate the work. Don't let it 03:11.100 --> 03:17.100 make the decisions. Step four, and this is the one everybody skips, pull every single authority to its 03:17.100 --> 03:23.200 primary source. Open the case, open the statute, not the AI summary, not a second AI query, 03:23.200 --> 03:28.400 the actual document. Confirm it exists and then confirm it says what you're about to say it says. 03:28.400 --> 03:33.400 Both tabs. Step five, run the finished product through a real citation checker, 03:33.400 --> 03:39.100 the Lexis or Westlaw tool as a backstop, belt and suspenders, but understand it's the backstop, 03:39.100 --> 03:44.800 not the check itself. Step six, keep a light record of what you did, which tool, what you verified, 03:44.800 --> 03:50.500 what consent you got. Your future self and rule 5.1 will thank you if anyone ever asks. Six steps, 03:50.500 --> 03:54.900 you could teach them to a summer associate in about 10 minutes. The problem was never that this is hard. 03:54.900 --> 04:00.800 The problem is that the tool is so smooth, it tempts you to skip straight from step one to done. And then 04:00.800 --> 04:07.600 the verification step itself. The boring, non-negotiable one we started the hour with. And the one thing I most 04:07.600 --> 04:12.000 want you to walk out with. When the work is done, before it goes anywhere, every authority gets confirmed to 04:12.000 --> 04:18.200 the primary source, not the AI summary, not to a second AI query. We established why that's something 04:18.200 --> 04:23.700 close to worthless. And checking one AI against another, by the way, is not good enough. Track it 04:23.700 --> 04:30.000 down to the actual case, the actual statute, the actual regulation. Opened, read, confirmed to exist, 04:30.000 --> 04:36.100 and confirmed to stand for the precise proposition you're citing it for. And to still be good law. That 04:36.100 --> 04:40.800 second half matters as much as the first because the subtler and increasingly common failure isn't 04:40.800 --> 04:46.100 that the case doesn't exist. It's the real case cited for something it doesn't hold. If you also run 04:46.100 --> 04:50.400 the finished product back through a Lexis or Westlaw citation checker as a backstop, wonderful. But 04:50.400 --> 04:55.500 understand clearly that automated check is the belt and your own eyes on the primary source are the 04:55.500 --> 05:01.200 suspenders. You need to wear both. And one more practical note, and it's a good one for this room 05:01.200 --> 05:06.700 specifically. New Jersey's new requirement takes effect January 1st of 2027 and first applies to 05:06.700 --> 05:12.400 reporting cycles ending December 31st, 2027. Every active New Jersey attorney will have to earn a 05:12.400 --> 05:19.000 technology related CLE credit. One every two years subject to the usual exemptions and inactive status 05:19.000 --> 05:24.200 rules. My honest view, it's smart, but it's a starting point, not an endpoint. One credit every 05:24.200 --> 05:28.800 two years is the floor for a technology that is remaking the practice underneath us in real time. 05:28.800 --> 05:34.800 So treat the requirement as the minimum that it literally is, not as the education that it isn't. 05:34.800 --> 05:40.000 Go past it. This hour for what it's worth is built to count toward it. But don't let a check box be the 05:40.000 --> 05:44.700 reason you stop learning about the most consequential tool to hit this profession in a generation.