Latest / CLEs You Actually Want to Hear / AI Chats Aren’t Privileged: What California Lawyers Need to Do Now
Transcript
- 0:03Hello, everyone. Welcome to AI Chats on Privileged,
- 0:07What California's Lawyers Need to Do Now. I'm
- 0:11Max Goosen, president of the HBA barrister section,
- 0:15and I'm delighted to introduce our speaker, Usama
- 0:17Koff. Usama is a partner in Fisher Phillips Irvine
- 0:21office and a certified information privacy professional.
- 0:26Usama co -chairs the firm's AI, data, and analytics
- 0:30practice group. as well as the privacy and cyber
- 0:33practice group. His primary areas of practice
- 0:36are workplace privacy and data security, employee
- 0:40defections, unfair competition, trade secret
- 0:44theft, and corporate espionage and artificial
- 0:47intelligence. As co -leader of the firm's consumer
- 0:51privacy team, Usama advises clients on compliance
- 0:54with state consumer privacy laws, including the
- 0:57California Consumer Privacy Act, and the California
- 1:00Invasion of Privacy Act. He regularly conducts
- 1:04trainings and publishes articles relating to
- 1:06consumer and employee privacy and data security.
- 1:09All right, Usama, turning it over to you. Thank
- 1:12you, everyone, and thank you to the Beverly Hills
- 1:16Bar Association for hosting this CLE. I'm Usama
- 1:21from Fisher Phillips and I'm thrilled, maybe
- 1:23slightly terrified to be here to talk about AI
- 1:28confidentiality and the attorney client privilege.
- 1:32If you've ever worried that your client might
- 1:35accidentally waive privilege by CCing the wrong
- 1:38person, imagine what happens when they CC chat
- 1:42GPT. So today's session, AI chats aren't privileged,
- 1:48what California lawyers need to know is not really
- 1:50meant to terrify you, though that may be an unavoidable
- 1:54side effect for which I apologize. It is really
- 1:57to help all of us understand how these new tools
- 2:01intersect with our ethical duties and how to
- 2:04use them safely. If you have questions at any
- 2:08time, please put it into the Q &A. I will do
- 2:11my best to monitor those questions and I may
- 2:15stop and answer them live while as we're talking
- 2:18or towards the end when we have a little time
- 2:22for Q &A. The AI is everywhere. Our browsers,
- 2:32email, legal research tools that we're using,
- 2:36we as lawyers use it to check tone in our messages,
- 2:41to sometimes draft. messages to clients, to internal
- 2:47messages, external messages, draft any kind of
- 2:50communication, to draft discovery responses,
- 2:54to summarize discovery responses, to summarize
- 2:57transcripts. There are now even new tools that
- 3:00could be used at depositions to monitor what
- 3:04is happening and identify inconsistencies in
- 3:08what the witness just talked about. from a prior
- 3:13deposition or in the same deposition. There are
- 3:17a lot of other uses that are all over the legal
- 3:22industry, every facet, not just litigation, but
- 3:25transactional work as well. But when you paste
- 3:31part of a client email or a draft brief into...
- 3:37an AI tool, you may have just shared it with
- 3:42a third party and that's that therein lies the
- 3:45issue. And that third party does not owe confidentiality
- 3:48and any obligation of confidentiality. The single
- 3:52act of sharing the information with a third party
- 3:56can waive privilege, expose work product, and
- 3:59potentially violate the rules of professional
- 4:02conduct. Some of the things we're going to talk
- 4:05about today include guidance that has already
- 4:07been put out by Committee of the State Bar, as
- 4:10well as by the ABA and other associates. related
- 4:16to a lawyer's ethical duties when it comes to
- 4:19using AI. But I think most importantly is that
- 4:22we all should familiarize itself with the technology
- 4:24itself at the end of the day. Use it, test it,
- 4:28understand how it works. Otherwise, you probably
- 4:34have to consult an expert on how to use it before
- 4:37you use it. But generally, it's a good practice
- 4:39for each of us as lawyers to personally use these
- 4:43tools to really... understand what are the limitations,
- 4:45what are the risks, and how to not use it and
- 4:49how to use it most productively. Today, we are
- 4:57going to cover why AI interactions generally
- 5:00fall outside the attorney -client privilege.
- 5:03We're going to cover confidentiality and work
- 5:06product risks under California law. We'll talk
- 5:09about how to safeguard your practice and your
- 5:12bar card. while still using AI productively.
- 5:16In other words, how to avoid being the next Mada
- 5:20versus Avianca headline. That's a case where
- 5:24a lawyer used an AI tool and resulted in coming
- 5:28up with some very fake citations. Now, this presentation
- 5:33is not about hallucinations in chat GPT and other
- 5:37gen AI tools. That's a whole nother topic that
- 5:41we can delve into. However, as a result of some
- 5:44of the cases involving AI hallucinations where
- 5:46lawyers have put fake citations into brief submitted
- 5:51with the court is in some of those there was
- 5:55the potential or, you know, commentators have
- 5:58made comments about how there's always the potential
- 6:00for disclosing information and in some of these
- 6:03cases if you dive down it's quite possible that
- 6:07in coming up with some of the analysis, some
- 6:09of the information, some of the writing, that
- 6:11information was shared or disclosed with that
- 6:15third party about a client. So client confidence
- 6:19could have been disclosed in some of these cases
- 6:22involving AI hallucinations. But separate issue
- 6:27is the lawyer's duty to make sure that you're
- 6:31not putting fake citations into the record. That's
- 6:34a whole nother reason. related to building enough
- 6:38safeguards in the use of gen AI tools and to
- 6:41ensure that those tools are used appropriately
- 6:44for the purpose for which they are built and
- 6:46tested. And there are now sophisticated gen AI
- 6:52type tools that are trained specifically on actual.
- 6:56that exists and they only look within that closed
- 6:58environment, as opposed to just giving you what
- 7:01you want and giving you the answer that you're
- 7:03looking for. And for example, co -counsel is
- 7:07a good example of that, which can be within a
- 7:11closed environment and it's looking at only at
- 7:16Westlaw or whatever search tool you connect it
- 7:20to so that everything is limited and it is not
- 7:23creating a new law. But you always have an obligation
- 7:25to test it. I don't want to dive down too much
- 7:28into that, but that's one of the safeguards you
- 7:32should think about as we go through the presentation.
