Latest / CLEs You Actually Want to Hear / Generative AI and State Tax Administration with Travis Thompson and Brian Oliner
Transcript
- 0:02Welcome to Generative AI and State Tax Administration.
- 0:07I'm Philipp Berendt, share of the tax law section
- 0:09of the Beverly Hill Bar Association. The tax
- 0:12section is sponsored by Hochmann, Sork and Torscher
- 0:14Perez. The Hochmann, Sork and Torscher Perez
- 0:17PC advantage is experienced and I for innovation
- 0:20and genuine dedication to clients, their families
- 0:23and their business. The firm is also active and
- 0:26successful in handling civil tax and criminal
- 0:29tax controversies. and related litigation in
- 0:32federal and state courts. I'm delighted to introduce
- 0:35our speakers, Brian O 'Liner and Travis Thompson.
- 0:39Travis Thompson is director at Fenmore's Business
- 0:42and Finance Practice Group, where he advises
- 0:44clients on complex tax matters, including litigation,
- 0:47refund claims, voluntary disclosures, and cross
- 0:50-border compliance. He has represented clients
- 0:53in high -profile cases involving FBA, cryptocurrency,
- 0:56and international tax issues, and frequently
- 0:59writes and speaks. on the intersection of tax
- 1:02law and technology. Tervis also served as founder
- 1:06and chair of the ABA tax sections, tax practice
- 1:10and technology committee. Brian O 'Liner is general
- 1:14counsel at the Federation of Tax Administrators,
- 1:18where he advises on tax compliance and enforcement
- 1:20issues, drawing on over 14 years as principal
- 1:24counsel to the Comptroller of Maryland. His background
- 1:27spans government service, private practice, and
- 1:30frequent speaking on tax administration. He is
- 1:34a graduate of the University of Maryland School
- 1:36of Law and University of Maryland Baltimore County.
- 1:39And now let's get started. We can't wait to see
- 1:41what you have prepared to share for us today.
- 1:45Thank you, Phillip. Okay. Did everybody see that?
- 2:00Yep. Very good. Well, hey, thank you to the Beverly
- 2:03Hills Bar Association for allowing us to speak
- 2:07today. My name is Travis Thompson. I'm a director
- 2:10at Fennimore, based out of Walnut Creek, California.
- 2:14Really excited to be speaking with Brian today
- 2:18about generative AI, AI generally, and state
- 2:23and tax. state tax administration. I think every
- 2:28great conversation about what's going on in the
- 2:31tax world related to the AI starts with what
- 2:37is AI. But before we do that, let's just go through
- 2:42our brief agenda. I'm going to talk a little
- 2:45bit about what's going on in the world, why all
- 2:48this AI talk context is important. And practitioners
- 2:54around the United States need to understand why
- 2:57is this moment so important for the purposes
- 3:00of artificial intelligence and tax? I'm gonna
- 3:05talk a little bit about what the feds are doing
- 3:08and the executive orders from the White House
- 3:12related to AI, especially through the Biden and
- 3:15Trump administrations. And then we'll discuss
- 3:18a little bit about what the IRS is doing. and
- 3:22now the CDTFA, and then Brian will cover some
- 3:26more of the state policy around AI and what we'll
- 3:30see moving forward. So look, I always start with
- 3:37this particular slide. It's very important because
- 3:41AI is not so simple to define, right? But the
- 3:47best, simplest definition that we may have, is
- 3:51machines acting in ways that seem intelligent.
- 3:55I'm sure if we all asked every tax practitioner
- 4:02in Beverly Hills or in California what artificial
- 4:05means, we would likely get a very similar definition
- 4:08as to what artificial meant. It's that intelligence
- 4:13definition. that I'm sure we would get very different
- 4:18responses to, especially in this day and age.
- 4:22So the best definition that I was able to find
- 4:26for intelligence came from someone who studied
- 4:32AI at MIT named Howard Gardner and at Harvard.
- 4:37And he said that intelligence is the ability
- 4:40to solve problems or to create products. that
- 4:44are valued within one or more cultural settings.
- 4:47So just keep that in mind as we move forward.
- 4:51AI is here to stay. And if you're not using it
- 4:57at this moment, then I highly encourage you to
- 5:00get trained up as soon as possible because it's
- 5:05a major paradigm shift in the way that we practice
- 5:07tax, both federal and state. Just real quick.
- 5:15There's a lot of convoluted messages out there
- 5:18about what AI is But more importantly where it
- 5:21fits within the definitions of artificial intelligence
- 5:25especially this new generative AI Piece, you
- 5:30know perplexity chat GPT those modules and so
- 5:36what I like to do is just break it down real
- 5:39quick into You know the categories of AI and
- 5:44why that's important. So first off, I think many
- 5:48practitioners, just like in the 80s when Terminator
- 5:52came out, think AI is gonna take over our jobs,
- 5:57it's gonna become self -aware, it's gonna be
- 6:00like Hal in 2001, it's gonna be like Terminator,
- 6:04and that type of AI is... disputed at a very
- 6:11high level amongst academics and professionals
- 6:14and falls within a category called general AI.
- 6:19Many, if not all, the practitioners of artificial
- 6:24intelligence, the actual builders of it around
- 6:27the globe, many believe it'll never ever occur,
- 6:30but there are these other professionals and subset
- 6:34of folks that believe it will. But as of now,
- 6:38There's no evidence to suggest that general AI
- 6:43will ever become a thing. If you ever want a
- 6:46great thought experiment around this, I would
- 6:50ask you to go read what's called the Chinese
- 6:55Room Argument by Professor Sorrell, former Professor
- 6:59Sorrell at Berkeley. He talks about how computers
- 7:03don't have minds. He came up with this thought
- 7:05experiment in the 80s. And this is really kind
- 7:08of the main argument against general AI. For
- 7:13our purposes, what we like to talk about for
- 7:17AI is within the narrow AI scope, right? And
- 7:21narrow AI is really just machine -based systems
- 7:24designed to address very specific problems like
- 7:28playing chess. And I just want everybody to know
- 7:30that to date, All applications of AI are what
- 7:35are called narrow AI applications. So what are
- 7:42narrow AI applications? Well, robotics is definitely
- 7:46something we all know, right? Automated devices
- 7:50that perform physical tasks in the real world.
- 7:54We see that in Amazon centers, and we're very
- 7:58familiar with what robots are. The second one,
- 8:03natural language processing, you've probably
- 8:04heard of, we've all been using it. If I were
- 8:08live right now, I would ask the audience to put
- 8:12up their hand and tell me how many people own
- 8:17dictionaries on their desk. And many people,
- 8:21usually about one out of a hundred that say they
- 8:23have a dictionary on their desk. And that comes
- 8:26because we all now value spell check as a natural
- 8:29language processing AI platform. The utility
- 8:33of spell check is much greater than actually
- 8:37having a dictionary on our desk or being an expert
- 8:42speller. So we've all been using natural language
- 8:44processing. Whenever we type something in Google,
- 8:49that's a form of natural language processing,
- 8:51especially when it knows what we're going to
- 8:53type next. So keep that in mind that this has
- 8:57been part of our culture for a very long time.
