Tax Notes logo

Artificial Intelligence Isn't Here Yet, But It's Already Changing Tax: Transcript

Posted on Dec. 16, 2020

Information technology hasn't yet advanced to the point of "true" artificial intelligence, but the combination of huge datasets and increased computing power is nevertheless transforming the practice of tax law. In a December 9 webinar, "Taxing Issues: Artificial Intelligence and the Future of Tax," Benjamin Alarie of Blue J Legal, Jeffrey Saviano of EY, and Sarah Lawsky of Northwestern University Pritzker School of Law discussed how technological shifts are shaping the field of tax and leading to new breakthroughs in statutory analysis, decision making, and other aspects of tax.

Tax Analysts President and CEO Cara L. Griffith moderated the discussion, which can be viewed on Youtube. A transcript is below.

Cara Griffith: Welcome, everyone. I'm Cara Griffith, the president and CEO of Tax Analysts. I’m so pleased you’ve joined us today for what I know will be an interesting and informative discussion on the use of artificial intelligence in tax. AI has the potential to improve taxpayer compliance, make tax code enforcement more efficient, and even help practitioners address legal matters by predicting outcomes. It truly is a brave new world. We are thrilled once again to have a large audience, not only from all over the country, but also from around the world. As we all know, technology has no borders. I really look forward to the discussion today. I have a lot to learn about various forms of AI and how they can be used in tax. I also look forward to your questions, and I guarantee you this panel is well-qualified to discuss and answer them. Today's event is the sixth in Tax Analysts’ new series of public discussions that we call “Taxing Issues.”

We launched the series as part of our 50th anniversary celebration. And through it, we are bringing the tax community together for bipartisan discussions on the future of tax policy with leading policymakers and experts. And as some of you have heard me say before, with COVID-19 continuing to wreak havoc all across our country, we will hold these discussions in a virtual format for the foreseeable future, but we welcome your feedback on how we can make them more interactive. You can send your thoughts to, and now on to the topic at hand. Today, we're moving outside of traditional tax policy to explore the world of AI and how advanced technology can be used to make us more efficient and effective. Innovation is a word that we hear a lot these days, even in fields like law and tax, which historically have not been cutting edge. Although we captioned today's discussion as “AI in Tax,” the term AI is quite broad, and it has several components.

AI is also a little bit of a misnomer in that there isn't true intelligence. AI refers to a computer that can perform human-like functions. With AI, a computer system can take in large amounts of data, establish relationships between the data, or create algorithms that help it determine the right answer or the best way to accomplish a task. Also under the AI umbrella are things like machine learning, which is like it sounds: the act of teaching computers to be progressively more accurate with the task. It's a logical, condition-oriented thought process in which the computer quote-unquote learns by detecting data patterns and relationships. There's also robotic automation. This is the more traditional type of computer programming that we're much more familiar with. The computer is programmed that if A is true, then do X; if A is not true, then do Y, but within all of this, there are many variations.

There's a large variety of applications that all use some form of AI, but how does AI apply to the tax community, and are tax practitioners and administrators ready to use AI to solve the problems? It depends on who you ask. Automation has been used for years in tax preparation, but when it comes to analyzing legal questions and accurately predicting answers, practitioners may be somewhat hesitant to jump on board, but AI and, more specifically, innovation is coming fast and furious. So we all need to better understand the benefits, the opportunities, and the risks. So to help us sort this out, we have three outstanding speakers, and I'm going to introduce them in the order in which they will speak. I've asked them each to speak for about five minutes, and when they have finished, I'll begin a discussion with a question or two. As I noted at the outset, we welcome your questions.

Thank you to those who emailed questions in advance. And during the event today, please use the chat feature to submit your questions. I'll get to as many of them as time permits. We will hear first from Benjamin Alarie, co-founder and CEO of Blue J Legal, a company that uses machine learning and AI to predict legal outcomes. Ben is also a law professor at the University of Toronto. Next we'll hear from Jeff Saviano, EY’s global tax innovation leader. Jeff has the monumental task of driving innovation to help the firm find and commercialize new tax services and technology forms. Finally, we'll hear from Sarah Lawsky, a law professor at Northwestern School of Law and the associate dean of academic programs. Sarah has taught a multitude of tax courses, and her research has focused on tax law and how to apply formal logic and AI to the law. I truly couldn't ask for a better panel to help us lead through these discussions. I'm really looking forward to it today. And so, Ben, I turn that over to you.

Benjamin Alarie: Phenomenal. Thanks, Cara. It's really my pleasure to be joining the discussion today. And I'm going to spend my few minutes talking a little bit about the history of where tax technology has been in terms of research, talking about a little bit where it is today and how we're using analytics to analyze tax positions at Blue J. And then I'm going to speculate a little bit about where things are likely to go next in the future. So first take a few steps back, then talk about the present approach using technology, and then speculate a little bit about the future. I think there's a slide that we could bring up that nicely illustrates where we've been and where I think we are going. In around 1800 — in 1798 to be precise, 222 years to this month actually — William Pitt the Younger introduced the first income tax in England.