- 7:37All right. And I see we are already getting some
- 7:39questions, which we are going to hopefully answer
- 7:43on some of these topics that we'll cover. But
- 7:46if they're not answered, I'll certainly go back
- 7:48to them and start talking to them. But let's
- 7:50set the stage first. The first place to start
- 7:54is the model rules and the California equivalence.
- 7:58And one of the themes out there from the ABA,
- 8:02for example, that has published a formal opinion
- 8:05on this issue is that we don't need a new rule
- 8:09to address use of AI by lawyers. Existing rules
- 8:14and existing duties and guidance is plenty sufficient.
- 8:19to address all these issues if we really understand
- 8:22them and talk about them and apply them to this
- 8:25new technology. So there's no need for establishing
- 8:30new law and new rules for lawyers because the
- 8:33rules are already there. And what are the rules?
- 8:35Number one, competence. It includes technological
- 8:39competence. And that goes back to the issue that
- 8:42I mentioned earlier that each of us as lawyers
- 8:46need to have a the technological understanding
- 8:49of the tools that we are using in the practice
- 8:52of law. I would draw the analogy to eDiscovery
- 8:56tools with guidance out there in California that
- 9:02says lawyers need to either know how eDiscovery
- 9:07tools work when they use them, or they need to
- 9:10engage an expert who knows eDiscovery and will
- 9:13be able to identify the right tool to use and
- 9:17the technological process for culling through
- 9:20data and identifying what is relevant and what
- 9:23isn't. But that's the analogy I would draw is
- 9:26that same technical competence duty applies to
- 9:30AI. You can't just be the average user. You have
- 9:37to be a lawyer user, which means you got to know
- 9:39it. You have an actual duty to know how it works.
- 9:43And if you don't know how it works, then get
- 9:45some technical training, obtain training. And
- 9:49in fact, some of these tools will tell you how
- 9:53they work, right? Ask ChatGPT how best to use
- 9:56ChatGPT, ask ChatGPT how it is built. It'll answer
- 10:00the questions for you, right? So you can understand
- 10:05how the tool you're using actually operates if
- 10:08you ask the right questions. But beyond that,
- 10:11you can obtain some specific training on that
- 10:14issue as well. The second is confidentiality,
- 10:17that's the big one. And this is where there's
- 10:21always a risk of uploading client information
- 10:23to public AI. It means you're now disclosing
- 10:29it to a third party. So anytime you are... uploading
- 10:35information, you can scrub it once you've scrubbed
- 10:38it in a way that doesn't identify that potentially,
- 10:41but may not necessarily be the right answer every
- 10:43time, depending on the information that is still
- 10:46in the document. But that's just one way to minimize
- 10:49the risk. But really, one of the things that
- 10:53I would recommend would be engaging tools on
- 10:56an enterprise level, which makes sure that It
- 11:02is a closed system, a closed environment, where
- 11:05the tool is not being trained on your client's
- 11:08data. The model is not being trained on your
- 11:10client's data. And the AI developer and provider
- 11:14is not sharing and receiving that data and using
- 11:17it to train their model. And perhaps the data
- 11:21doesn't even leave your system. If it's a license
- 11:25to use it within your environment, then the data
- 11:29doesn't really leave your system in the same
- 11:31way that you put your discovery documents into
- 11:36an e -discovery relativity instance. You're not
- 11:41disclosing it to a third party. The third party
- 11:43may be hosting it for you and providing the infrastructure
- 11:46for it, but it's not a disclosure that would
- 11:48result in the waiver. Similarly, AI tools can
- 11:52be engaged and integrated into your IT infrastructure
- 11:58in the same manner as eDiscovery tools. The next
- 12:04one is confidentiality of information as well
- 12:10as communications with a client that's 1 .4,
- 12:141 .6. 5 .3 is responsibilities regarding non
- 12:20-lawyer assistance. So making sure that AI vendors
- 12:26are properly supervised, right? An AI vendor
- 12:31counts as a non -lawyer assistant. We're essentially
- 12:34as lawyers responsible for our assistance. you
- 12:38know, actions, right? When they pertain to the
- 12:44work that we're putting, you know, in front of
- 12:46the work product we're putting in a public filing
- 12:49or our responsibilities and communications with
- 12:53our clients. So just like you're responsible
- 12:56for your assistant or your paralegal, you're
- 12:59responsible for what... the output of the AI
- 13:03tool and how you use it. And so if you think
- 13:06of the AI tool as an assistant and use it as
- 13:08an assistant, that could really bring home the
- 13:15idea that you're still responsible for that at
- 13:18the end of the day. The ABA business law today
- 13:25put out an article the stresses that competence
- 13:29means knowing where your data goes. The ABA also
- 13:34put out another report that warns that even AI
- 13:38note takers can capture client competences. So
- 13:42that brings up another issue, which is AI note
- 13:45takers that automatically pop up when you join
- 13:47a Zoom or a Teams. And I've always been wary
- 13:51of them. And so now what I do and I recommend
- 13:55lawyers do is you know, your client joins a call
- 13:58and there's a AI tool that automatically is there,
- 14:02immediately kick it out of the meeting. because
- 14:05you don't know whether that tool is really using
- 14:10the data to train the model. You don't know if
- 14:12the information is going to end up in the hands
- 14:15of a third party. There's always that risk, unless
- 14:18if you have properly engaged the tool, and again,
- 14:21it's a closed system, and all of the safeguards
- 14:25have been built into that process. But your standard
- 14:29commercially available tool out there that just
- 14:32joins for you automatically and starts taking
- 14:36notes and summarizing what everyone is saying,
- 14:39there is risk in that. And so I would heed that
- 14:44warning to maintain the confidentiality by not
- 14:46relying on them, which in some cases, this is
- 14:52a challenge for some clients. They think that...
- 14:55Well, an AI note taker can result in efficiencies.
- 15:01Now you have someone who's taking notes for you
- 15:04that's very accurate, that's tested. You lose
- 15:07that because now you're relying on your own note
- 15:10taking as a lawyer. But that's what we're...
- 15:13we're paid to do. We're good at taking notes
- 15:16and ensuring that we don't miss the details.