- 9:01And so you've been using AI, whether or not you've
- 9:04known it or not. The most important one, though,
- 9:08for the purposes of today, for the purposes of
- 9:10the overall conversation about generative AI
- 9:14is machine learning, which is just really powerful
- 9:19computer systems that crunch data faster than
- 9:23they've ever been crutched faster than it's ever
- 9:25been crutched before. And they use this data
- 9:29to predict and automate learning so that the
- 9:33systems themselves can predict and learn in the
- 9:36future. And so this forms the basis of what our
- 9:41conversation is around AI, what's called large
- 9:45language models that are a subset of this machine
- 9:47learning. But really, these large data sets that
- 9:52exist are what are required in order for generative
- 9:56AI and artificial intelligence through machine
- 9:58learning to actually produce and work properly.
- 10:04This is a graphic I like to show folks. It's
- 10:08not very detailed, but it gives you a sense of,
- 10:12you know, where we're at in this generative AI
- 10:15chat GPT world. inside of the deep learning,
- 10:19machine learning and artificial intelligence
- 10:21scope, you see it falls under the category of
- 10:25narrow AI or machine learning. So what's all
- 10:30the fuss? Like, why is this a big deal? Well,
- 10:34unlike previous AI systems, the thing that makes
- 10:38generative AI so special is its ability to create.
- 10:44And when you really take a look, at everything
- 10:48that has come before open AI and chat GPT. None
- 10:54of those predictive algorithms ever were able
- 10:57to take data sets and then create, you know,
- 11:02paragraphs using language. The most important
- 11:08caveat here in ways that seemed cogent or reasonable,
- 11:14especially when it came to the tax code or law
- 11:19in general, right? But that is why this is so
- 11:23special, is generative AI can create unlike previous
- 11:28AI. And that obviously poses a whole set of issues
- 11:33for lawyers and accountants and practitioners
- 11:36that they need to understand and deal with because,
- 11:41and I'm sure you've heard about it, Yes, it can
- 11:43create, but that doesn't mean it's accurate.
- 11:46And I'm talking about the concept of hallucinations
- 11:50and what happens when the systems that are not
- 11:55based on law in the code create fake cases that
- 12:02attorneys have been putting inside of briefs
- 12:06that are submitted to courts. So keep that in
- 12:09mind, right? That is why we're having this conversation
- 12:14today because anything that can create in language
- 12:19can be very valuable as a tool for state and
- 12:24federal tax agencies. Now, aside from its ability
- 12:31to create, like what is this other thing that's...
- 12:34that's happening on the back end. Why now? And
- 12:37really it just comes, why the rise in artificial
- 12:40intelligence all of a sudden? Well, it comes
- 12:43down to the word, the one word, which was data
- 12:46or data, as some people say, right? I'm a Star
- 12:49Trek fan, I use the word data. So look, there's
- 12:54never been ever in the history of technology
- 12:59or even mankind, the amount of accumulated data
- 13:02as there is right now in our culture, right?
- 13:09Literally, if you look at the amount of data
- 13:14that Google and everybody is storing, it's more
- 13:17now in the last five minutes than it's probably
- 13:19stored in the last hour, right? And then within
- 13:22the last hour, it's probably more than it's been
- 13:24stored in the last week. And the reason for that
- 13:28is, is it's never been less expensive and the
- 13:31computers have never been faster to actually
- 13:34process this type of data. And so in order for
- 13:38artificial intelligence systems to run thoroughly
- 13:41and effectively, they actually need it. They
- 13:44need it. When I tell people that I don't look
- 13:47at this as a cell phone anymore, I look at it
- 13:51as a data accumulator. with lots of apps with
- 13:54companies that are competing for me to interact
- 13:57with those apps. Those apps are what generate
- 14:01the data to feed the artificial intelligence
- 14:04systems on the backend to the various companies,
- 14:08Facebook, social media, whatever. It doesn't
- 14:10matter, right? Everybody is competing for you
- 14:14to interact with their respective apps because
- 14:17they want the data that you provide to make their
- 14:20systems on the backend faster and more reliable.
- 14:25AI is worthless without data. It's worthless.
- 14:28So the more of it that it has, the better that
- 14:31it is. And if you really stop and think about
- 14:34it, who has the most proprietary data out of
- 14:38all of the agencies, right, in either state or
- 14:42the federal government? It's tax data. The IRS
- 14:45and in California, the California Franchise Tax
- 14:48Board have billions, trillions of terabytes of
- 14:53data that is not being used effectively to determine
- 14:57or enforce the tax code. So just keep that in
- 15:00mind that these agencies are getting smarter
- 15:03and faster with their technology tools in order
- 15:08to process the data that they previously haven't
- 15:11processed. And because of it, it's supposed to
- 15:14make the administration of the tax laws more
- 15:18efficient. and better for society as a whole.
- 15:21But as we will talk about later, there's going
- 15:24to be some hiccups, right, especially when it
- 15:26comes to ethics and, you know, bias, but also
- 15:31just, you know, enforcement efforts as a whole
- 15:34and who's being targeted and who isn't. So, next
- 15:40slide. And again, the computing power, which
- 15:43I already talked about is literally the fastest
- 15:46it's ever been. So data computing power allowed
- 15:50for generative AI to even be a thing. And now
- 15:53we have CHAT GPT and how that plays in with the
- 15:57overall scheme of the tax code and state and
- 16:03federal agencies using it. That's what we'll
- 16:06talk about today. Rather than create separate
- 16:14slides for each of the various executive orders
- 16:19that seem to come out of the current administration
- 16:21now every other day, I just decided to list all
- 16:27of the executive orders that pertain to AI from
- 16:35Biden and Trump too. And as you can see, it's
- 16:40rather substantial. So the Biden administration,
- 16:44when they were first teeing up their executive
- 16:48orders around AI, were very concerned. They had
- 16:53just come off having the Inflation Reduction
- 16:55Act funding, but they were also very concerned
- 16:57with the concept of bias, especially racial bias.
- 17:01And what they did is they directed all of the
- 17:04agencies in the federal government to essentially...
- 17:07write up reports that they would submit to the
- 17:10White House on how they were using AI and how
- 17:14they planned to eliminate bias and how they plan
- 17:18to manage the future of the implementation of
- 17:22AI. And a lot of it centered around the concept
- 17:25of ethics, but also just to get an overall summary
- 17:30of how the agencies were using this technology.
