Latest / CLEs You Actually Want to Hear / AI and White Collar Crime: Enforcement Trends, AI-Washing Liability, and Building a Compliance Framework
Transcript
- 0:02Welcome everybody here to AI and White Collar
- 0:06Crime Enforcement Trends, AI Washing Liability,
- 0:10and Building a Compliance Framework. I'm Mark
- 0:13Hogue, Chair of the Beverly Hills Bar Association's
- 0:16AI in the Law section, along with my co -chair
- 0:18Amanda Harris. I'm also the founder of hoagelaw
- 0:22.ai, where I service fractional GC to AI -first
- 0:26startups and SaaS companies while spending most
- 0:30of my nights and weekends vibe coding apps and
- 0:32other projects, because of course I do. I am
- 0:36delighted to introduce today's speakers, Jennifer
- 0:38Park, Matthew Yelovich, and Christian Granger
- 0:40from Cleary, Gottlieb, Steen, and Hamilton LLP.
- 0:46So Matthew is a partner at Cleary Gottlieb, where
- 0:48his practice focuses on government and international
- 0:50investigations, defending companies and individuals
- 0:53in high -stakes domestic and international enforcement
- 0:56actions. Before joining Cleary, he served as
- 0:59deputy chief and appellate chief of the criminal
- 1:01division in the U .S. Attorney's Office for the
- 1:03Northern District of California, where he oversaw
- 1:06approximately 100 federal prosecutors. As an
- 1:09assistant U .S. attorney for nearly 10 years,
- 1:11he prosecuted corporate fraud, health care fraud,
- 1:14money laundering, and public corruption matters.
- 1:16He's clerked for the Honorable Rena Regge of
- 1:18the US Supreme Court, excuse me, the US Court
- 1:20of Appeals for the Second Circuit, and has taught
- 1:23at Stanford Law School and UC College of the
- 1:26Law in San Francisco. Jennifer is also a partner
- 1:30at Cleary Gottlieb, where her practice focuses
- 1:32on white collar defense, enforcement actions,
- 1:35crisis management, and complex disputes. She
- 1:38advises global corporations and financial institutions
- 1:41on interactions with regulators, including the
- 1:43DOJ, SEC, CFTC, and various international authorities.
- 1:48She's been recognized by Benchmark Litigation,
- 1:51Chambers USA, and Global Investigations Review,
- 1:54among others, and has published on DOJ compliance
- 1:57guidance updates addressing AI. She serves on
- 2:00the boards of Legal Services New York City and
- 2:03George Mark Children's House. And Christian Granger
- 2:07is a litigation and enforcement associate at
- 2:09Cleary Gottlieb, and prior to joining Cleary
- 2:11Gottlieb, Christian served as a legal scholar
- 2:14with the SEC Enforcement Division in San Francisco
- 2:17and has also extended to the District of Nevada.
- 2:23Christian earned his JD from the University of
- 2:25California, Hastings College of the Law, now
- 2:27UC Law SF, where he served as Senior Production
- 2:30Editor of the UC Law Journal and was named champion
- 2:34of the David E. Snodgrass Moot Court Competition.
- 2:37His recent publications include the Whistleblowing
- 2:39and Focus series, covering developments in whistleblowing
- 2:42laws and AI. And with that, let's get started.
- 2:50Well, thanks very much, Mark, for that generous
- 2:52introduction. Just a pleasure to be with all
- 2:56of you today. Thank you to the Beverly Hills
- 2:58Bar Association for having us. We are going to
- 3:02talk today about artificial intelligence and
- 3:06enforcement actions. In particular, we've seen
- 3:11a real shift over the last few years in terms
- 3:16of how the federal authorities think about AI
- 3:21and we're still seeing a lot of uncertainty as
- 3:25to how that plays out. We're going to walk through
- 3:28where we were, where we are now, and where we
- 3:31expect to be, as well as some practical advice
- 3:35that can help for those of you who are in -house
- 3:38or who are advising corporate clients on compliance
- 3:43programs in this era of uncertainty. Our goal
- 3:48at the end is that you'll walk away with a firmer
- 3:52understanding of both the recent history in terms
- 3:57of executive actions on artificial intelligence
- 4:01and the enforcement landscape and sort of to
- 4:05-do list items for making sure your compliance
- 4:08infrastructure is up to snuff on that front.
- 4:12Next slide, please. So part of the presentation
- 4:18today is going to be walking through these areas.
- 4:22First, recent AI enforcement, and that will cover
- 4:27different examples of how federal authorities
- 4:30have been thinking about artificial intelligence
- 4:32in the enforcement landscape. We'll then talk
- 4:35about the Trump administration's approach to
- 4:38AI enforcement and some of the nuances that have
- 4:41developed there. We'll talk about AI enforcement
- 4:45tools, which is to say tools that federal authorities
- 4:49are actually employing in their cases in bringing
- 4:52enforcement actions and in running investigations.
- 4:56We'll touch on AI and privilege, which has become
- 5:00a recent news item in the last month or two in
- 5:04particular due to some court orders in one in
- 5:08enforcement action and then one in a civil action.
- 5:11And then we'll finish with building a compliance
- 5:14framework and some practical advice on things
- 5:17to make sure you're going through when doing
- 5:20that. If we have time at the end, we're happy
- 5:23to answer any questions. Next slide, please.
- 5:28And I will turn it over to my colleague, Jen.
- 5:30Right. So when we say recent AI enforcement,
- 5:34I just want to preview that. We're going to talk
- 5:35a little bit about what was happening under the
- 5:38prior Biden administration before the Trump administration
- 5:42took over. There are some things that the administration
- 5:45had in common when it comes to AI. And then there
- 5:47are some stark differences between the two. So
- 5:52under the Biden administration, and then continuing
- 5:55somewhat, both the Department of Justice and
- 5:58the Securities and Exchange Commission have focused
- 6:01on AI in three distinct buckets that we're going
- 6:06to look at together. And those buckets are enforcement
- 6:09buckets. So this is ways to think about how AI
- 6:12can be used potentially as part of a crime or
- 6:16as part of a regulatory offense that the SEC
- 6:18might be interested in. And roughly speaking,
- 6:23you can think about them in this way, which is,
- 6:25has the AI been used as sort of the weapon of
- 6:28the crime? Has AI been involved in a lie, right?
- 6:34So has AI enhanced the ability to lie? And lastly,
- 6:40has AI been sort of the object of a crime, meaning
- 6:43have you tried to infiltrate the AI technology?