- 15:19But sometimes, depending on the circumstances,
- 15:21you may need to have someone else, an actual
- 15:24human with you to help you take notes and maintain
- 15:27the information confidentially. But every circumstance
- 15:30obviously is different. But AI note takers can
- 15:34certainly result in some risk to be concerned
- 15:38about. And this is the. ABA formal opinion 512
- 15:50that was issued July 29, 2024. And it really
- 15:56comes down to three things, they say. Lawyers
- 15:59remain fully accountable for all work product,
- 16:01regardless of how it is generated. The existing
- 16:05rules of professional conduct are sufficient
- 16:07to govern AI use in the legal practice. And AI
- 16:10is here to stay, it is not going away. So we
- 16:13all better jump on. on board and understand it
- 16:17as lawyers. Obviously, they delved into a lot
- 16:19more, but if you were to summarize the takeaways,
- 16:23I think these are the three key takeaways from
- 16:26that opinion. All right, attorney -client privilege
- 16:30and work product at a glance. Just a quick summary.
- 16:34It's a review for all of you, I'm sure, but just
- 16:37so we're setting the stage for the next discussion
- 16:40here. Privilege protects confidential communications
- 16:45for the purpose of legal advice between an attorney
- 16:49and the client. And so it protects communications
- 16:53and the contents of those communications, including
- 16:56if the communication is between non -lawyers
- 17:02at a company, for example, forwarding a lawyer's
- 17:06communication. That's still attorney -client
- 17:07communication now being discussed internally
- 17:10between people discussing that legal advice.
- 17:13that could also extend to that communication
- 17:18as well. We're still talking about communications
- 17:21that are privileged because they pertain to either
- 17:25requesting legal advice, providing legal advice,
- 17:28gathering information for the purpose of providing
- 17:30or requesting legal advice. Anything within that
- 17:34scope would be covered by the privilege. But
- 17:37the privilege is then waived. when the information
- 17:42or the contents of those communications are shared
- 17:45with a third party. The work product is obviously
- 17:51a little different because work product doesn't
- 17:54cover communications. Although communications
- 17:57may contain work product, and they usually do
- 17:59if they're substantive, The work product doctrine
- 18:03protects attorney impressions, thoughts, strategies,
- 18:06opinions, notes, things that the lawyer takes
- 18:10that they may never share with their client because
- 18:14it's their own process and their own thoughts
- 18:17and notes. These are things that don't necessarily
- 18:20belong to a client in a client file if you never
- 18:22shared them and they're not part of the communication.
- 18:25They are yours as an attorney. They are your
- 18:27attorney work product and you're the one that
- 18:30holds the the privilege as the lawyer to protect
- 18:34that information. On behalf of your client if
- 18:36it contains client information and client confidences,
- 18:39but it's also you're the holder of the privilege
- 18:43as opposed to the attorney client communications
- 18:46where the client is the holder of the privilege
- 18:49and only the client can waive those communications.
- 18:52The work product, if it contains client information
- 18:54as well, the client is the holder of the privilege
- 18:58and you would not be able to wave attorney work
- 19:02product protection if the attorney work product
- 19:05contains client confidences and information that
- 19:09the client is the holder of. There's certainly
- 19:11that overlap here. All right. So another rule
- 19:17to mention and discuss is the evidence code 954
- 19:22regarding the scope of the privilege. A client
- 19:25has the privilege to refuse to disclose and to
- 19:28prevent another from disclosing a confidential
- 19:30communication between a client and a lawyer.
- 19:33And a confidential communication means one that
- 19:36is made in the course of the relationship and
- 19:38in confidence, meaning there has to be an attorney
- 19:40-client relationship between the two or in a
- 19:44potential relationship. Like if it's a prospective
- 19:47client and they are reaching out for potential
- 19:49legal advice to evaluate whether to hire the
- 19:51lawyer or not, that's still a protected communication.
- 19:54As long as it's maintained confidentially in
- 19:57a confidential setting, there's an expectation
- 19:59of privacy in that communication. And there isn't
- 20:01a third party that that is outside of that relationship.
- 20:07The disclosure to a third party like an AI vendor
- 20:10can defeat the confidence requirement. Next is
- 20:20the California Work Product Doctrine under CCP
- 20:232018 .030. It talks about absolute protection
- 20:28versus qualified protection. Qualified protection
- 20:32for attorney work products sometimes can involve
- 20:36your notes about communications with third parties.
- 20:38For example, if you interview witnesses, there
- 20:41are courts that find that there's a qualified
- 20:43work product, a protection for interview notes
- 20:47with third parties, if those notes would reveal
- 20:50your thoughts and strategies and analysis of
- 20:53your client's position in a case based on what
- 20:56a third party witness is telling you, even though
- 20:58you don't have an actual attorney -client relationship
- 21:01with that third party. The absolute protection
- 21:04would cover writings containing attorney impressions,
- 21:07conclusions, opinions or legal research and so
- 21:10on. Again, disclosure to a non -agent waives
- 21:14the protection and allows the opposing counsel
- 21:17to seek discovery of AI chat records. And there's
- 21:22been a couple of key cases that we should talk
- 21:25about. The first is Tremblay versus OpenAI, decided
- 21:30by the Northern District of California in August
- 21:33of 2024. There... the court held that attorney
- 21:41crafted, the court held that attorney crafted
- 21:45prompts to gen AI, like chat GPT, constitute
- 21:49opinion work product rather than fact work product.
- 21:52And that's a key distinction. And as a result,
- 21:55you know, prompts, you know, may be discoverable
- 22:00if they are fact, work product versus opinion
- 22:04prompts. Here, in this case, the prompts contain
- 22:07counsel's mental impression and strategic thinking
- 22:10about how to interrogate AI systems and anticipation
- 22:13of litigation, making them virtually undiscoverable
- 22:17unless opposing parties can show that the attorney's
- 22:24mental oppressions are actually an issue in the
- 22:26case and demonstrate a compelling need for the
- 22:29material. So how do you demonstrate a compelling
- 22:31need for the materials? I guess it depends on
- 22:33every case and whether that is called into question.
- 22:41brings me back to the hallucination cases where
- 22:43a lawyer creates a prompt and based on the prompt,
- 22:48a bunch of fake citations are generated and then
- 22:51they use that. Well, maybe the prompt at that
- 22:53point and whatever the lawyer told the gen AI
- 22:56tool is an issue, right? So they may end up having
- 22:59to potentially disclose to the court. what the
- 23:01prompt was that led to all of these fake citations
- 23:04being generated. So that's one example of where
- 23:07it might be an issue where, yeah, your work product
- 23:10can be discoverable if it becomes an issue. The
- 23:21Tremblay case has been cited and followed. And
- 23:28it's going to lead to more cases that would cite
- 23:32to it. But it's probably the first place that
- 23:34I would start in analyzing this issue if you
- 23:37are facing it in litigation. Another court that
- 23:42has looked at this is Concord Music versus Anthropic.