- 17:34at the time. I think the Biden administration
- 17:37only gave the agencies like six months to respond
- 17:42after the executive order came out. You can go
- 17:46read what Treasury put together for its department
- 17:51and the Internal Revenue Service if you like.
- 17:56But for the purposes of our conversation today,
- 17:59it's super in -depth and you know, rather lengthy
- 18:03and it wouldn't fit into us focusing on the states.
- 18:07So when Trump came in, you may have seen in the
- 18:15news and through the various tax notes, publications,
- 18:18or whoever, that their focus, especially teaming
- 18:22up with Doge and every leader it seemed like
- 18:27in Silicon Valley on AI, their focus really was
- 18:31on getting a grasp in their way on how the federal
- 18:36government would use artificial intelligence
- 18:38in the past. So it immediately rescinded all
- 18:41the Biden executive orders on AI. But then it
- 18:44wrote its own AI executive orders that came out
- 18:50right when Trump came into office. You can see
- 18:54January 20th, he didn't even waste any time.
- 18:56January 20th, 23rd. boom, boom, he comes out
- 19:00with two executive orders. Now, when you go read
- 19:02these, they're not dissimilar from what President
- 19:09Biden and his team had put together. They're
- 19:12just not. They use language about eliminating
- 19:17woke decision -making inside of AI and other
- 19:22various aspects of what you're hearing out of
- 19:25the current administration related to how government
- 19:29is working. But the reality is most of the AI
- 19:34executive orders that they've been putting together
- 19:37were really put out there because, one, Doge
- 19:42was asking for it, but also, two, they were looking
- 19:46for ways to make the IRS more efficient and downsize
- 19:50its workforce. by implementing these executive
- 19:55orders, they have and are attempting to find
- 20:00ways to use technology with the current Inflation
- 20:04Reduction Act funding still, find ways to use
- 20:07technology to streamline the processes at the
- 20:11IRS and throughout the federal government as
- 20:13it relates to AI. Now that they have the funding
- 20:16to do that, and they thought this was an ample
- 20:20opportunity to eliminate actual humans from the
- 20:26workforce in order to implement those particular
- 20:30systems. So you hear them going back and forth
- 20:34on this, but they are fully focused now on implementing
- 20:40various AI systems into the IRS in order to enforce
- 20:49the future tax laws. and current tax laws. So
- 20:54AI is going to be very prominent in the future
- 20:58of the IRS. And as practitioners, we need to
- 21:02understand the language that obviously encapsulates
- 21:07that and how we can best represent taxpayers
- 21:11moving forward. So right here, Removing barriers,
- 21:20I think this is actually a typo, sorry. But this
- 21:27is what I just talked about. I just had the Trump
- 21:32administration eliminating all the Biden administration's
- 21:37AI policies, but also they'll use words like
- 21:41America's dominance, promote human flourishing.
- 21:46this very colorful language to do the exact thing
- 21:49that the Biden administration was doing with
- 21:52their executive orders around AI. And of course,
- 21:57the concept of being free of ideological bias,
- 22:01they added the word ideological, whereas the
- 22:04Biden administration simply used the word bias,
- 22:07which is, you know, talked about often inside
- 22:10AI circles when it comes to government use of
- 22:15AI. This is a comment from Treasury Secretary
- 22:20Bessen on May 6, 2025. We are taking the IRS
- 22:25back to where it was before the Inflation Reduction
- 22:28Act bill substantially bloated the personnel
- 22:32and the infrastructure. I believe through smarter
- 22:36IT, through this AI boom, that we can use that
- 22:39to enhance collections. So I included that as
- 22:44just, you know, validation of the points that
- 22:48I had made earlier. And 2023, to go back a little
- 22:55bit, maybe you guys don't know or never heard
- 22:58of the RAS department inside of the IRS. That
- 23:02stands for Research and Applied Analytics and
- 23:05Statistics Group. And that, you know, prior to
- 23:08Trump coming in was the most technologically
- 23:12advanced group inside of the IRS. Where they
- 23:16tested new AI systems where they managed use
- 23:19cases? Where they were trying to determine, you
- 23:24know with vendors what the what the best possible
- 23:28systems were Moving forward for the IRS. I call
- 23:32these guys the Big Bang Theory guys. I mean super
- 23:35smart MIT Harvard Yale computer engineers, Berkeley,
- 23:40Stanford, all folks that had computer engineering
- 23:44backgrounds or economics backgrounds could write
- 23:46algorithms and work in technologically advanced
- 23:50spaces. So this is where the IRS was going with
- 23:54its hiring practices prior to Trump II. And you
- 23:59can see that, you know, they were sending out
- 24:03requests for, you know, information or requests
- 24:06for proposal on various aspects of AI. This is
- 24:10just related to LB &I, where they were looking
- 24:13to implement AI to accurately select returns
- 24:19and reduce the no change rate. So looking to
- 24:25RAS, it's kind of a cool department. Admittedly,
- 24:29I can't speak to its level of sophistication
- 24:33right now because of all the changes at the IRS.
- 24:36but research and applied analytics and statistics
- 24:39department, which used to be run by the guy that
- 24:43brought data analytics into the MBA, used to
- 24:48be a pretty cool spot for what the IRS was doing
- 24:51with AI. We're still waiting for some updates
- 24:55to see how they're doing moving forward. These
- 25:02are just some applicable authorities at the IRS.
- 25:05has used around AI and, you know, in the IRM.
- 25:10We had the OMB memorandum. I won't read this
- 25:15whole slide, but really, like, this is just to
- 25:18show you that, you know, internally they've put
- 25:22parameters around AI usage. They're very conservative
- 25:27with their implementation of it, even to this
- 25:30day. And the reality is, When TIGDA went back
- 25:37and looked at how they were actually using it
- 25:40and then report November 12, 2024, these are
- 25:46various AI -related systems that are both in
- 25:51use and in development at the IRS. You can see
- 25:5868 AI projects in total. I can tell you that
- 26:05the IRS Criminal Investigations Unit has been
- 26:08using Palantir and other data analytics systems
- 26:11for quite some time, you know, going on seven,
- 26:14eight years now, in order to build cases within
- 26:19CI. This was all very, I think, very surprising
- 26:23to folks, the way that the particular report
- 26:25laid out exactly how they were using artificial
- 26:30intelligence or looking at using it. But, you
- 26:34know, this is meant to show you that it's one
- 26:38of the top priorities at the IRS. 12 of those
- 26:43projects were to improve customer service. 27
- 26:48of those were to improve the enforcement efforts
- 26:51of the IRS. And then 29 of those just with improving
- 26:57the overall operations of what was going on at
- 27:01the IRS. and improving the efficiencies internally.
- 27:09Well, is the IRS using generative AI? They're
- 27:13exploring it. They've made some recent comments
- 27:15about using it. They're in the early stages.