And of course, that was a paper-based system. All of the legislation was based on paper. Fast-forward more than a hundred years and the first U.S. federal income tax — also paper-based of course — a completely analog system. And the income tax has, of course, been predominantly paper-based for the majority of its existence. But around the 1960s, 1970s, we started to see the emergence of the use of digital technology to reproduce what used to be on paper, and for a long time was exclusively on paper, into electronic databases. And so this was the move from analog to digital, and it was a huge undertaking to move the entire substrate of the law from paper into bits and bytes on computers. I think now 50 years on if you plant the flag at 1970, 50 years on, I think we can say that's largely been successful. It's really happened. We have made the transition from words on paper to words predominantly on screens.

Of course, many people like to print out what they see on their screen. I know I do because paper has some advantages over seeing things on the screen, but I would say that transition is largely done. What's happening next is a movement that kicked off really around 2015, where we're seeing the emergence of computational methods for figuring out the content of the law and how it actually applies to particular tax situations. And so I think this represents another step change, a really important change in the substrate of how we think about tax law and how we conduct research in tax law. And so this 2015 date coincides with the founding of Blue J Legal. And I think it is an important day. It really reflects that time when we started to see interest in using machine learning to answer practical tax questions.

I think it's time for the next slide, please. On the next slide, you can get a sense of the process by which you can use machine learning. And in this case, it's supervised machine learning to engineer a solution to a particular tax question. So what we're talking about here is identifying a particular legal situation or tax question that's predominantly based on facts and circumstances where it's easy to identify the issue or the question; it's very difficult to answer that question. So what I have in mind are examples, like does the economic substance doctrine apply to a particular case, or is a particular financing instrument going to be treated as debt or equity for tax purposes, or finally, is a worker an employee or an independent contractor? And there are dozens of these characterization questions throughout tax law, not just the income tax law, but also in state and local taxes as well.

And so what you can do is collect all of the rulings, all of the cases, and these become training data for teaching algorithms how to characterize new situations. And so this actually becomes an iterative thing over time. You collect hundreds and hundreds, potentially thousands, of examples about how courts have ruled on particular situations. Those facts and circumstances of each one of those particular cases becomes a training example. And you can train an algorithm to pattern match: identify the patterns implicit in those facts and circumstances, leading to particular outcomes in all of those different cases. And what that means is you are now able to scenario test a new set of facts and circumstances trained on those examples and actually get a predicted outcome.

So if we move to the next slide, please. This is really powerful. So this is where we are at now; this already exists. And you can use this kind of technology to test dozens of different kinds of situations that previously would have been very ambiguous and would have required a ton of time to go through and manually locate through keyword-searching potential rulings or cases or decisions that might affect your thinking about the situation. Now you can synthesize all of those materials very, very quickly — that a system will elicit the facts from you, run it through the algorithm, and give you a prediction of what the likely outcome would be along with its score on the likelihood of the situation being characterized one way or the other way: a list of rulings where the facts and circumstances were similar to the facts and circumstances that you're running with your scenario. And also it'll give you a plain language explanation of why that's the outcome. So this exists right now.

In the one minute that remains, I'm going to talk about where I think things are going. And so this has already advanced fairly significantly over the past five years. I think we're going to see even more accelerated change going forward. I think we're going to see tools that address the thicket of very intricate rules in the code and in the regs emerging. And I think Sarah will have something to say about the power of those kinds of approaches to understanding tax law in a few moments. I think we're going to see systems that learn from the structure of the code and the regs, but also learn from the behavior of researchers who are interacting with the code and the regs to influence how you should approach your own tax research problems.

And I think in the long run, we're going to see really rich conceptual models of law, of entities, of businesses, of people, of individuals, of nonprofits, of other organizations, which will then interact with machine learning systems trained on all of the law to allow for really fast experimentation in policymaking. And I think that's going to lead to very rapid development of the law in a major, major complexification of the law, which is going to actually reinforce the necessity for all of these technological tools necessary to cope with this complexity. So I'm going to pause here and turn it back to Cara.

Griffith: And I have to tell you it is absolutely fascinating, and I'm astonished by what you've been able to do with Blue J. It's really quite impressive. Jeff, let's turn it over to you for a few minutes.

Jeff Saviano: Thank you. Well, first thank you for inviting me to the panel today, and it's been so great to get to know this group in our preparation. What an exciting topic. And it's so great going after you, Ben. Ben and I spent some time together just a year ago in our office. We hosted the local TEDx organization, TEDxBeaconStreet. And we can send around links to the audience today, but you'll love the TEDx talk that Ben delivered about a year ago on this topic. And we were so thrilled to have you as part of that program. This is really an exciting time. I want to talk about some of the practical applications of AI solving tax problems today, and then I'll touch on some of the geopolitical and governance issues that we see because what is different about tax from other disciplines is the necessary integration of the public and private sectors — that the laws come from governments, and they set the policies and administer the laws, and, of course, taxpayers, whether they’re business or individuals, in compliance. And in many cases the taxpayer-tax authority relationship defines those citizen-government relationships. So I want to touch a bit on how that comes together to actually proliferate these systems.