- 23:47And it's a case decided in May of this year,
- 23:502025. And in that case, court found that certain
- 23:55prompts settings and outputs from Claude AI were
- 23:59protected, rejecting the argument that they were
- 24:02automatically discoverable. The court in this
- 24:05case emphasized that these materials were generated
- 24:08during the investigative process by the lawyer,
- 24:12and therefore they qualified for attorney work
- 24:15product protection. The protection can be waived
- 24:18when a party relies on that material in their
- 24:21pleadings or in their motions. So that's the
- 24:25key that the prompts, the settings, the outputs
- 24:29from a Gen .ai tool, if you rely on them in your
- 24:34pleadings or motions, then they can potentially
- 24:38be discoverable. So how do you not rely on them,
- 24:41right? Well, that's where, use the Gen .ai as
- 24:47an assistant, truly as an assistant. It can be
- 24:50a first run. But you're still responsible for
- 24:53critically thinking through the arguments, the
- 24:56summary, verifying the facts, verifying the citations,
- 25:01and everything before you put it in front of
- 25:03the court so that you never put your interaction
- 25:07with an AI tool at issue. So that's one way to
- 25:11do it, to put a barrier. Implied waiver of attorney
- 25:21client privilege and attorney work product occurs
- 25:25when a party's conduct, such as using an AI platform,
- 25:28places privileged information at issue or makes
- 25:31it relevant to a claim or defense. For example,
- 25:35if an attorney uses an AI tool to process or
- 25:38analyze privileged communications, For example,
- 25:42you have a bunch of communications between you
- 25:44and a client, and you put those communications
- 25:46into the tool. And the tool's design or operation
- 25:49involves sharing the data with the third parties.
- 25:52So maybe that AI tool is sharing data with third
- 25:55parties, or now the information is out in the
- 25:59ether, and it's now in the hands of a third party.
- 26:07It could be construed as a voluntary disclosure.
- 26:09thereby waiving the privilege. And herein lies
- 26:14the danger of using GEN .AI for summarizing client
- 26:19communications. maybe one way to do it is to
- 26:23redact all of the information before you put
- 26:26it in. So that requires tedious manual work potentially
- 26:30to redact everything before you put the information
- 26:35into the GEN .AI tool and that before you ask
- 26:39for the summary. Courts have And there's been
- 26:47courts, obviously before the advent of AI, that
- 26:50have addressed this implied waiver issue. But
- 26:52generally, courts have emphasized that fairness
- 26:55principles guide the application of the implied
- 26:58waiver, particularly when the privilege is used
- 27:00as both a shield and a sword, or when the opposing
- 27:03party is denied access to information that is
- 27:05vital to their case. So there's going to be general
- 27:07fairness guidelines that are addressed here as
- 27:12we deal with this new AI era. Before I talk about
- 27:18the CCPA, I want to mention that word product
- 27:21often includes notes and drafts. If you upload
- 27:24them to an AI tool that stores data off site,
- 27:29then potentially you've disclosed them. And so
- 27:32where the data is stored is very important in
- 27:35understanding what tool you're engaging. If you're
- 27:40a law firm, Just like you're engaging an eDiscovery
- 27:45vendor, for example, know where your data is
- 27:48stored and know who has access to your data.
- 27:53Engage in the proper vetting of that vendor.
- 27:58Bring that same expertise and experience you've
- 28:01developed in engaging and onboarding an eDiscovery
- 28:05vendor and apply it to an AI tool. It's the same
- 28:08thing. It's just another vendor who's going to
- 28:10have access to potentially confidential and sensitive
- 28:13information. So you have an obligation to properly
- 28:16vet that vendor to ensure that they are securing
- 28:20the data, maintaining it safely. They're not
- 28:23disclosing it to third parties and they agree
- 28:26to comply contractually with these obligations.
- 28:30But for AI tools, one of the best ways is to
- 28:32make sure that you're in a closed system where
- 28:34the data isn't going to a third party. The CCPI
- 28:44is just another complicating layer, right? So
- 28:48there's ethical standards, there's ethical rules
- 28:50and rules of professional responsibility. Then
- 28:52there's the CCPI. It doesn't apply to every lawyer.
- 28:56It applies probably to all of the large law firms
- 29:00out there and any law firm that meets the threshold.
- 29:03of being a business operating in California that
- 29:06operates in California that collects information
- 29:09about people that are California residents. So
- 29:12if you meet the threshold, then this consumer
- 29:15privacy law applies to you. So why do I bring
- 29:18it up in this context? It's that a law firm that
- 29:22had, you know, any attorney at any level, share
- 29:30information with a client's information with
- 29:34a GEN .ai tool. And let's say that information
- 29:36happens to be sensitive data, you may have a
- 29:40potential CCPA problem on your hand. And that's
- 29:44because the CCPA requires a covered business
- 29:48to maintain appropriate security measures to
- 29:51ensure that there is no unauthorized disclosure
- 29:55or access to certain sensitive data. Social security
- 30:01numbers, medical information, things along those
- 30:04lines, and they're defined in the laws to what
- 30:07actually is covered in this part of the statute.
- 30:13So it's certain types of sensitive data. If the
- 30:16documents that are being uploaded contain that
- 30:18information, it potentially could be a data breach.
- 30:22right? Not just under the CCPA, but under other
- 30:24data breach laws. But that's very concerning,
- 30:28especially because there is a private right of
- 30:30action under the CCPA that can subject you to
- 30:34liability. And the liability doesn't require
- 30:37proof of harm, although there could be major
- 30:40and significant harm from disclosure of a client's
- 30:43sensitive and confidential information. even
- 30:48if there is no actual harm that is provable under
- 30:51the CCPA, there are statutory damages that are
- 30:54recoverable regardless of harm. And, you know,
- 30:58so that's why this is significant. If you're
- 31:01disclosing information to an Edge NAI tool that,
- 31:05you know, You certainly have to worry about the
- 31:08ethical obligations and potential disclosure
- 31:10of information in violation of those ethical
- 31:14duties. But now you may have a data breach on
- 31:16your hands. So that's the reality. So when you
- 31:22hit enter on any kind of AI tool and submit your
- 31:28prompt, AI systems often log the text for quality
- 31:32control. No training modes, still cached data.