- 27:19The IRS chief technology officer said that while
- 27:26AI is well established, generative AI is different
- 27:29for the reasons that we talked about earlier.
- 27:32but they are testing whether or not GenAI could
- 27:36possibly help in customer service situations.
- 27:40They're worried about accuracy, they're worried
- 27:42about bias, they're worried about questionable
- 27:45sources. There are ways to combat that by narrowing
- 27:52down the field of data that the AI is using to
- 27:55come up with his answers. But as of right now,
- 28:00they're And, you know, anecdotally, I can tell
- 28:03you in conversations with personality at the
- 28:06IRS, they are not, to my knowledge, as of now,
- 28:11using generative AI applications at the IRS.
- 28:22America's AI action plan. Well, this was kind
- 28:24of recent. But yeah, so. this came out really
- 28:32to as like an announcement to the states. I put
- 28:40this in here because essentially what the federal
- 28:43government is saying in this particular order
- 28:46is any discretionary funding that should be considered
- 28:54for the states for AI related instances, should
- 29:01be hindered if the states themselves are implementing
- 29:08policies that go against the federal government
- 29:11related to the federal government's AI policies.
- 29:15According to the White House, not all the states
- 29:18are doing that. Some states are effectively implementing
- 29:22AI systems, but this is really kind of a shot
- 29:25across the bow to the states. to let them know
- 29:31that the possibility of pulling funding to the
- 29:35states based on their not getting on board, so
- 29:39to speak, with the administration's full -fledged
- 29:42efforts to promote AI inside of the United States
- 29:46could possibly reduce funding to the states.
- 29:54This is also extremely recent. There was a tax
- 29:57nuts article that came out. Gosh, I want to say
- 30:02last week that dug into how the California Department
- 30:06of Tax and Fee administration was using generative
- 30:10AI technology. And they made an announcement
- 30:14that as of August 22nd of last, yeah, last month,
- 30:19they launched their generative AI solution. This
- 30:21was the result of three contracts that were signed
- 30:25by Governor Newsom earlier in the year with a
- 30:30company called SimSoft Solutions LLC to implement
- 30:34generative AI systems inside of not just the
- 30:39CDTFA but other agencies. According to the CDTFA,
- 30:43they are using it for customer service interactions
- 30:47and essentially, You know using the internal
- 30:52data of the CDT FA only is what you know I took
- 30:58from this the customer service agents at the
- 31:01CDT FA will use the tool and to answer questions
- 31:07for customers and They will have the power of
- 31:12whether or not to Actually give the taxpayer
- 31:16on the other line or the other end of the phone
- 31:19that answer. So they will be the person, they
- 31:23will be the guardrail, so to speak, of any answers
- 31:27that the generative AI comes up with. One could
- 31:33imagine in situations where you have entry -level
- 31:37employees working in customer service positions
- 31:40at the taxing agencies, how this could be an
- 31:44extremely effective tool to helping taxpayers,
- 31:50especially unsophisticated taxpayers that can't
- 31:54hire lawyers or accountants, answer questions
- 31:58about the tax code, about laws related to the
- 32:02tax. They can't give legal advice, but they can
- 32:05give overall references to materials and forms
- 32:11and other things that can help taxpayers in an
- 32:14efficient manner, and they can do it rather quickly.
- 32:18as overall the utility of this particular tool
- 32:21inside of the customer service department at
- 32:27the CDTFA, I think is going to be a very, very
- 32:31good thing. But there are experts out there that
- 32:34are not so enthusiastic about the CDTFA's use
- 32:43of this technology. So it'll remain to be seen.
- 32:50You know how it goes. Now I want to pass it over
- 32:54to Brian to talk about what the Federation of
- 32:59Tax Administrators thinks about all this. Brian,
- 33:07you're on mute. There we go. Thank you, Travis.
- 33:16The Bar Association, thank you for inviting me.
- 33:19As indicated before, at the top of the program,
- 33:23I am general counsel with the Federation of Tax
- 33:25Administrators. Before I joined FTA, I was in
- 33:30the Maryland Attorney General's Office, was their
- 33:32general counsel, representing the Maryland Comptroller's
- 33:36Office, which is our Department of Revenue in
- 33:38the state of Maryland. For some 14 years as general
- 33:41counsel and a number of years before that, just
- 33:44among the regular soldiers of the office. As
- 33:48a caveat, when I start any of these presentations,
- 33:50because I tend to get opinionated sometimes,
- 33:53I have to say these are my opinions. They're
- 33:55not the opinions of the FTA. Should I go down
- 33:58that road? I mean, Travis gave us a nice introduction
- 34:02that points to the fact that we've all been using
- 34:05AI in a traditional, simple sense, if you will,
- 34:11for decades. Spell check. is the biggest example
- 34:15that always comes to my mind. And it's evolved,
- 34:18obviously, over the years to the point now where
- 34:20it's gotten a little bit further. So it anticipates
- 34:22what we're going to type. We're going to text
- 34:24a message or we're writing a letter. It starts
- 34:27throwing out potential text for us and those
- 34:30things. And departments of revenue similarly
- 34:33have been using AI in some format for a long
- 34:36time. Think back to the early, the 1980s and
- 34:421990s when there was what they would call rule
- 34:45-based systems and expert models that early forms
- 34:48of smart automation that would get applied, codifying
- 34:52human rules to structure problems and therefore
- 34:55filtering tax returns where you would see obvious
- 34:59anomalies for the administrations. Machine learning
- 35:03and analytics, statistical methodologies that
- 35:05allow for computers to spot patterns. Natural
- 35:09language processing, which I think Travis mentioned.
- 35:11the robotic process automation. OCRs, optical
- 35:17recognition software was a big advance, but that
- 35:20is AI. When you scan a document into your Adobe
- 35:23and it can use OCR to actually recognize words
- 35:27and text and you can search it, that's AI at
- 35:30work and the simple. Now we're into the generative
- 35:32AI. Before I jump into my slide here, the one
- 35:36thing I want to point out with when Travis was
- 35:39talking about America's A .I. action plan back
- 35:44on July 23rd. This is kind of a reaction to what
- 35:49was in H .R. 1 in the reconciliation bill. There
- 35:53was a provision in the original text of the reconciliation
- 35:56building that would have put a moratorium on
- 36:00states to do anything in the generative or the
- 36:04A .I. world in terms of regulating, putting guideposts
- 36:07in, guardrails, I think is the term we would
- 36:10use. to if you want to use those systems, here
- 36:14are the guardrails that you got to be doing.
- 36:16That got stricken from the reconciliation bill.