- 6:47Just to drill down a little bit on that, when
- 6:50AI is the instrumentality of the fraud, what
- 6:53we're really talking about is hyperscaling crime
- 6:57using AI. And you'll hear the term hyperscaling
- 7:00used by folks within the Department of Justice
- 7:02and the SEC. So their view is that These are
- 7:06crimes we've seen before, they're fraud, they're
- 7:09price fixing, they're market manipulation, but
- 7:11AI allows you to hyperscale them. Think about
- 7:14your standard fraud against the elderly, for
- 7:17example, these sort of call up the elderly and
- 7:19try to get that money. Well, with AI, you can
- 7:22do that thousands or millions of times a day.
- 7:25Second category is the lie category. And this
- 7:28is what is referred to as AI washing. And this
- 7:31is where individuals or companies tell you something
- 7:34about their AI products or services that are
- 7:37powered by AI to attract customers, to attract
- 7:41investors. And that's just not true. And the
- 7:45last category is AI is the object of fraud, where,
- 7:48as I mentioned, AI is the target. So you're trying
- 7:50to steal someone's AI trade secrets or AI patented
- 7:54technology. Typically, these cases have a national
- 7:57security dimension. And we'll talk a little bit
- 7:59more about the current administration's focus
- 8:02on the national security aspects of artificial
- 8:05intelligence. We can go to the next slide. Thank
- 8:08you. When we talk about AI as the instrumentality
- 8:14of fraud, I want to just focus a little bit on
- 8:16some statements that were made under the Biden
- 8:18administration by the Department of Justice.
- 8:21Lisa Monaco, who is the prior deputy attorney
- 8:24general, made a number of public statements over
- 8:27the course of 2024 about how the department was
- 8:29thinking about AI in the context of both white
- 8:33collar and non -white collar crime. And two important
- 8:35things come out of that. One is that she said,
- 8:39that if you're going to use AI as an instrumentality
- 8:43of a crime, the government may seek to enhance
- 8:46penalties against you as a result. She even compared
- 8:50AI to a firearm, for example. So the Department
- 8:53of Justice can seek higher penalties when a firearm
- 8:56is used in a crime. And she said, AI is just
- 8:58like that. Use AI. We may decide to seek higher
- 9:02penalties because it poses a serious public risk.
- 9:06The second thing she said is that If a company
- 9:10wants to get credit from the Department of Justice
- 9:12for cooperating with it, and or credit for having
- 9:15a robust compliance program, both are which are
- 9:18really important factors when the Department
- 9:21of Justice is considering whether to charge a
- 9:23company, how to charge a company, or how to resolve
- 9:26with a company, they're gonna ask how your compliance
- 9:30program thinks about AI and addresses the risks
- 9:33of AI. So two ways in which AI comes up, in the
- 9:38context of AI being the instrumentality of a
- 9:41fraud. So let's talk about an example. What does
- 9:43this look like actually in practice? And we can
- 9:45go to the next slide to do that. There you go,
- 9:49United States versus Smith, super interesting
- 9:52case here. This is from September of 2024. I
- 9:55will say that this status is that this person
- 9:57has been indicted and that hasn't gone any further
- 10:00yet. So no trial yet. So we'll see how this turns
- 10:02out. But the allegations are as follows, that
- 10:05Michael Smith decided that he was going to concoct
- 10:09a scheme to make money from music royalties.
- 10:12So he was going to both use artificial intelligence
- 10:16to create millions of songs, each of which had
- 10:21slightly different names on them. And I'll come
- 10:23back to why he did that. And he was also going
- 10:27to use AI driven bots. to listen to those songs
- 10:31on various streaming platforms. And the reason
- 10:33he wanted to use the bots to listen to the songs
- 10:37is that generates royalties. You get paid by
- 10:39these streaming services by the number of times
- 10:41a song is listened to. And so if you can get
- 10:44a song to be listened to millions of times, you're
- 10:46going to get more money. I'll come back to sort
- 10:50of now why he was changing the names. Mr. Smith
- 10:54allegedly knew that each of these streaming platforms
- 10:57had technology to try to detect fraud. And they
- 11:00would look at spikes in the number of streams
- 11:03of songs. or number of streams that just seem
- 11:06unrealistic. So for example, a song has a billion
- 11:09listens in one day, unless you're Taylor Swift
- 11:12or Beyonce, that's probably not realistic. And
- 11:15so he changed the names of each of the songs
- 11:18and had a limited number of listens to each of
- 11:20the songs. So again, This is actually just fraud,
- 11:25right? You made fake music to generate fake royalties,
- 11:28but you hyper -scaled it with artificial intelligence.
- 11:32We'll follow this closely. I'd encourage people
- 11:34to pay attention to if there's a guilty verdict,
- 11:36what the sentencing looks like, and if the government
- 11:38does in fact seek enhanced penalties because
- 11:41artificial intelligence is utilized. I think
- 11:45I'll turn it over to Matt now to talk about a
- 11:47different kind of AI crime. Thanks, Jen. Next
- 11:51slide, please. Thank you. So another aspect of
- 11:55AI and enforcement is when AI was the object
- 12:00of the crime. And here is actually an area where
- 12:04I think both the past administration and the
- 12:06current administration are probably going to
- 12:08have a decent amount of overlap in their interest.
- 12:11And that is protecting American technology companies
- 12:16and American AI developers from economic espionage
- 12:20or IP theft that may occur, particularly that
- 12:23involving foreign actors. Here's an example case,
- 12:27United States versus Ding, which was charged
- 12:30in the fall of 2024 when I was still at the U
- 12:35.S. Attorney's office by the U .S. Attorney's
- 12:36office in the Northern District of California.
- 12:39And that case involved allegations that the defendant,
- 12:44who was a software engineer, had actually been
- 12:47slowly stealing AI technology from his employer
- 12:53and moving those trade secrets to the People's
- 12:58Republic of China to two different entities with
- 13:00which he was associated. The investigation of
- 13:07this crime was extremely fast for a white -collar
- 13:10case, and I flagged that because I think that
- 13:14that is a trend that you're going to see in cutting
- 13:19edge AI technology cases where the government
- 13:21is trying to protect the technology. I think
- 13:24there's every incentive for both the government
- 13:26and the private sector together to move quickly
- 13:29on cases like this. At the time, it was a case
- 13:34that was... highlighted by the Attorney General
- 13:37of the United States. The indictment was announced
- 13:40by then Attorney General Merrick Garland at a
- 13:44conference for white collar criminal lawyers.
- 13:48And his personal involvement in the announcement,
- 13:50it's very rare for an attorney general to announce
- 13:53a federal indictment, I think signaled the last
- 13:56administration's interest in this type of case.