- 31:36and vendors can be subpoenaed. There's rarely
- 31:40a real deletion right. So what you type today
- 31:44can reappear tomorrow. Clients ask, you know,
- 31:48can't we tell the chatbot not to remember? And
- 31:52I'd say sure, and tell your toddler not to repeat
- 31:55what you said in the car. You know, good luck
- 31:57with that. So it's really important that technology,
- 32:03that reality about this technology that whatever
- 32:06you type in, it's there, right? Your ability
- 32:10to remove it. after you've hit submit is virtually
- 32:15none at this point. Now, we may get to a different
- 32:19point down the road where someone addresses that
- 32:23issue and rates safeguards, and now you may have
- 32:27the ability to remove information from the hands
- 32:31of the vendor. But once you've submitted it to
- 32:33the vendor, it's there. It's in their hands.
- 32:37The waiver of the privilege occurs upon disclosure.
- 32:40work product loses protection if shared with
- 32:42a non -agent. There is no common interest or
- 32:46agency relationship with the AI provider, so
- 32:48you can't really claim protection of the information
- 32:51under that issue. Even if off, data may be cached
- 33:01for debugging, like I mentioned, and the third
- 33:04-party vendors can potentially access or be subpoenaed
- 33:07for records. It's really important to understand
- 33:12that you may not have any meaningful right to
- 33:15deletion or control of the information once it
- 33:18has been submitted. There are a lot of common
- 33:23uses of AI chat tools by lawyers. This includes
- 33:28drafting memos. motions, discovery requests,
- 33:31like I mentioned earlier, summarizing depositions
- 33:34or client emails, brainstorming legal arguments
- 33:37or case strategy, checking citation formats or
- 33:41tone polishing, putting in a case saying, hey,
- 33:46can you look up, can you summarize the case?
- 33:51Putting in a case and saying, do I have the right
- 33:54citation for this? Or find other cases like this.
- 33:59And beyond just simply using the other standard
- 34:03research methods that you have through Lexis
- 34:06or Westlaw or other tools that are out there.
- 34:09So there are a number of ways that AI tools can
- 34:15be used by lawyers. Now we have to make sure
- 34:18as lawyers that we are using AI tools responsibly.
- 34:23And so realistic examples of confidential data
- 34:27that lawyers may input include client identifiers
- 34:30and case names, right? So make sure that you're
- 34:32removing identifiers and case names. I even try
- 34:39to ensure that I'm only using my personal email
- 34:42with any chat tool and I'm not connecting anything
- 34:44that I put. with my firm. And that's just to
- 34:49create a separation. So anything I put in there,
- 34:52I say, I'm a privacy lawyer at a big law firm.
- 34:55And I'm going to do XYZ, and I'm thinking of
- 34:59writing a message to so -and -so without naming
- 35:04them, of course, to someone who is in this type
- 35:08of business. Give me a few starting paragraphs
- 35:15or give me a couple of points that I can start
- 35:17and I can work from. So that's just an example
- 35:20of making sure that I'm not using client identifiers,
- 35:24a case name. so that it's not linkable back to
- 35:31a specific client or particular matter that you're
- 35:33working on. Factual summaries of harassment or
- 35:38termination allegations. You put in your witness
- 35:41interviews from a harassment investigation or
- 35:44workplace investigation, and you put it into
- 35:46the AI tool and ask for a summary. Prepare the
- 35:50first report. Here are all 15 witnesses I interviewed.
- 35:54Prepare the report. What should be my findings
- 35:57based on what everyone said? That is going to
- 36:01be ripe with potentially disclosing. privileged
- 36:04information or confidential information that
- 36:06is sensitive in nature, especially based on these
- 36:09kinds of allegations. Draft settlement agreements
- 36:12with dollar figures. Well, a lot of these agreements
- 36:15are going to have confidentiality obligations,
- 36:17right? Not to mention that there are changes
- 36:20and things that are going back and forth with
- 36:24the opposing party. disclosure of these drafts,
- 36:29including work product drafts that haven't been
- 36:31shared with anybody, putting those kind of drafts
- 36:34into the AI tool can risk having the information
- 36:39be no longer protected. Internal work product
- 36:44is another example. Litigation strategy notes
- 36:47that your client doesn't see, but putting that
- 36:50into an AI tool and asking for a summary, asking
- 36:53to organize it. asking to make sense of it, for
- 36:56example. That could be risky unless, of course,
- 37:00you implement the right safeguards. And I've
- 37:03already talked about some of those safeguards,
- 37:05like ensuring that the data isn't going somewhere
- 37:09else, that it's a closed system, and the model
- 37:13isn't being trained using your information or
- 37:16your client's information. Another example here
- 37:20is client email. text that is quoted verbatim
- 37:24for tone suggestions. Well, if it's a verbatim
- 37:28quote, it can be found. It can be identified
- 37:31later and tracked to whoever said it. So that's
- 37:38just another risk that is out there. There is
- 37:44certainly practical exposure when using AI tools.
- 37:48The exposure can be subpoena or discovery risk.
- 37:52Opposing counsel may request AI chat logs as
- 37:56evidence of waiver or work product disclosure.
- 38:00There is a risk of erosion of the privilege.
- 38:03Once disclosed, the privilege cannot be clawed
- 38:06back. Work product loss, conversations may be
- 38:10discoverable as nonprivileged materials. There's
- 38:13also the client trust impact, reputational and
- 38:17malpractice consequences, which could be quite
- 38:22devastating to your practice. In fact, I may
- 38:26think of this one as probably the most important
- 38:29one. We all thrive for client trust. This is
- 38:33a very trust -based industry. You start with
- 38:38having trust with your clients. If you don't
- 38:40have that, you don't have anything. frankly.
- 38:43So I would actually put that at the top as the
- 38:46worst possible scenario. You lose the trust of
- 38:50your client. Other things that can happen are
- 38:54bad, but it's not as bad as you've lost the trust.
- 38:58You lose the trust, then what are you doing as
- 39:02a lawyer? So those are all significant risks
- 39:08to address. How AI undermines work product protection
- 39:13includes uploading drafts. I've seen a question
- 39:17here about uploading drafts or legal strategy
- 39:20that can convert them into shared documents.