- 36:19I wouldn't be surprised if it doesn't come back
- 36:21in legislation down the road. But that AI action
- 36:25plan is the direct result of that. OK, we didn't
- 36:28get it in the bill. But now if you states try
- 36:30to do anything to regulate AI in any manner,
- 36:33and we're not talking about regulating it out
- 36:35so it can't be used. But just talking about some
- 36:38of the safeguards I'm going to go through in
- 36:40a moment to protect consumers, to protect private
- 36:44entities, to protect the public, to make sure
- 36:47it's done right and done well so that we can
- 36:50actually achieve the goals that we want to. These
- 36:55things just are trying to stop the states from
- 36:57trying to do those things that the federal government
- 37:00Congress is not doing. and it's of concern to
- 37:06the state. So across state tax centers administration,
- 37:11you've got multiple uses of generative AI that
- 37:14they're looking at and starting to consider.
- 37:17So like I've listed a number of them up here,
- 37:20consumer contact and service delivery. This again
- 37:22was what again, something Travis was talking
- 37:24about. You call up the, especially during tax
- 37:29season, you call your revenue agency and say,
- 37:31yeah, I've got a question. about what I can do
- 37:34and how I can do it in a form, using generative
- 37:37AI tech programs, if they're able to do it, will
- 37:42actually be able to listen to what a customer
- 37:44is asking on the other end. And it could be a
- 37:46practitioner as well as a regular taxpayer trying
- 37:49to do their own return and get them to the right
- 37:52person, gets into somebody with the knowledge
- 37:55and or access forms. explanatory documents that
- 37:59that Department of Revenue has and get it to
- 38:01the customer in real time, freeing up the human
- 38:05services, the human resources to be able to address
- 38:08more complex issues. That's a goal that's going
- 38:11to be positive for everybody. It's going to help
- 38:14the practitioners because they're going to be
- 38:16able to have, there will be people with knowledge
- 38:18that they can get through to. It's going to help
- 38:20the regular, the average taxpayer, the average
- 38:23business owner who's just got a quick question.
- 38:25How do I fill out my this month's sales and use
- 38:28tax return. And if they can get somebody to get
- 38:31some way to get a simple answer quickly, freeze
- 38:33up them, freeze up you. Staff support, knowledge
- 38:37management, same idea. This is the behind the
- 38:42scenes use of AI services to help the people
- 38:48that work at the agencies also to access the
- 38:51information. You're always onboarding new people.
- 38:54You're always advancing and moving up the chain
- 38:57up the ladder as you become more knowledgeable.
- 39:00But the ability to get the Department of Revenue
- 39:03employees to their sources of information sooner
- 39:07and quicker educates them, gets them doing the
- 39:11job right and helps again with you. Excuse me.
- 39:17Return processing. I think we all agree that
- 39:21this is one of the. the core services and functions
- 39:26of a department of revenue. Machine learning
- 39:35models can rapidly validate returns, rapidly
- 39:41check those returns with rules and filters that
- 39:45have been put into place to evaluate whether
- 39:50the returns are proper, and by using generative
- 39:53AI, they can learn from the responses they get
- 39:56back. They become better at that, maybe identifying
- 39:59other roles that can help move returns through
- 40:01processing faster, getting refunds out to taxpayers
- 40:04quicker, getting things processed, and if there's
- 40:07a question, asking it sooner of the taxpayer
- 40:10or the practitioner. Instead of waiting months,
- 40:13maybe it'll be days that you'll get something
- 40:15back. Those are the goals, those are the aspirations,
- 40:18but this is what these types of tools can do.
- 40:21Fraud detection and the shadow economy. Again,
- 40:26these tools can dig down deeper to analyze vast
- 40:32amounts of data to see the unusual patterns.
- 40:35There are third -party platforms out there that
- 40:37allow tax agencies to uncover undeclared and
- 40:42underreported income by combining those types
- 40:45of databases into what the revenue agencies are
- 40:48using. Again, revenue agencies just want everybody
- 40:52to be reporting what they're required to report
- 40:55under the law and paying what they're required
- 40:58to pay under the law. No more or no less. Just
- 41:01like that's all your clients want to have to
- 41:03do. But if we can weed out the bad actors by
- 41:06having filter systems and other detection devices,
- 41:10again, it helps the taxpayers that just want
- 41:13to do right by getting the bad apples, if you
- 41:17will, out of the way. audit compliance and enforcement,
- 41:21same idea, same thoughts as what we're looking
- 41:24at there. Collection and debt recovery. Again,
- 41:27there are platforms out there that you can tie
- 41:30into that help a revenue agency determine whether
- 41:33or not a particular taxpayer who has a debt has
- 41:37the ability to pay. And if so, what's a reasonable
- 41:42collection program. The taxpayer, again, most
- 41:44of them, they know they may owe something. They
- 41:47just don't know their ability to pay is compromised.
- 41:51And the ability to enter into a payment program,
- 41:55for example, that's going to work for both the
- 41:57agency and the taxpayer is good for both parties.
- 42:00It doesn't overburden the taxpayer and just set
- 42:02them up for failure. And to be able to access
- 42:05those other platforms to be able to create that
- 42:07information to do that is what we want to be
- 42:10able to do. Policy modeling and forecasting.
- 42:14This is, I think, very critical in the idea of
- 42:17using these generative AI tools when your legislature
- 42:21goes in and says, oh, we want to do this new
- 42:24tax. And this is how we expect, this is how we,
- 42:28the legislators, thinks it's going to work and
- 42:31the type of income and revenue it's going to
- 42:33produce and the side effects of that, the cost
- 42:36to retailers or to businesses or to individuals.
- 42:39They want to be able to model that out. These,
- 42:42again, these tools for modeling and forecasting
- 42:45that are starting to come about that are already
- 42:48on existing platforms are going to make that
- 42:51modeling and forecasting better. So again, so
- 42:54that legislature is much better informed about
- 42:56whether this is the right type of type of revenue
- 42:59raising, or maybe they want to eliminate a tax.
- 43:03And it should that be appropriate. Again, modeling
- 43:05and forecast become critical for that. And these
- 43:10new tools are going to help with all that. Data
- 43:14analytics and cross matching. We've been doing
- 43:16this for a long time within revenue agencies
- 43:18where, you know, the returns come in and you're
- 43:21matching it to W2s and 1099s. Again, as the tools
- 43:25get better, the cross matching gets better and
- 43:28helps everybody in the long run. That's what
- 43:30you want to do. And then, of course, communication
- 43:32and outreach. And this goes to, this isn't necessarily
- 43:37the call centers where you're getting calls in,
- 43:40but if you now just send a letter to a taxpayer,
- 43:44traditionally with more limited tools, it tended
- 43:47to be a script letter. It was the same letter
- 43:50that's gonna be going to Travis as it's going
- 43:52to Philip, even though that they're slightly
- 43:54different based upon their returns, but because
- 43:57they were limited in the way they could customize
- 44:00a letter. These generative tools, as the platforms
- 44:03improve and get better, are going to help revenue
- 44:07agencies be able to create more customized letters
- 44:11to go to taxpayers so that they're prepared and
- 44:13know how to respond to them or to answer questions.