- 13:59But as I said, I expect that that type of enforcement
- 14:03interest will you know, has bipartisan appeal
- 14:06for the current administration as well. Unlike
- 14:09Smith, Ding has already gone to trial. He was
- 14:12convicted in January of 2026 of stealing trade
- 14:17secrets and of economic espionage, which is stealing
- 14:22the trade secrets with the intent of aiding a
- 14:26foreign country. Next slide, please. So another
- 14:33Another bucket of enforcement actions that Jen
- 14:38mentioned at the start is AI washing. Like Jen
- 14:42said, a lot of these AI -related cases end up
- 14:47being frauds. The subject of the fraud is artificial
- 14:53intelligence. It's not that AI is an instrumentality
- 15:00to supercharge the offense. And it's not that
- 15:03AI is the object of the trade secret theft or
- 15:06any kind of crime. It's that AI is the subject
- 15:10of lies for companies who want to get more investments
- 15:14or want to retain investments and realize that
- 15:18as the entire economy shifts towards an AI -centric
- 15:22model across industries, there's every incentive
- 15:27to speak glowingly about the use of AI in your
- 15:32business. Now, the folks in these cases were
- 15:38these are two example cases in United States
- 15:42versus scale charged in June of June and September
- 15:48of 2024 are these two defendants, Gale and Junko.
- 15:52The defendant in scale was ultimately pled guilty,
- 15:57whereas the defendant in the Junko matter remains
- 16:00just under indictment. But these are both classic
- 16:03examples of investment fraud. So you have two
- 16:09individuals who in soliciting investments from
- 16:13private persons just lied about the AI capabilities
- 16:19of their company. So for scale, the claim was
- 16:24that they had digital employees that could replace
- 16:28five human employees and save your company $100
- 16:33,000 a year. And based on that statement, raised
- 16:39over $40 million in investor money. falsely represented
- 16:47that the AI component would automate a lot of
- 16:51the work that your human employees were doing.
- 16:56And in fact, the founder also engaged in embezzlement
- 17:00against the company. So some side projects that
- 17:03didn't fare well for him. On Juco, this case
- 17:08was indicted in the Southern District of New
- 17:10York. It's still ongoing. There's another case
- 17:13where it was fraud to raise money. But the fraud
- 17:18was related to AI. So the owner of the company
- 17:22lied about the capabilities of the AI technology
- 17:26that the company was running and said that they
- 17:32could match employers with diverse job candidates
- 17:35using seven different AI algorithms, as well
- 17:39as a fully automated recruiting solution. In
- 17:43fact, the company had no such technology, but
- 17:47the company raised $21 million from 27 different
- 17:51investors. the allegations there, and I should
- 17:56say they are allegations because he remains indicted,
- 18:01are that the underlying technology actually wasn't
- 18:04anywhere close to what was being presented to
- 18:08investors. So these cases are, again, just illustrative,
- 18:11and I'm sure you'll see more of them going forward,
- 18:13and I know Christian in a moment will talk about.
- 18:15AI washing in civil enforcement actions. But
- 18:20they're illustrative of the fact that you don't
- 18:22need complex technology to get an AI washing
- 18:26charge. I mean, you can have simple fraud that
- 18:30takes advantage of so much investor interest
- 18:33in AI technology and in the explosive growth
- 18:37in AI capabilities. So this is just another facet
- 18:44of how AI and the enforcement matrix are interacting.
- 18:48I will turn it over to Christian. MoviePass and
- 18:55Tadris Capital, for example, are two of the earliest
- 18:59AI washing enforcement cases brought by the DOJ
- 19:01and show how the pattern of prosecuting AI -focused
- 19:04fraud was established before, as we'll talk about
- 19:07later, it became a stated priority. For MoviePass,
- 19:10in November of 2022, MoviePass's CEO was charged
- 19:14with securities fraud and wire fraud for misrepresenting
- 19:17that MoviePass's AI tools were capable of analyzing
- 19:21and monetizing subscriber data. when in reality
- 19:24the company didn't have an AI -enabled platform.
- 19:27And then in January of 2025, the MoviePass CEO
- 19:30actually pled guilty to defrauding most investors.
- 19:33So again, here the timeline is kind of noteworthy
- 19:36because the MoviePass conduct predates the current
- 19:38AI boom. AI washing cases aren't limited to just
- 19:41this generative AI era, and instead it appears
- 19:44that any misrepresentations about AI, whenever
- 19:47they occur, are fair game. Second example is
- 19:50Tadris Capital. So in September of 2023, Mina
- 19:53Tadris, the founder of Tadris Capital, was charged
- 19:56with wire fraud and investment advisor fraud.
- 20:00Tadris falsely represented that his fund used
- 20:02AI -driven algorithmic trading to generate guaranteed
- 20:05returns. But in reality, it was just a hedge
- 20:09fund. I mean, a Ponzi scheme run is a hedge fund,
- 20:11using new investors' money to pay the promised
- 20:13returns. In February of last year, Tadris pled
- 20:17guilty to investment advisor fraud. And in August
- 20:20of last year, Tadris was sentenced to 30 months
- 20:22in prison. So there it was an AI washing on top
- 20:26of a Ponzi scheme. Using AI enabled investment
- 20:29claims is the hook, which made investors invest
- 20:32nearly $6 million with less than 1 % of those
- 20:34funds actually going to legitimate investments.
- 20:37They just funded exotic travel and luxury goods.
- 20:40If you wouldn't mind going to the next slide.
- 20:44Perfect. And then three civil examples. So Delphia
- 20:48and Global Predictions are two of the earliest
- 20:50AI washing cases brought by the SEC, and Presto
- 20:54automation occurred just last year. So in Delphia
- 20:57in March of 2024, the SEC fined Delphia $225
- 21:01,000. There the firm falsely claimed again that
- 21:04it used AI to analyze client spending and social
- 21:07media habits to tailor their investment approaches
- 21:10and advice when No such data informed its investment
- 21:13practices, nor did they have the tools to back
- 21:15that up. And again, this is one of the first
- 21:17like of the SEC's AI related actions. Also in
- 21:21March of 2024, the SEC find global predictions
- 21:25$175 ,000. They're the firm falsely advertised
- 21:29that it developed an AI driven investing algorithm
- 21:33that use client data to provide expert AI driven
- 21:36forecasts. This was just a robo advisor offering
- 21:40advisory services to retail clients. But in reality,
- 21:43it was not an AI enabled platform and they did
- 21:45not offer any tailored investment advice. And
- 21:49then a more recent example is Presto Automation.