- 39:24Of course, treat disclosure to third -party tech
- 39:26providers as an intentional waiver. And this
- 39:29is becoming more of an issue. The protection
- 39:34under the code evaporates when confidentiality
- 39:38is not maintained. The responsible AI use in
- 39:43law practice includes addressing these particular
- 39:48safeguards that I have here on the slide. But
- 39:51first, and probably one of the more important
- 39:52ones, is to use an enterprise -licensed AI tool.
- 39:56For example, ChatGPT Team, Microsoft Copilot
- 40:01with data isolation, as opposed to having a personal,
- 40:12an individual premium version. Even that, if
- 40:18you're a law firm, I wouldn't include, I wouldn't
- 40:20put any confidential client information in it
- 40:24because there's always the risk of disclosure
- 40:26that is a third party you're interacting with.
- 40:30They're not your agent. There's no common interest.
- 40:37You're not signing a contract, hiring them as
- 40:40your service provider and vendor like an e -discovery
- 40:43vendor. They're just an AI tool and AI provider
- 40:47that you're signing up according to whatever
- 40:49terms of use they have or end user license agreement
- 40:52that they have. So if you invest in the enterprise
- 40:56version of all of these tools, that's really...
- 41:01the best way to maintain the privilege and to
- 41:05maintain confidentiality. Obviously never input
- 41:08client identifiers or facts. Treat AI output
- 41:13as non -confidential drafts, right? Because they
- 41:18are non -confidential if it's the output of an
- 41:21AI tool. It is because it could potentially be
- 41:28discoverable. Review all the results. for accuracy
- 41:32and hallucinations and ensure that you're not
- 41:35putting anything out there outside of your firm
- 41:38that hasn't been reviewed by a human. The technology
- 41:43can help us do our job faster, better, cheaper,
- 41:49but it is not a replacement for... the lawyer
- 41:55exercising their own due diligence and their
- 42:00own duty of competence, their duty to supervise
- 42:04more junior attorneys and paralegals and assistants
- 42:07to ensure that you're not putting out work product
- 42:10that is full of even a single, not just talking
- 42:19about full of hallucinations, even one hallucination
- 42:22is bad enough, right? I think my audio cut out
- 42:34there for a second, so I am back. So it is the
- 42:40duty of the lawyer to ensure that everything
- 42:44that you put out that is a public filing, anything
- 42:47that leaves your firm has been reviewed for accuracy,
- 42:52has been checked. Before the advent of AI, It
- 42:58was always your duty to check and cite. That's
- 43:01why the best practice whenever we file a brief
- 43:04is there's always someone who's going to double
- 43:06check these cases to ensure that they're still
- 43:09good law. When you're citing a case, you want
- 43:11to make sure you're citing it for the right proposition
- 43:13and you're not just putting a citation in a brief
- 43:16in front of a judge. something that the case
- 43:20doesn't stand for, right? Or putting in the wrong
- 43:23citations. So there's always someone that had
- 43:25to have been checking these, right? Now you have
- 43:28some technology that can help check those in
- 43:31a more efficient manner. But it's still your
- 43:33duty to ensure that everything has been checked.
- 43:36And every day, we are hearing more and more about
- 43:38these hallucinations in legal briefs. And they
- 43:42really create some significant risk. But for
- 43:44purposes of today's discussion, the risk is not
- 43:47just in creating false, fake case citations,
- 43:53but in disclosing confidential information about
- 43:56clients, as well as work product of the attorney.
- 44:00The last issue here, the last recommended practice
- 44:05for responsible AI use is to disclose AI use
- 44:09to clients when appropriate. There's always been
- 44:13a debate among lawyers ever since AI came on
- 44:18the scene that do you really have to disclose
- 44:23to your clients that you're using AI? You don't
- 44:29have to tell them every little detail of how
- 44:30you do your job of every program. Hey, by the
- 44:34way, we use Microsoft and we use this and we
- 44:36use that. You've never had that level of... specific
- 44:41obligation to disclose the specific tools to
- 44:43them, right? You can go and engage in eDiscovery
- 44:48vendor to engage in forensics and eDiscovery.
- 44:50Did you have to necessarily disclose to a client
- 44:55that, hey, by the way, when you work with us,
- 44:58your data is going to be put into relativity
- 45:04and we have relativity with this vendor or with
- 45:06that provider. I'm not sure what the answer is,
- 45:11because it really depends on where the client
- 45:14is, depends on the ethical obligations, your
- 45:17contract with the particular client. But there's
- 45:23the law and there's what is actually required.
- 45:25And then there is the best practice. Best practice
- 45:28is always transparency. Clients appreciate it.
- 45:36It's important that they know that they can trust
- 45:38you. It's one way to establish trust. Hey, I'm
- 45:42using AI tools to do my job in a more efficient
- 45:47manner so I can serve you better. Do you have
- 45:50any concerns? If you have any concerns, let's
- 45:53run through them. Hey, we've made sure that the
- 45:57AI tools that we are using are in a closed system.
- 46:01The data isn't going somewhere else. So all of
- 46:03these things can be addressed upfront with your
- 46:07clients. The California State Bar has a standing
- 46:13committee on professional responsibility and
- 46:15conduct, and they've come up with guidance for
- 46:18use of GEN .AI. And it covers a lot of different
- 46:22things. And I just wanted to highlight some of
- 46:25the key issues that some of the key guidance
- 46:30that they put in that document. It's a very short
- 46:32document. I recommend that all of you read it
- 46:35because it covers how the existing rules apply
- 46:39in the context of using GenAI. But some of the
- 46:43key things are, you should review the terms of
- 46:45use of your GenAI apps. As lawyers, unless if
- 46:50you're in that kind of field that involves reading
- 46:58and writing terms of use and contracts, if you're
- 47:02a very contract -oriented lawyer, When it comes
- 47:06to our own personal uses of apps, if this was
- 47:12a live audience and I say, who here has ever
- 47:14read the end user agreement for pick an app on
- 47:18your iPhone? Have you read it? I guarantee no
- 47:21one will ever raise their hand. So this should
- 47:25be different for lawyers. So the idea is we need
- 47:28to hold ourselves to a higher standard and we
- 47:30should actually review the terms of use. If you
- 47:33are at a larger firm and you're engaging in AI
- 47:36tool on an enterprise level, then obviously you're
- 47:38going to have a team of people whose job is,
- 47:40or part of their job is going to be vetting that
- 47:42vendor and reading those terms and understanding
- 47:44them, and understanding what are the limitations.