- 44:17They're not going to get a generic answer. They're
- 44:19going to get something that's better and more
- 44:22directed to their actual inquiry. Travis, if
- 44:25you could knock that next slide up. So security
- 44:31and safeguards. So now these are what I'll go
- 44:33through now. What are I think you need to know
- 44:36as to what the agencies are thinking? They want
- 44:39to put these tools in place that are being told
- 44:41by their by the higher ops to other other organizations.
- 44:46You need to put these tools in place. But as
- 44:48revenue agencies, unlike some other agencies,
- 44:51there's a lot of limitations or considerations
- 44:55that have to be taken into account. when you're
- 45:00going to develop these programs. So starting
- 45:02with the security and safeguards. I'm sure folks
- 45:05here on the call are familiar with IRS publication
- 45:091075. 1075 controls how the agencies or how somebody
- 45:16is going to handle federal tax information. And
- 45:18why is that important? As I'm sure again, you
- 45:21know, tax, your State Department of Revenue has
- 45:24working agreements with the Internal Revenue
- 45:27Service for sharing of data. And when it comes
- 45:29from the IRS to the states, that's federal tax
- 45:32information or FTI, which is protected under
- 45:34multiple provisions of the Internal Revenue Code,
- 45:37can only be disclosed under certain conditions.
- 45:40And when it is disclosed to a third party, in
- 45:42this case, I'm talking about your state revenue
- 45:45agencies, there are a number of protections that
- 45:48have to be in place. These include protections
- 45:52to the access, who has access to the information,
- 45:55how the information is labeled. how it is in
- 45:59fact stored. I remember when they were doing
- 46:02audits in the Maryland Comptroller's Office and
- 46:04the IRS does that every three years or so. They
- 46:06come in and they look at how they're doing in
- 46:10safeguarding. I think there was a three lock
- 46:13rule. There had to be three locks between the
- 46:15public and FTI. And that was the lock on the
- 46:19door, the lock on the file cabinet, and maybe
- 46:21another lock into the building. Those things,
- 46:24they have to be taken into consideration, and
- 46:26we can't ignore them just because now we're using
- 46:28generative AI tools. So we have to be thinking
- 46:31about that, as well as, finally, transmission
- 46:35of the data and then disposal and destruction
- 46:37of the data. These have to be taken into consideration
- 46:41when you're going to start using these large
- 46:43language models that are going to use the data
- 46:45that Travis was talking about, the trillions
- 46:47of data. of bytes of data that your state tax
- 46:52revenue agencies have. If you're going to use
- 46:54those to inform other programs and other systems,
- 46:58these safeguards from Publication 1075 have to
- 47:01remain in place. The other thing is open government
- 47:04laws. You have open records, open meetings, FOIA.
- 47:08All these are out there. So state agencies are
- 47:12constantly getting FOIA requests that may or
- 47:16may not trigger information that is confidentiality
- 47:22laws prevent from disclosure, and you have to
- 47:24have the ability to make sure that you are protecting
- 47:28that and or editing that. So there always has
- 47:31to be a human resource, some human being in place
- 47:36to be able to review things to make sure these
- 47:38systems are working properly. Violation of these
- 47:45disclosure rules. They're criminal offenses and
- 47:48state employees can go to jail if they improperly
- 47:52disclose confidential state or federal tax information
- 47:55So they don't take these rules lightly and they're
- 47:58very concerned about it And when I was counseled
- 48:00to the comptroller's office, I was always making
- 48:03sure that our employees were not doing something
- 48:06wrong I'd rather have them come and ask questions
- 48:10to make sure they're not disclosing something.
- 48:12So now we have to think about those kind of guardrails
- 48:15in a revenue agency to be able to take advantage
- 48:19of these new tools that are out there while still
- 48:22protecting and safeguarding the information that
- 48:24is going to be used. And that comes to one of
- 48:29the things I have listed there is the HITL, which
- 48:33is just another acronym for human in the loop
- 48:37oversight. And that's what I was just talking
- 48:39about. You need to have humans in the loop to
- 48:41make sure that the things are being done properly.
- 48:47Let's say you create a digital way to screen
- 48:51for information, but you have to continuously
- 48:53be updating those screening tools to make sure,
- 48:56A, they're working, two, they're not missing
- 48:58something that they need to be revised or reworked
- 49:01in some way. You need to be able to classify
- 49:04and segregate the outputs from AI and the inputs
- 49:08from AI. So you need to be able to distinguish
- 49:10between what's sensitive taxpayer information,
- 49:13what may be not taxpayer information, but is
- 49:15personally identifiable information, like social
- 49:18security numbers, names, addresses, something
- 49:21like that. And then the administrative records,
- 49:23which are usually open to public inspection.
- 49:26You need to be able to distinguish with that.
- 49:28And that all again, like I say, these are the
- 49:30rules that are in place today. that we need to
- 49:32make sure we're continuing with Department of
- 49:34Revenues to develop these policies that stay
- 49:37in place as we move into the AI world. You also
- 49:43need to be able to account for the resource -intensive
- 49:48customization required if you use a cloud -based
- 49:52or a third -party storage system. Let's say that
- 49:56your Department of Revenue wants to now start
- 49:58stopping staying we're not going to store all
- 50:01our data on servers at our particular IT center,
- 50:05but we're going to go to a cloud -based system.
- 50:08Well, they're not going to be using and creating
- 50:10their own cloud -based system. They're probably
- 50:12going to be using it from a third -party provider.
- 50:15So they have to be very concerned about making
- 50:17sure those third -party providers have the appropriate
- 50:20security protocols in place to protect that data.
- 50:28This area right here is probably one of the defining
- 50:31challenges for the agencies, that balancing of
- 50:35security and disclosure obligations with the
- 50:39great potential that AI is going to provide.
- 50:46So, I mean, kind of I would summarize all this
- 50:50as to how you all help the departments of revenue.
- 50:54As you're hearing me describe this, these are
- 50:58human resource intensive programs and policies
- 51:02that need to be put in place. So when your legislatures,
- 51:06or whomever it is, is pushing for AI in the tax
- 51:09administration area, they have to understand
- 51:11there's a cost that's associated with that. And
- 51:14if the private sector wants to have and take
- 51:16advantage of the efficiencies that are going
- 51:19to be built in with this system when it does
- 51:21this, be partners with your departments of revenue
- 51:24to be able to make sure everybody understands
- 51:27that, that there's going to be long -term efficiencies
- 51:29and productivity that's going to come out of
- 51:31this, but there is a cost to being able to do
- 51:34it. That cost will be outweighed over the long
- 51:36run, but there are going to be upfront cost.