- 21:52So in August of 2024, Presto Automation was delisted
- 21:55from the NASDAQ following a significant drop
- 21:58in its share price. And in January of 2025, Presto
- 22:01consented to a cease and desist order cooperated
- 22:04with the SEC who declined to impose any civil
- 22:06penalties, despite finding that the company made
- 22:09materially false and misleading claims in public
- 22:11filings and public statements in 2021 and 2023.
- 22:15There, the company claimed that its AI voice
- 22:17product eliminated the need for human order taking
- 22:20and drive -thrus when the company was aware that
- 22:23most orders still required human intervention.
- 22:26And like the other examples that Matt and Jen
- 22:28covered, the statutory framework here is not
- 22:30novel. We're talking about wire fraud, securities
- 22:33fraud, and investment advisor fraud. None of
- 22:35these statutes mention AI, but all of them reach
- 22:37no misrepresentations made to obtain money or
- 22:41defraud investors. I think it's over to Jen to
- 22:43talk about the Trump administration. Yeah, so
- 22:46before we turn to talking about the Trump administration,
- 22:47I just kind of want to summarize what you've
- 22:49heard, which is that, you know, prior to the
- 22:51Trump administration, we had a Biden administration
- 22:54that was quite focused on the risk and dangers
- 22:56posed by artificial intelligence in particular
- 23:00spaces, right, to the everyday consumer, right,
- 23:03to national security. um and to investors and
- 23:08so the question everybody's mind is right so
- 23:10what is the trump administration doing around
- 23:13ai and has the tide changed are we going in a
- 23:16different direction i think what you're going
- 23:18to hear from us is the jury is probably still
- 23:20out uh on that question but there's definitely
- 23:23at least been a tonal shift from the top about
- 23:27artificial intelligence. We are going to talk
- 23:30about some very specific executive orders that
- 23:33relate to artificial intelligence under the Trump
- 23:35administration. But before I do that, I just
- 23:37want to flag that there's a host of executive
- 23:40orders that have been issued by this administration
- 23:42that don't specifically relate to AI, but that
- 23:46ultimately could relate to AI. Let me give you
- 23:48an example. The administration now has executive
- 23:51orders relating to utilizing the False Claims
- 23:55Act to pursue government waste, fraud, and abuse.
- 23:57The False Claims Act is the statute that makes
- 24:00it unlawful to tell a lie to the government in
- 24:03order to procure a government contract. One of
- 24:07the areas of focus for the government under the
- 24:09False Claims Act is lies related to artificial
- 24:12intelligence that led to the government giving
- 24:15you a contract or buying a service or product
- 24:18from you. So that's sort of a separate whole
- 24:20area of executive orders that aren't specific
- 24:23to AI, but that tangentially will relate to it.
- 24:26But let's turn to ones that are specific to AI.
- 24:28If we can go to the next slide. So this is an
- 24:31overview just of the administration's approach
- 24:35to artificial intelligence. What you'll see here
- 24:37is that there have been a series of executive
- 24:39actions that have outlined a vision for how AI
- 24:43should be governed at the federal level and what
- 24:46roles the state should play. There is a definite
- 24:49pro... AI stance that comes into each of these
- 24:53executive orders. The underlying message is that
- 24:57America needs to lead when it comes to artificial
- 25:00intelligence. in order to be competitive and
- 25:03to sort of win the AI war, particularly against
- 25:06China, but other countries as well. And that
- 25:10the government is going to use all of its various
- 25:13agencies to make sure that that happens. And
- 25:16so these orders cover the alphabet soup of executive
- 25:20agencies and directing the federal government
- 25:23to streamline bureaucracies that artificial intelligence
- 25:27can advance more quickly, but also considering
- 25:30the risks of doing that to make sure we're protecting
- 25:33Main Street against those risks. So let's turn
- 25:37to the next slide. I'll talk about one of the
- 25:39executive orders. So first executive order we'll
- 25:43talk about is titled Removing Barriers to American
- 25:46Leadership and Artificial Intelligence. The title
- 25:49kind of tells you everything I think you need
- 25:50to know about the executive order. This is an
- 25:53executive order that laid out that the government
- 25:56was revoking the Biden administration's executive
- 25:59order related to artificial intelligence that
- 26:02was more risk -based and risk -focused and was
- 26:05going to prioritize deregulation and eliminating
- 26:09bias against artificial intelligence and artificial
- 26:12intelligence training. It also said that there
- 26:16was going to need to be an action plan. So it
- 26:18directed a variety of different agencies to sit
- 26:21down, come up. with an AI action plan. The action
- 26:25plan itself, which Matt will touch on in a second,
- 26:27I think is lengthy. And I would encourage the
- 26:31folks who have put comments into the Q &A to
- 26:34take a look at it, because it addresses some
- 26:36of these questions around how are we going to
- 26:38protect certain people from AI and what are our
- 26:40priorities around that protection? Matt, you
- 26:45want to take over? Sure. Thanks, Jen. In December
- 26:51of just this past year, the Trump administration
- 26:54issued yet another executive order, this one
- 26:57entitled Ensuring a National Policy Framework
- 27:00for Artificial Intelligence. This is really an
- 27:04effort to protect the growing AI industry from
- 27:11different state -by -state regulation and So
- 27:16the purpose of the order is to direct different
- 27:20parts of the federal government to do what it
- 27:23can to remove barriers to the AI industry's growth.
- 27:28And the barriers the administration has identified
- 27:31are mostly regulatory in nature or the threat
- 27:34of state regulation, because as I'm sure you've
- 27:37all seen, there's been a lot of news about different
- 27:41states trying to think through what kinds of
- 27:44regulations might be appropriate for protecting
- 27:48consumers or others in their state. Most prominently,
- 27:53of course, California has had a very public debate
- 27:55about that. And the executive order tries to
- 27:59get in front of that by directing to the extent
- 28:04can federal agencies to issue regulations and
- 28:09try and preempt state authority in this area.
- 28:13So in particular, I want to highlight that the
- 28:16executive order comes right out and says that
- 28:20it favors a minimally burdensome national standard.
- 28:23This is the kind of language that actually is
- 28:26pretty rare in a federal executive order and
- 28:29that it so clearly takes a minimalist regulatory
- 28:33approach. it directs the FTC to issue a policy
- 28:38statement on the application of the FTC Act and
- 28:43how unfair and deceptive acts and practices which
- 28:46are prohibited under the FTC Act, how that interacts
- 28:48with AI models and the growth of AI. We expect
- 28:54that that will end up being significant for companies
- 28:58that are making AI -related marketing representations
- 29:04or representations directly to consumers, and
- 29:06it may give consumers and the FTC a roadmap for
- 29:11civil enforcement that might follow once the
- 29:15FTC acts. It also will probably be persuasive
- 29:21to courts when they're thinking about whether
- 29:23state laws that that interpret similar language
- 29:28in state acts to the FTC Act's language, whether
- 29:33it applies to certain AI -related conduct, that
- 29:39may be informed by the FTC's own view of what
- 29:42the act, the FTC Act involves. Similarly, and
- 29:46last point on this, December executive order
- 29:50directs the secretary of commerce to issue a
- 29:52policy notice, specifying the conditions under
- 29:55which states may be eligible for funding under
- 29:57the broadband equity access and deployment program.