- 47:49What are the guardrails? What are the risks?
- 47:52Have they addressed security, data security?
- 47:54Have they addressed where the data is being stored?
- 47:59Who has access to the data? Training the model
- 48:01is probably one of the most important issues
- 48:04here. So those are all things that you should
- 48:07address when you are reviewing the terms of use
- 48:11of your GEN .AI apps. Number two from this guidance
- 48:16is an emphasis on consulting with technical experts.
- 48:22And I've mentioned this a couple of times already,
- 48:24but I'll reemphasize it here. The State Bar Committee
- 48:29recommends that you engage experts if you don't
- 48:33know. And if you don't understand these tools,
- 48:37that you actually engage experts. The experts
- 48:39can be Another lawyer that's actually dealt with
- 48:42this, an AI expert, an AI lawyer, the expert
- 48:44could be a technical consultant. There are many
- 48:47technical consultants out there that work with
- 48:49law firms to help them with AI governance and
- 48:54dealing with AI tools, how to onboard them properly,
- 48:59how to vet them properly. There are technical
- 49:02experts out there. pick any kind of, you're a
- 49:11personal injury lawyer. Do you even have the
- 49:14time to become your own expert on the tool, right?
- 49:19You should. a consultant expert to advise you
- 49:24on how to implement and how to use this technology
- 49:26properly. And there are consultants out there
- 49:29that focus on this industry. You can also utilize
- 49:32a particular vendor that you are seeking to implement.
- 49:41Like I said earlier, you can get information
- 49:45directly from ChatGPT about how How best to use
- 49:50it? What are the limitations? It'll answer some
- 49:53of the questions. Now, can you trust the answers
- 49:57all the time? They could be hallucinations, so
- 50:00you still need to get some independent review
- 50:04and advice on that. The third recommendation,
- 50:09again, among money, I think they have like over
- 50:1110 different things they addressed. I picked
- 50:14the four that are most relevant for our discussion
- 50:16today. Never input confidential information without
- 50:19proper confidentiality and security protections.
- 50:23So what are those confidentiality and security
- 50:25protections? That's gonna depend on your client's
- 50:28expectations, the type of information at issue,
- 50:30the type of technology that you're using. all
- 50:35these factors. And then the last one is you got
- 50:38to properly supervise junior attorneys. And that
- 50:45really goes without saying, because the last
- 50:47thing, if you're a partner and you have an associate
- 50:50working for you, the last thing you'd want is
- 50:54for a brief to get filed that you didn't lay
- 50:57eyes on properly, you didn't test, and that you
- 51:01didn't verify that your junior associate who
- 51:05prepared it had actually vetted every case. Having
- 51:12proper training is another way for doing this.
- 51:18If you're not a solo practitioner and you have
- 51:21lawyers working under you and paralegals, ensuring
- 51:25that they receive the appropriate training. Now,
- 51:27how do you find this kind of training? Who provides
- 51:31it? Again, go back to the technical expert. There
- 51:34are consulting firms out there that will provide
- 51:36training for your team on how to properly use
- 51:39AI tools and how to understand it. This includes
- 51:43having a policy. So if you don't have a written
- 51:45policy for your team that they can read, they
- 51:48can reference, that explains to them what the
- 51:51rules are, what the guardrails are, that's the
- 51:55first step. A policy, training, and then as well
- 51:59as proper supervision to make sure that you've
- 52:02asked the questions every time, that nothing
- 52:04leaves the door without ever being checked for
- 52:09hallucinations, as well as to ensure that you
- 52:13haven't disclosed confidential information. Excuse
- 52:20me. Privileged means private. AI isn't. Public
- 52:26AI chats means disclosure to a third party. Some
- 52:29of this is a repetition. Use AI only under vetted
- 52:34control systems. Be the lawyer who understands
- 52:37both the law and technology that'll give you
- 52:40an edge over others who either understand one
- 52:44or the other. kind of the takeaways. But there
- 52:50are a lot of questions here. I'm going to get
- 52:51to the questions. Legislators are discussing
- 52:54safe harbors for lawyer -directed AI and treating
- 52:58AI vendors as agents under confidential contracts.
- 53:03Data privacy advocates are also out there trying
- 53:07to address these issues. They want default deletion
- 53:10and subpoena notice rules. But that could be
- 53:13years. before the legislature does anything in
- 53:18that regard. So until then, we got to stay cautious
- 53:23and keep clients from sending midnight chat GPT
- 53:26queries. AI is really an amazing assistant, but
- 53:29a terrible confidant. So treat it like a summer
- 53:32associate. Helpful, fast. But don't trust it
- 53:36unsupervised. If you remember nothing, privilege
- 53:40means private and AI isn't. Now, let me go over
- 53:44some of the questions that are here in the next
- 53:46seven minutes that we have. So first question
- 53:54was, no work product protection even with disguised
- 53:59names. Well, disguised names will disguise the
- 54:02client, but it's still your input as an attorney
- 54:06into this. So you've put your impressions, your
- 54:10thoughts, your analysis into the prompt itself,
- 54:14into the AI tool. You've interacted with it.
- 54:16Even if you're interacting with it without mentioning
- 54:19client identifiers, the work product protection
- 54:23may be gone with respect to the interactions
- 54:26in the event that it's at issue. Right? So we've
- 54:31talked about the Trumbly case where the court
- 54:32said it's virtually undiscoverable. Well, yeah,
- 54:35virtually undiscoverable. But what if now the
- 54:39prompt and what you put in there is an issue?
- 54:42Because it could be in some cases, which is a
- 54:46risk that I wouldn't take with my own work product,
- 54:50even if I'm not including any client identifiers.
- 54:54Next question, how can one make confidential
- 54:56uploads to ChatGPT and Cloud AI? I have the $20
- 55:00subscription to both. The answer is you need
- 55:05to have the enterprise version. So it needs to
- 55:08be a closed system, enterprise version, closed
- 55:11system only. The $20 subscription premium version,
- 55:15yeah, that's great, that's helpful. It means
- 55:17maybe they're not using the information to train
- 55:19their model. There are some additional protections,
- 55:23but at this point, when it comes to attorney
- 55:25work product and attorney client privilege, I
- 55:28still wouldn't trust it at this point. Do you
- 55:33see waiver problems if you publicly upload documents
- 55:36already filed in a case and request responses
- 55:38to those documents? If the information is already
- 55:43publicly available and you're asking CHAT GPT
- 55:47to summarize a publicly available document, you're
- 55:51not disclosing attorney client privilege information,
- 55:58then the only issue is back to attorney work
- 56:01product. Like I talked about earlier, there's
- 56:04always that risk that it could potentially be
- 56:06discoverable if it becomes an issue. There's
- 56:08fairness standards, fairness considerations,
- 56:10things that may come up in those cases that put
- 56:14your work product at issue. So that's the risk.