- 51:40Travis, if you can pop over to that last slide
- 51:42for me. These are the considerations from my
- 51:45perspective as counsel to, if I was counseled
- 51:48to a revenue agency, there's legal and ethical.
- 51:51Again, we've talked about the data privacy and
- 51:54protection and making sure those laws are being
- 51:56complied with. When I talk about intellectual
- 51:59property rights, a lot of times these AI tools
- 52:05are being provided by third party contractors
- 52:08to the agency. Something maybe they're upgrading
- 52:12their systems, maybe they have a data warehouse,
- 52:14at least that's what we call it in Maryland,
- 52:16in which data is stored. but it's their outside
- 52:19contractors, third parties that are helping to
- 52:21put these in place. And they're using the data
- 52:25in there to create as the large language model
- 52:27to create the signs to address and identify anomalies.
- 52:35Who owns that intellectual property once that
- 52:37database is created? It sounds simple. From my
- 52:42perspective, obviously, it should be the state
- 52:44owns that IP and owns that thing, that large
- 52:47language model that did that. But from the contractor
- 52:51standpoint, if not only are they using the data
- 52:54from their contract with the state of California,
- 52:57but they're using the data they're generating
- 52:58from the state of Iowa and from North Carolina,
- 53:03the larger the language models, in theory, the
- 53:06better the the identifications are going to be
- 53:10out there. So they want to be able to use that
- 53:13data also. So believe it or not, there actually
- 53:15are fights that arise as to who owns that language
- 53:19model and the database from which it was created.
- 53:23And as you can guess, that can be complicated.
- 53:26Liability and accountability. You use an AI,
- 53:30a generative AI tool. Travis calls up the desk
- 53:33and says, can I take my dog as an exemption?
- 53:36And for some reason, the generative AI says,
- 53:40absolutely, you can take that cute poodle, and
- 53:42it can now be your next dependent on your tax
- 53:45return. I'm not real familiar with California
- 53:48law, but I'm going to guess that's not allowed
- 53:50under California law. Whose liability and accountability
- 53:54is for that? If you're using a third party vendor
- 53:58that's helped you to create this, are they going
- 54:00to be the one that's going to have to Assume
- 54:04the liability for that or is that your Department
- 54:06of Revenue? My guess is that there's a compromise
- 54:10in there. Ultimately, a taxpayer is going to
- 54:13have to pay the tax because it was wrong. But
- 54:15interest and penalty, if they get assessed, that
- 54:18has to be determined who will be liable and accountable
- 54:20for that type of thing. And the states need to
- 54:23work that out with the vendors and whoever they're
- 54:26working. So if you're seeing These are issues
- 54:29that we're working on behind the scenes. You're
- 54:31never going to see it, but these are the things
- 54:32we're concerned about from the revenue department's
- 54:37perspective. The anti -discrimination laws, Travis
- 54:41talked about this. This is really the bias area.
- 54:43There are protected classes out there. There's
- 54:47a number of stories out there that you've probably
- 54:49seen in the press and the study that came out
- 54:52from the Biden administration about how Certain
- 54:57models were created, but they ended up targeting
- 54:59certain populations just because those populations
- 55:02happen to be prevalent in that particular tax
- 55:08benefit. I think of the earned income tax credit,
- 55:11for example. There's a certain population that's
- 55:15going to be taking the earned income tax credit.
- 55:18They end up getting overemphasized for audits.
- 55:21And then they start looking at zip codes. And
- 55:23they look at the zip codes where people live.
- 55:25And they start going after that. And that gets
- 55:27targeted. And that tends to be, it's dangerously
- 55:31close to targeting people because of a protected
- 55:33class as opposed to being neutral. So we have
- 55:37to be very conscious of that. And again, there
- 55:39needs to be human expertise involved in developing
- 55:43the models that are going to be inputted into
- 55:45a system so the output isn't overly concentrating
- 55:51on populations just because they end up being
- 55:54in a protected class in some manner. I talked
- 55:58about regulatory compliance. This is where simply
- 56:00making sure whatever those statutes and regulations
- 56:02that are in place are being complied with through
- 56:07the use of the AI. And again, it may be third
- 56:10-party vendors that you're using in this area.
- 56:13On the ethical side, again, there's nothing magical
- 56:16in these terms I've used here, fairness and equity.
- 56:19That's a basic trademark or under a pillar of
- 56:25any tax system is you want it to be fair and
- 56:28you want it to be equitable that everybody gets
- 56:30treated the same. Everybody has the same information,
- 56:33the same knowledge, so that they will at the
- 56:35end of the day be able to know that the rules
- 56:38that apply to me apply to you. We still need
- 56:41to make sure that's happening in this world with
- 56:43AI, that that continues. Transparency and explainability.
- 56:48Transparency is pretty obvious there. You want
- 56:49the models. If you're going to have a large language
- 56:51model that's evaluating data, you want to be
- 56:55transparent on how it's doing that evaluation.
- 56:59What are the rules that have been put in place
- 57:01through the model to evaluate that data? Explainability.
- 57:07When you want to be able to explain how it's
- 57:09happening, you want to be able to explain it
- 57:11not just for technical experts, but for people
- 57:15like myself who aren't technical experts. And
- 57:18my guess is you want it too as practitioners.
- 57:20If your client comes in and has been assessed
- 57:23and wants to fight that assessment, they want
- 57:26somebody to explain to them and you exactly what
- 57:31was it and how was it that that assessment was
- 57:34done. And if it was using one of these AI generative
- 57:37programs, they want to be able to explain how
- 57:40that happened to be able to test to make sure
- 57:42that the results are legitimate and correct.
- 57:44So you need to be able to explain that. Human
- 57:49autonomy and consent, again, this goes to making
- 57:52sure that AI, we understand that AI is there
- 57:56to augment. It's not there to undermine or replace
- 58:00the human people input into these programs. That's
- 58:04critical, I think, to making these, to fill in
- 58:09the fairness and equity piece. that human autonomy
- 58:15and consent needs to be part of it. Accountability
- 58:18and oversight, harm prevention. Again, we're
- 58:20getting back to these are the considerations
- 58:22that, again, the systems need to be accountable
- 58:25to the users. And when we talk about the users,
- 58:29let's use an audit program, for example. The
- 58:32user is not just the auditor. The user is also
- 58:34the taxpayer at the end of the day who is going
- 58:37to see the results of that audit. The systems
- 58:40need to be accountable to the users, whomever
- 58:43they may be. There needs to be that oversight,
- 58:46because this is an iterative process. You may
- 58:48have a program today that uses the parameters
- 58:521, 2, and 3. And as it uses those, we need to
- 58:55go back to make sure those parameters 1, 2, and
- 58:573 are appropriate. And we shouldn't be adding
- 59:00x, y, and z to that. That's where we are. That's
- 59:04what the Departments of Revenue are taking into
- 59:06consideration as they try to implement these
- 59:08programs that are being asked by them by all
- 59:11the heads of agencies, multiple players in the
- 59:16environment. I'm not sure if we had any questions
- 59:22out there, but there's two in there at least.