- 30:01I think in general, everyone's eyes are on FTC
- 30:04on this. And I would not be surprised and Jen
- 30:10and Christian can jump in if they disagree, but
- 30:12I would not be surprised to see. more executive
- 30:14orders along the lines of what we've just described,
- 30:18because you've seen the pace and volume of those
- 30:21executive orders be quite high when it relates
- 30:23to AI under this administration. Next slide,
- 30:26please. Christian, I think you're on mute. Thanks,
- 30:33Matt. So US versus Sanager is the first AI washing
- 30:38case brought under the Trump administration as
- 30:41Jenna Matta both previewed. This is one of the
- 30:43common threads that's continued from the last
- 30:45administration until the present. And here the
- 30:48SEC and DOJ filed parallel actions, but on the
- 30:52DOJ front in March of 2025. The U .S. Attorney's
- 30:57Office for the Southern District of New York
- 30:58charged Albert Sanager with securities fraud
- 31:00and wire fraud. The indictment alleged that Sanager
- 31:03raised over $40 million by claiming his shopping
- 31:06app Nate used cutting edge AI. He claimed a 93
- 31:11to 94 percent automation rate for a single click
- 31:14shopping website and touted a custom built deep
- 31:17learning model. In reality, the automation rate
- 31:20was stuck at 0 % and the transactions were processed
- 31:23manually by contract workers overseas. So Sanager
- 31:27instructed employees to keep overseas contractors
- 31:29secret, controlled information shared with investors,
- 31:33and expectedly the company ultimately failed,
- 31:35leaving investors with new total losses. In December
- 31:38of 2025, Sanager pled not guilty to securities
- 31:41fraud and the case is still ongoing. However,
- 31:44notably on a policy point, the prosecutors framed
- 31:47the Sanager case as promoting American technological
- 31:49leadership, stating that Sanager's misrepresentations
- 31:53and AI misrepresentations more broadly not only
- 31:57victimize innocent investors, they divert capital
- 32:00from legitimate startups, make investors skeptical
- 32:02of real breakthroughs, and ultimately impede
- 32:05the progress of AI development. And this framing
- 32:08is important because it's consistent with the
- 32:10lens that Jen described with Executive Order
- 32:1314179, which is aimed at removing barriers to
- 32:16American leadership in AI, which includes ensuring
- 32:20that money is going to the trustworthy and noteworthy
- 32:24businesses that are not abusing or misstating
- 32:27their AI capabilities. So the government's not
- 32:29just prosecuting fraud. It's prosecuting fraud
- 32:31that threatens American AI competitiveness. And
- 32:35then I think it's over to Matt to talk about
- 32:37some enforcement tools that the US is using.
- 32:40Thanks, Christian. So we're going to touch briefly
- 32:44on how the federal government is actually using
- 32:47AI in its enforcement capacity. So I'll touch
- 32:51on two specific examples related to the Department
- 32:55of Justice's use of AI and the FBI's use of AI
- 32:59in their investigations and prosecutions. And
- 33:02then I'll turn it back over to Christian to cover
- 33:04the SEC and the Treasury Department's use of
- 33:07AI. in their work. So the first thing I want
- 33:11to mention is the healthcare initiative. So in
- 33:13June of 2025, the Department of Justice announced
- 33:17that it had charged 324 defendants with healthcare
- 33:20fraud, totaling $14 .6 billion, which is the
- 33:25largest healthcare fraud takedown in Department
- 33:27of Justice history. Obviously these are huge
- 33:31numbers and most of the time when these kinds
- 33:33of announcements are put together, there is sort
- 33:35of a scramble to find related cases that might
- 33:38be part of that initiative. But here there was
- 33:42a through line in the announcement in that. the
- 33:46head of the criminal division highlighted the
- 33:48use of the department's new tool, which is the
- 33:51healthcare fraud data fusion center, which utilizes
- 33:55AI data mining and advanced metrics to detect
- 34:00suspicious activity, such as overbilling and
- 34:02shipping materials to non -existent addresses.
- 34:05Now, you probably, if you work in this space
- 34:08or advise in this space, recall that under the
- 34:11past administration, there was a lot of talk
- 34:13about big data. at the Department of Justice
- 34:16and using data to make cases or pick out suspicious
- 34:22patterns, in particular in bills that were going
- 34:25to the government. So healthcare is the most
- 34:27obvious example because the government pays out
- 34:30such a large amount of its dollars in healthcare
- 34:32dollars. The government has a lot of data as
- 34:35to what claims are being submitted to it in order
- 34:39to pay those claims out. there have been for
- 34:43a number of years, lists of outlier billers.
- 34:48Billers who are in the top 10 in a certain category
- 34:51or for a certain code in the healthcare industry.
- 34:54But this goes a step further, right? Because
- 34:57this Healthcare Fraud Data Fusion Center that
- 34:59was announced actually is going to have all of
- 35:03that data centralized. And instead of just sort
- 35:06of top 10 lists or an agent sort of pouring over,
- 35:11different codes to figure out if anyone's an
- 35:13outlier on charts. AI is going to do that analysis
- 35:17and flag suspicious activity, and it may not
- 35:20rely on simply who is the highest biller of a
- 35:23certain code in a way that's outsized. It may
- 35:25rely on patterns of activity that are subtler,
- 35:29which I think will aid in the department's work.
- 35:32The Federal Bureau of Investigation, the FBI,
- 35:35has confirmed that it also is using AI tools
- 35:38to detect criminal activity and, in particular,
- 35:42national security threats. So the FBI is using
- 35:46AI to sift through huge amounts of data, and
- 35:51they confirm that yet again in January of this
- 35:53year. And that presents, again, a further step
- 35:58in the direction of sort of AI being run on large
- 36:03data sets without the need for constant investigator
- 36:07involvement, which as the federal government
- 36:12goes through rounds of budget savings or cuts
- 36:16in federal employees, you might think that that
- 36:19would naturally lead to decreased enforcement
- 36:21risk. But here, I think they're trying to make
- 36:24up ground by embracing the use of technology
- 36:27in the form of AI to do that data analysis. in
- 36:31a way that might be helpful as they lose agents.