- 56:16And I don't think it's worth taking. Next question.
- 56:21Does the Plus Membership Chat GPT notice that
- 56:25OpenAI doesn't use their workspace data to train
- 56:29its models, provide adequate privacy for document
- 56:32uploads? No. It means that they're not using
- 56:38the data to train the model, but does this mean
- 56:41that you can give it as a third party your confidential
- 56:45client information without breaking the privilege?
- 56:49No, I think it can still break the privilege
- 56:51by uploading information. Is an AI like an assistant?
- 56:58Well, exactly so, yes. So my communications with
- 57:01an assistant are protected. It's an assistant,
- 57:04but it's a third party. They're not your employee,
- 57:08they're not your agent. But we say as an assistant,
- 57:11not as an employee, as an agent, but use it as
- 57:16an assistant. That is the best use of a GenAI
- 57:18tool is they're someone who can assist you to
- 57:23get started, but not the one who will do the
- 57:27work for you, who will just get you the final
- 57:29and here's... it's something you can actually
- 57:32rely on at the end of the day. It's a place to
- 57:36get started, to facilitate, to help you on the
- 57:39journey, in the process of getting to where you
- 57:42need to be. But that's why I say use them as
- 57:46an assistant. But they're still a third party.
- 57:55It says question for later. If copying private
- 57:58material to ChatGPT can create a privilege waiver
- 58:02problems, then what about firms with a long -term
- 58:06practice of automatically transferring client
- 58:08files to a commercial cloud service for backup?
- 58:12I don't think that creates a privilege issue.
- 58:18When you have a server, whether it's in the cloud
- 58:23or not, when you engage cloud -based services
- 58:26to store data. you're engaging them as a vendor,
- 58:30just like any discovery vendor, that I mentioned
- 58:33earlier, same kind of vetting that you would
- 58:36engage in to ensure proper protocols, cybersecurity
- 58:41protocols are in place to ensure that no one
- 58:43has access to the data. And they just become
- 58:46a storage facility for you. So when you hire
- 58:49someone to store data for you, they are in that
- 58:55capacity, your agent in storage. and maintaining
- 58:57the data. It's as if the data is just in your
- 59:01building, but now it's in someone else's building
- 59:05and you've hired someone to guard it and to safeguard
- 59:08it. So it's different. It's not the same thing
- 59:12as a chat GPT and data being out there with a
- 59:18third party. you're not engaging ChatGPT as a
- 59:23data processor, for example, to use CCPA terms,
- 59:27they're not a service provider. A service provider
- 59:30on the CCPA or a data processor under GDPR or
- 59:34other state privacy laws that use the processor
- 59:37terminology is a vendor that is agreeing contractually
- 59:40to not use the data for purposes outside the
- 59:44scope of the services they're providing, to not
- 59:46sell the data, to not share the data, to... to
- 59:49cooperate with your obligations related to privacy
- 59:51law and a bunch of other restrictions. You're
- 59:54not doing that with the $20 subscription to ChatGPT,
- 59:58but you are doing that with a vendor that is
- 1:00:01a cloud -based hosting service to host your data
- 1:00:05somewhere. You would have a DPA with them. You
- 1:00:09would have that kind of data protection agreement
- 1:00:11with that kind of vendor. So these are two different
- 1:00:13things. We are running out of time and there
- 1:00:16are a lot of questions. So I'm going to kind
- 1:00:21of scroll through and get to at least one or
- 1:00:25two final questions. Here's a question. As AI
- 1:00:33becomes baked into apps like MS Word, an attorney
- 1:00:37forgets to disable it. It is also baked into
- 1:00:40your phone and newer computers. It is now your
- 1:00:43computer fair game. These things happen with
- 1:00:47updates, so it can be challenging to keep up.
- 1:00:51Yeah, I've seen that too, but AI is getting baked
- 1:00:55into existing apps. It doesn't change the fact
- 1:00:57that the... those apps that you're using, like
- 1:01:02Word and Outlook and other things, they're still
- 1:01:05engaged through a service provider type of model.
- 1:01:10And so they're still... segregation of data,
- 1:01:14there's still protection of data, no disclosure,
- 1:01:17it's still like hiring a vendor. So all of that
- 1:01:20technology that is in use within that same closed
- 1:01:22environment and system doesn't necessarily mean
- 1:01:25that you're giving data to a third party. As
- 1:01:27long as it's still a closed system enterprise
- 1:01:29type application, I think you'll be fine. All
- 1:01:35right, wow, there are a lot of questions. All
- 1:01:40right, I'm gonna have to get one final question.
- 1:01:45How is co -counsel safer? Because co -counsel
- 1:01:49would be engaged as an enterprise type of system.
- 1:01:55It's a closed system as well in terms of what
- 1:02:03data it references. So if you're doing research
- 1:02:07using co -counsel, It's been trained on all the
- 1:02:14legal research repositories out there, and it's
- 1:02:20limited to those repositories that you link it
- 1:02:24to. So it's not going to create new things. You
- 1:02:27still have to review it, but it's still a closed
- 1:02:30system. It's still important to not include confidential
- 1:02:39client information, but if you've engaged that
- 1:02:44tool at an enterprise level in a closed system,
- 1:02:49then you have at least mitigated the risk that
- 1:02:52you are disclosing client information when you're
- 1:02:55interacting with it. I appreciate that there's
- 1:02:59a lot of questions, and I'm sorry that I was
- 1:03:00unable to get to all of them. I've certainly
- 1:03:04hit a button here with the number of questions
- 1:03:07that have come in. So it's an important topic.
- 1:03:10I would love to continue the conversation going.
- 1:03:12But thank you all for attending. I'm happy to
- 1:03:17chat with you at any point as well. Feel free
- 1:03:20to reach out. And just as a colleague in the
- 1:03:24bar, I'm happy to share. knowledge and to continue
- 1:03:28to help any lawyer with this issue. Thank you
- 1:03:32all and have a good day.