- 59:29I don't see any. If you click on the Q &A, somebody
- 59:33asks, is it really the low and middle income
- 59:35taxpayer that is targeted with AI review, since
- 59:38high income taxpayers have a more complicated
- 59:40return? They're all target. Targeting is not
- 59:43a word that I like to use. We're not targeting
- 59:46them. The systems look for anomalies and returns
- 59:51and identifies those anomalies. And then if that
- 59:55is anomalies, suggests that there's a problem
- 1:00:00with the return, the return will be questioned.
- 1:00:03Is it just low and middle income tax payers?
- 1:00:05Not necessarily, but depending on the type of
- 1:00:08tax that you're talking about, it could be. So
- 1:00:12yes, I think it's MIT or one of the larger institutions
- 1:00:19that do these types of evaluations have noted
- 1:00:23where there is bias in income taxes administration
- 1:00:26that you need to be aware of. Yeah, that study
- 1:00:29was done by Stanford. Thank you. And I was in
- 1:00:36the room. It was actually North Carolina. It
- 1:00:38was a very specific state, a very specific tax
- 1:00:44credit that they looked at because at the time,
- 1:00:47North Carolina was the only state that actually
- 1:00:51tracked race data. Now they don't track that
- 1:00:57anymore. And obviously the IRS doesn't track
- 1:01:00it, and the other states don't track it either.
- 1:01:02And so the reason Stanford used that state was
- 1:01:07because of that. It was because they were able
- 1:01:09to at least get some quantifiable data about
- 1:01:12race in that study. You're right, earned income
- 1:01:17tax credit. What the study found was that, yeah,
- 1:01:23because of the credit itself, Yeah, I mean, there's
- 1:01:35income parameters on the credit which are naturally
- 1:01:38going to be looking at certain. It was a natural
- 1:01:41occurrence of bias because of who took the credit.
- 1:01:45I can tell you anecdotally I was in the room
- 1:01:48after this study came out with a bunch of IRS
- 1:01:51personnel when we were talking about this study
- 1:01:54in Washington DC. and they were very upset at
- 1:01:59this particular study just because, you know,
- 1:02:04they themselves felt like they weren't part of
- 1:02:08any type of bias in their enforcement efforts.
- 1:02:13I think the study was meant to show that sometimes
- 1:02:17there's unintended bias whether or not it's intentional
- 1:02:21or not. And so I want to be very clear here.
- 1:02:26It was a very specific study done by Stanford
- 1:02:30that was highlighted. The commissioner at the
- 1:02:34time definitely confirmed it to the Senate Finance
- 1:02:38Committee that it had been occurring, but it
- 1:02:42was very specific. And now as it relates to race
- 1:02:45data, I think it would be very difficult to determine
- 1:02:50bias based off that because no states are measuring
- 1:02:53it and the federal government isn't either. So
- 1:02:58I want to be very clear about that. I would agree
- 1:03:03with you and I think it can be taken unfairly
- 1:03:06but nonetheless it is something that is of concern.
- 1:03:11There's one other question in the discussion
- 1:03:14there from Ruth Rollett. How seriously can we
- 1:03:18take all this supposed security when the entire
- 1:03:22Social Security Database was uploaded off the
- 1:03:25SSA, from the Social Security Database. That's
- 1:03:29separate. That's a different story. I personally,
- 1:03:34from the State Tax Administration perspective,
- 1:03:37we are extremely conscious of and serious about
- 1:03:41security and protecting taxpayer information.
- 1:03:47Brian Alliner. is very supportive, at least of
- 1:03:50the two lawsuits that are out there trying to
- 1:03:52protect and prevent the sharing of data with
- 1:03:56the Department of Homeland Security. And if for
- 1:04:01no other reason, there's two reasons. One, as
- 1:04:06tax administrators, we've lived on a promise
- 1:04:09that was made to taxpayers that your data is
- 1:04:13safe. And that feeds into the second reason,
- 1:04:16because if your tax information isn't safe, our
- 1:04:20voluntary compliance system, where we rely on
- 1:04:25voluntary compliance for people to file their
- 1:04:26returns, is seriously in jeopardy. So from a
- 1:04:31state tax administrator's perspective, we take
- 1:04:34it seriously. What they're doing at the federal
- 1:04:37level is beyond us. We can't control it. But
- 1:04:39at the state level, we are constantly talking
- 1:04:42about how to make sure that state tax information
- 1:04:45stays secure and doesn't get compromised. Yeah,
- 1:04:49I'll speak to the Fed side, because I think what
- 1:04:52you... Look, Ruth, I can't answer that effectively
- 1:04:56either. What happened, you saw, and when the
- 1:05:00executive orders came out, what happened when
- 1:05:02the current administration came in was they immediately
- 1:05:06started... focusing on AI as it related to the
- 1:05:10various agencies. I cannot tell you in all certain
- 1:05:15terms exactly what's going on with the privacy
- 1:05:18aspect of taxpayer data right now. I can't because
- 1:05:23there's been so much turnover. It's been hard
- 1:05:25to figure out who to speak with, who to invite
- 1:05:28to panels or whether or not they're even invited
- 1:05:30in the future. I will say this. I think there
- 1:05:35has been a large subset of people that have joined
- 1:05:39the efforts of the current administration and
- 1:05:42learned about the law that they didn't know about,
- 1:05:4526 USC 6103, that when they got on board and
- 1:05:50started, you know, looking at taxpayer data and
- 1:05:54have been asked to share certain data, they've
- 1:05:57realized that if they work for the IRS, that's
- 1:05:59a crime. And in doing so, I think a lot of them
- 1:06:04have stepped back. I think a lot of people have
- 1:06:08rejected the requests because of 6103. And I
- 1:06:15think that's why we're starting to see so much
- 1:06:17turnover at the commissioner level. So while
- 1:06:20I cannot speak, you know, intelligently about
- 1:06:25the privacy efforts of the IRS, I can tell you
- 1:06:286103 still exists. And if taxpayer data is shared
- 1:06:32by an IRS employee, it's a crime. So, not to
- 1:06:39end on a low note, but I really enjoyed speaking
- 1:06:44with everybody today. And there are more questions,
- 1:06:47please filter them through the bar association,
- 1:06:51or you can find me on the Fenimore website. But
- 1:06:54thank you so much, and I look forward to speaking
- 1:06:58with you again soon. Thank you all. Appreciate
- 1:07:01the time.