- 36:36Next slide, please. Awesome. So two more fun
- 36:41AI enforcement tool examples. First, the Treasury
- 36:45Department announced that its Payment Integrity
- 36:48Office prevented and recovered over $4 billion
- 36:51in fraudulent and improper payments in fiscal
- 36:53year 2024 through its use of enhanced fraud detection
- 36:57processes, namely machine learning AI. So more
- 37:00specifically through initiatives such as AI expedited
- 37:05check fraud identification, the Treasury Department
- 37:07recovered a billion dollars in fraudulent payments
- 37:09and stated that it's AI enabled risk based screening
- 37:13processes prevented another 2 .5 billion in improper
- 37:16payments. So the $4 billion figure is significant
- 37:19and established a new high watermark for the
- 37:21Treasury Department, which shows that this kind
- 37:24of AI -driven fraud detection is unlikely to
- 37:27go away, regardless of what administration's
- 37:29in office. Next, in August of 2025, the SEC created
- 37:35its AI Task Force, which will utilize, thank
- 37:39you, and enhance the efficiency of accuracy in
- 37:43its investigations to the use of AI and generally
- 37:45signal to more proactive posture with regard
- 37:48to the use of AI. The SEC AI Task Force means
- 37:52that securities investigations will increasingly
- 37:54be initiated and supported by AI detected anomalies
- 37:58and market data, disclosure filings, trading
- 38:01activity. And in short, the government's now
- 38:05essentially an AI powered investigator and compliance
- 38:08programs should be designed with the assumption
- 38:10that AI driven regulators will find anomalies
- 38:13that manual oversight may have missed in the
- 38:15past. And so controls need to be at least as
- 38:17sophisticated as the tools used to detect them.
- 38:20I think it's over to Jen. So we're going to be
- 38:23in turn to the next slide. We're going to pivot
- 38:24a little bit. We've kind of talked about the
- 38:29way in which AI was investigated under the Biden
- 38:33administration, the way it's being investigated,
- 38:37but also supported under the Trump administration
- 38:39and the ways in which the government is utilizing
- 38:42artificial intelligence in the enforcement space.
- 38:45We want to talk about the way we're all using
- 38:47artificial intelligence as we defend against
- 38:51these investigations and defend against these
- 38:54cases. And I don't just mean using artificial
- 38:57intelligence to like help write briefs. There's
- 38:59lots of cases right now about lawyers being criticized
- 39:03for not disclosing they've used artificial intelligence
- 39:05in court filings. We're gonna talk about the
- 39:08attorney -client privilege and work product doctrine
- 39:10because there's breaking news in this space.
- 39:13The first two cases at the federal level relating
- 39:16to the use of artificial intelligence and these
- 39:19two doctrines have come out. And we think it's
- 39:22important that we all pay attention to them.
- 39:25And they're also important because they feed
- 39:27into advice we're going to give you about building
- 39:29a compliance framework. So you need to understand
- 39:32how to obtain privilege over work product and
- 39:35your communications with clients. And companies
- 39:38need to understand that so they build robust
- 39:40compliance systems to protect those privileges.
- 39:43So just to remind us of what these are, the attorney
- 39:47-client privilege obviously protects one, communications,
- 39:50between a client and the attorney, and in some
- 39:53circumstances, people who are assisting the attorney.
- 39:57Notice I said people. We'll come back to that
- 39:59in a minute. Two, it's intended to be and is
- 40:02in fact kept confidential. So you haven't intentionally
- 40:07disclosed the information to your adversary,
- 40:10for example. And third, it's for the purposes
- 40:13of obtaining legal advice. The reason that you
- 40:15are having the communication is either to get
- 40:18legal advice from the person you are speaking
- 40:20with or to utilize in seeking legal advice from
- 40:25counsel. That's the attorney -client privilege.
- 40:28Work product doctrine, slightly different, governed
- 40:30right by the rules of evidence. And you have
- 40:35three standards here too. One is the materials,
- 40:38they have to be documents or tangible things.
- 40:41They have to be prepared in anticipation of litigation.
- 40:44There's lots of case law on what that means to
- 40:46anticipate litigation. There has to be a real
- 40:49genuine threat. And three, it has to be prepared
- 40:52by or for a party or its representative. So it
- 40:56has to be prepared by someone who's representing
- 40:58themself or it has to be prepared by counsel
- 41:01in anticipation of litigation. So super familiar
- 41:04test to all of us, but in the world of AI, when
- 41:07now we're asking ChatGPT and Claude and Legora
- 41:10and Vincent to name your AI tool, legal questions.
- 41:14Are those covered? So let's look at the first
- 41:16two cases. I'll talk about one and then Matt
- 41:18will talk about the second and then Matt will
- 41:21compare and contrast them so we can go to the
- 41:23next slide. So the first case we're gonna talk
- 41:26about is United States versus Heppner. This is
- 41:29a case that came out of the Southern District
- 41:31of New York Federal Court that was written by
- 41:35Judge Rakoff. Judge Rakoff ruled on this case
- 41:39very quickly. We think a little bit in parts
- 41:41that he could be at the leading edge of the conversation
- 41:44about how these doctrines work. And at the end
- 41:48of the day, what he held is that documents that
- 41:51Mr. Heppner, who is a criminal defendant produced
- 41:53using a publicly available generative AI tool,
- 41:57Claude, were not protected either by the attorney
- 42:00-client privilege. or by the work product doctrine.
- 42:04The basic facts are these that Mr. Heppner was
- 42:07a criminal defendant. He's a former financial
- 42:09services executive. He had received a grand jury
- 42:13subpoena. So he was aware that he was under investigation
- 42:17and he was later indicted for federal securities
- 42:20fraud and for wire fraud. While he was under
- 42:24indictment and before he was arrested by the
- 42:28FBI, he used Claude to develop and outline defense
- 42:33strategies and to assist in discussions with
- 42:35his attorneys. He essentially went to Claude,
- 42:38asked it some questions, and then the answers
- 42:40that Claude gave him, he shared with his counsel
- 42:43after he had received the grand jury subpoena.
- 42:46When the FBI raided his home and ultimately arrested
- 42:50him, they seized 31 documents that memorialized
- 42:54his communications with Claude. The government
- 42:57moved to admit those into evidence. The defense
- 43:01opposed, arguing that they were covered both
- 43:04by the attorney -client privilege and by the
- 43:07attorney work product. We can turn to the next
- 43:09slide. Judge Rakoff ruled against Mr. Hepner
- 43:15on both counts. He did so in a bench ruling first,
- 43:20and then he followed with a written opinion.
- 43:21I would commend everybody to read it, or you're
- 43:24welcome to go to Cleary Gottlieb's website and
- 43:26read the article that Matt and I wrote summarizing
- 43:29this. But on the attorney -client privilege,
- 43:32he said, no go. One, generative AI tools are
- 43:37not lawyers. It goes back to my point. It's not
- 43:39a person. It's not a human. And Rakoff emphasized
- 43:43that AI tools are not bound by fiduciary duties
- 43:46that are owed to clients, and there's no quote
- 43:48unquote trusting human relationships. So he said
- 43:52it failed at that very first prong. Chatbot is
- 43:56not your lawyer. Second, he said, it has to be
- 44:01intended to be, and in fact kept confidential.
- 44:03And here, unfortunately, Cloud's user privacy
- 44:07policy explicitly states that the parent company
- 44:11reserves the right to disclose your inputs, meaning
- 44:13the questions you ask, and the outputs, so the
- 44:17information that the AI tool gives you for training
- 44:20purposes and to give to other third parties,
- 44:23including potentially government. regulators,
- 44:26meaning that if a subpoena from the government
- 44:29is given to Claude for your inquiries and the
- 44:32output from your inquiries, the government will
- 44:34turn it over. The court also cited a prior case
- 44:39in Ray OpenAI, which is actually an intellectual
- 44:42property case, which in that case, the court
- 44:44had observed that AI users do not have a substantial
- 44:47privacy interest in their communications on publicly
- 44:51available AI platforms. I emphasize publicly
- 44:53available because we're not talking about a closed
- 44:57source AI system that many of our law firms and
- 45:01companies have, right, that is not public, it's
- 45:04private. And third, he said it wasn't for purposes
- 45:09of obtaining legal advice. The court found that
- 45:12the defendant, Mr. Heppner, had communicated
- 45:14with the AI tool for the purpose of discussing
- 45:17his case with counsel. And the court recognized
- 45:20that some AI tools may function as highly trained
- 45:23experts and can kind of be an agent of a lawyer.
- 45:27But it said that when you share nonprivileged
- 45:29communications with counsel, that doesn't make
- 45:31them privileged. And that is just black. letter
- 45:34law. You can't cloak a non -privileged document
- 45:38and privilege just by turning it over to your
- 45:40lawyers. So Mr. Heppner's communications with
- 45:43AI and the output were not privileged in the
- 45:46first instance, and by giving them to his counsel
- 45:48to assist his counsel in providing them an advice,
- 45:51it did not cloak them in privilege. So I want
- 45:55to turn it over to Matt to now talk about another
- 45:57decision that came out of a different court in
- 46:00a different context. in the space of AI and privilege.
- 46:04Thanks, Jen. So in Warner versus Gil Barco, the
- 46:08Eastern District of Michigan held that what could
- 46:14be viewed as similar conduct was protected by
- 46:18privilege. And I just want to break it down a
- 46:20little bit to get into some of the nuances. So
- 46:23in Warner, there was a pro se plaintiff, Ms.
- 46:26Warner, who was suing her employer for race -based
- 46:29discrimination. Her employer during discovery
- 46:33demanded that the plaintiff turn over Her entries
- 46:41into chat GPT and notes from and logs from those
- 46:46interactions. Interestingly, the plaintiff was
- 46:49pro se, so that sort of complicates the analysis
- 46:52here, but the pro se plaintiff resisted and cited
- 46:56the federal rules of civil procedure, which protect
- 47:00work product, citing the rule that in particular
- 47:04protects documents and tangible things that are
- 47:08prepared in anticipation of litigation or for
- 47:11trial by another party or its representative,
- 47:14citing that rule and saying, I'm a party and
- 47:18I use this tool in preparation and anticipation
- 47:21of litigation. The court, Magistrate Judge Patty
- 47:26held that the pro se plaintiffs chat GPT queries
- 47:30and AI responses were protected by the work product
- 47:34doctrine, finding that that those logs and entries
- 47:39actually reflected the litigants, the party's
- 47:42mental impressions and thought processes and
- 47:45therefore were protected by the rule. The court
- 47:49rejected the defendant's request for a production
- 47:51of all of these materials. So what you see here
- 47:54is, you see within a week of each other in February
- 47:57of 2026, two federal courts looking at similar
- 48:02fact patterns, right? Not the same, they have
- 48:04important differences, but similar fact patterns.
- 48:07which is when a litigant or defendant used an
- 48:12AI chat bot or service in connection with litigation
- 48:15and reaching quite different conclusions on whether
- 48:19the underlying material is protected by privilege.
- 48:23So overall point, and I'm going to get to some
- 48:25of the differences between the two cases that
- 48:27sort of. directed the outcome, but overall point
- 48:30is stay very attuned to this space because I
- 48:35suspect that there will be many more cases fleshing
- 48:39out these uncertainties and these nuances in
- 48:42ways that will add to all of our understanding
- 48:45of the law because this is just the first salvo.
- 48:48Next slide, please. So some of the differences
- 48:53between these two opinions, the Heppner case
- 48:56that Jen discussed, and then the Warner versus
- 48:58Gilbarco case that I just mentioned. In Heppner,
- 49:03the reasoning was that the work product doctrine
- 49:05was not intended to shield materials where they
- 49:09ended up in the attorney's possession, as Jen
- 49:12mentioned, but they weren't prepared by the attorney
- 49:15or their agent or at counsel's direction. And
- 49:18therefore, use of this third party tool, this
- 49:23AI tool, did not mean that it was protected by
- 49:27privilege. The reasoning in Warner was slightly
- 49:32different, which is, that the chat logs, as I
- 49:38mentioned, reflected the pro se litigants' mental
- 49:40impressions, and as the party reflected the thought
- 49:44processes in anticipation of litigation. And
- 49:50importantly, I think, and in a different sort
- 49:52of mode of reasoning, recognize that GenAI tools
- 49:56like ChatGPT are... tools of a litigant and are
- 50:02not actually third party services or another
- 50:05party. And you may see more and more teasing
- 50:08out of the differences there as courts grapple
- 50:11with this, because as Jen mentioned, of course,
- 50:15law firms or many large companies that have enterprise
- 50:18level agreements or terms and conditions with
- 50:23AI companies that restrict access to and reflect
- 50:27confidentiality of information that's entered
- 50:30into the AI tool are obviously differently situated
- 50:33than the public facing tools. But even the public
- 50:36facing tools, judges may end up looking at them
- 50:40slightly differently depending on the fact pattern.
- 50:44Next slide, please. Matt, could I just weigh
- 50:49in on that for one second? Please. I would just
- 50:51add, to kind of sum up that, that when a court
- 50:55considers whether the attorney -client privilege
- 50:57or the work product doctrine cover communications
- 51:00or documents and tangible things, it has always
- 51:03been a facts and circumstances test. Those cases
- 51:08and those decisions are almost always restricted
- 51:10to the facts that are before the court and serve...
- 51:14minimal presidential value, especially at the
- 51:16district court level. And so I would encourage
- 51:19everyone to keep that in mind as they're thinking
- 51:22about these cases. This is not breaking any new
- 51:25ground. The framework the court is using is the
- 51:28framework courts have always used. It's just
- 51:30applying them to the very unique facts and circumstances
- 51:33of using artificial intelligence. I totally agree
- 51:36with Jen's point. And one additional point would
- 51:40be, you also have to look at the consequences
- 51:43for access to justice that that second decision
- 51:46might have had if it had gone the other way.
- 51:49I think that there was probably some concern
- 51:53on the part of the judge who was facing that
- 51:56motion that if the judge did not allow a pro
- 52:00se litigant to use AI tools that were publicly
- 52:03available, it would have a pretty dramatic effect
- 52:09on future litigants and the relative fairness
- 52:12versus unfairness of pro se litigants having
- 52:15access to some of the same tools that litigants
- 52:19who are represented have. So I think that also.
- 52:21might have played a role in the difference. So
- 52:24we wanted to finish today with building a compliance
- 52:28framework and just sort of some very practical
- 52:31steps to think through when you're advising a
- 52:34client on how to put together a compliance program
- 52:39that really accounts for the growth in AI use
- 52:43and adoption. Or if you're in -house, how you're
- 52:46designing it yourself. And this is something
- 52:49that We talked to folks about quite a bit. And
- 52:52so I think we had just some practical advice.
- 52:55If you could turn to the next slide, please.
- 52:59So this is modeled as something of a checklist,
- 53:02but of course, each of these categories in and
- 53:05of themselves. have whole checklists that you'll
- 53:08wanna think through. But we just wanted to flag
- 53:10sort of big picture, what are things that companies
- 53:15should be thinking about, in -house teams should
- 53:16be thinking about, outside council should be
- 53:19thinking about and advising companies in how
- 53:21to build compliance systems that can tackle this
- 53:26area. So the first. thing we wanted to mention
- 53:28is third -party risk management. When engaging
- 53:32vendors that provide or use AI systems, it's
- 53:36important to apply the same compliance framework
- 53:42that you might apply to your own internal use.
- 53:45So you want to ask about their AI inventory,
- 53:48their risk assessments to the extent they have
- 53:52policies, procedures, and training that they're
- 53:54willing to share with you. making sure that there's
- 53:58some human oversight to those systems because
- 54:02third party AI exposure may end up being your
- 54:06exposure if they're an important part of your
- 54:11business as well. And so it could present an
- 54:14underappreciated area of organizational risk.
- 54:17The second thing we wanted to flag is, we think
- 54:21a starting place for most businesses and the
- 54:25compliance function, is to just take an inventory
- 54:27of what the AI use is at your business and figure
- 54:33out the systems that might be used. You can't
- 54:39really build a compliance framework in the modern
- 54:43age, in the AI era, unless you know what your
- 54:46employers are using, what folks in the business
- 54:49are using. And therefore, inventorying the AI
- 54:53tools that you use, are they enterprise grade,
- 54:56are employees using public AI tools? That can
- 55:00be important to your business and important to
- 55:04assessing third party vendors, products and services
- 55:07as well. Next, you want to understand potential
- 55:10risks. For each AI system that you identify in
- 55:15an inventory, you want to think through whether
- 55:18you can conduct diligence that addresses key
- 55:21threats to the business. Has the system ever
- 55:24been used as a vehicle for fraud or has that
- 55:29business or that type of AI ever been the subject
- 55:32of AI washing claims? How secure is the system
- 55:36against being compromised or exploited? Figuring
- 55:41out whether the AI tool you're using is or that
- 55:47a third party vendor is using is has security
- 55:52that would match your expectations for your own
- 55:54systems is really critically important because
- 55:58these systems are developing so quickly and and
- 56:02being deployed so quickly that security has to
- 56:06be at the forefront of folks thoughts when they're
- 56:09coming up with a compliance framework. In terms
- 56:11of human oversight, you want to make sure that
- 56:15real people are involved at some point in overseeing
- 56:19AI systems and are building controls for its
- 56:25use, particularly if you've identified risks
- 56:27in the previous step. Human oversight is a really
- 56:33important part of the compliance function, even
- 56:35as compliance teams understandably want to and
- 56:39do reasonably employ more and more AI tools to
- 56:42help their work. At the end of the day, human
- 56:49oversight really can't be replaced. And then
- 56:53the last two things I wanted to mention. would
- 56:55be to establish policies, procedures, and training.
- 56:59A compliance program really requires written
- 57:03policies that cover both appropriate and inappropriate
- 57:05use of every AI system and tool you've identified.
- 57:11paired with some meaningful training so that
- 57:14employees know the rules of the road, paying
- 57:17particular attention to the limitations that
- 57:20might apply to the permissible use of the system.
- 57:24In particular, our employees sharing confidential
- 57:28business information with AI systems or potentially
- 57:33privileged information outside of those enterprise
- 57:36level agreements that Jen mentioned earlier,
- 57:39that's worth flagging as something you want a
- 57:42written policy on, you wanna conduct training
- 57:44on, and have procedures when something goes awry
- 57:49or something is noticed. And then the last thing
- 57:52is to ensure that your external representations
- 57:56about AI use are consistent with all of the things
- 58:00we just said, with your inventory and your risk
- 58:03assessments, right? Because... when you're talking
- 58:06to investors, to clients, to regulators, to the
- 58:09public about how you as a business use AI. You've
- 58:14seen some of the potential pratfalls in what
- 58:19we just discussed in overstating AI use. But
- 58:22you want to also accurately disclose AI use to
- 58:25the extent it's not publicly known if you're
- 58:29talking to investors. So thinking through an
- 58:34accurate and consistent message across different
- 58:36stakeholders modulated to account for the risks
- 58:41that you've identified I think is really important
- 58:43because misalignment between you know what you
- 58:46say and what you're actually using or the capabilities
- 58:49of the technology you know may increase your
- 58:52exposure. So happy to, I don't see any open questions
- 58:58in the chat and we are exactly at time. So I
- 59:03want to thank everyone for tuning in and thank
- 59:06you very much to the Beverly Hills Bar Association
- 59:09for having us. It was a pleasure speaking with
- 59:12you today and I know I speak for Jen and Christian
- 59:15when I say thanks for tuning in.