But first, how is AI being used in tax today to solve issues? Well, you're probably using it and not even realizing that you are is one point I’d like to make. There’s a great Gartner quote in research that they estimate that 40 percent of all enterprise-wide technology applications by 2021 — that's next year, 2021 — will include AI capabilities within it. And just as Ben said, that is going to continue to aggressively grow in the coming years. From a tax perspective, we see taxes being used by governments to fight fraud, using it for anomaly detection, and how to parse incredible amounts of data to find the fraudsters and help prevent fraud from actually happening.

And we saw, unfortunately, as part of the stimulus in the U.S., tremendous attempts to defraud the government from stimulus payments. AI is a great tool for governments to use to actually find those data anomalies. We’re seeing it used to continue with some of the trends from government, better communication with taxpayers. Some governments, like the Australian tax office, has been a leader in using conversational agents, chat bots, in order to take the more common questions that taxpayers have upon calling authorities and looking for information. We're seeing it used to communicate better, to determine a more targeted way to communicate with taxpayers, and how important that is. I also agree with Ben on the importance of classification and prediction. As tax people, we think everything we do sometimes is really, really complicated, but it's amazing how you can break down a significant component of — what tax people do is we classify things. So much of the law is binary. Is it deductible or not deductible? Is the worker an employee or an independent contractor?

The machines’ AI systems are actually better than humans at parsing tremendous amounts of data to determine and make a prediction for one of those binary elements of the law. We did a bit of an informal study at EY when we launched our AI capabilities. We launched a lab based in Cambridge, and we do so much in collaboration with MIT and the Connection Science Lab. And when we, rather unscientifically, looked at how do we believe AI can most impact tax, one of the things we found is that as tax people, we read a lot, and we read a lot of the same documents, and we're extracting information from those documents that's very relevant for compliance and controversy in planning. So we think that natural language processing tools to extract information from documents and use it in a structured way to solve tax problems is one of the areas in the very foreseeable future where AI will play a significant role.

Now, just to give one example of that, we had the commissioner from one of the African nations in our lab and was talking about an effort to digitize thousands of tax returns over many years by just retyping and digitizing them manually. And that is a great problem for AI to solve. Those days as a young associate, thrown into a room to read hundreds of sales contracts and find the same six pieces of data — the machines, thankfully, are better at that than we are. I also just want to highlight some of these governance issues. One of the trends that we're following in tax technology is that we're very quickly entering a phase of multi-stakeholder platforms. I mentioned some of the public and private sector. And that's what makes tax so interesting to many of us is that we think these technology systems that have been developed one-to-one — develop technology for a government or for a corporate taxpayer — we're very quickly, and I think even accelerated coming out of the pandemic, [seeing] new multi-stakeholder platforms where multiple parties are convened in order to transmit data and AI, where learning will be of great assistance in not just capturing data, but the calculations and reporting mechanisms through those systems.

The governance of all of that, to have the change that we would love to see at scale, to come out of this pandemic that we're in and have new digital systems around the world, will require new governance systems that frankly just don't exist today. And I think an opportunity for multilateral organizations and, you know, great universities like represented today with Northwestern University and Toronto and the private sector and governments together finding some of these solutions. And we spent quite a bit of time on these new governance models to effect big change like this in the world. And so, excited to talk about more of that today, too.

Griffith: Thanks, Jeff. I have to tell you, EY has really been on the forefront in terms of innovation, and it's quite impressive, especially considering the size of the organization. It's not easy to get an organization to jump on the innovation bandwagon, and then the clients who deal with them. And then you made some great points there. And I am anxiously writing down some questions here, but first let's hear from Sarah. And I know that she's got some very interesting points as well.

Sarah Lawsky: Great. Thank you so much, Cara. So I want to start out by actually emphasizing a point that you made at the beginning when you put the caveat. So the artificial intelligence, it's not really intelligence, and I can't stress that enough. This is not intelligence. Maybe someday there will be developed actual artificial intelligence that is of some type of general intelligence, but that's not what we have right now. And we use “machine learning.” Nobody's learning, nobody's learning. It's just a machine. Nobody's learning. It's “deep learning.” It's not deep. They're talking about the structure of the underlying program. So this is not, you know, a robots coming for your job type of situation. What is going on? There's a couple things going on. There's, roughly speaking, we can divide what's happening into two groups, right? And one group is things like machine learning, which involves taking very large amounts of data and processing those large amounts of data very quickly.

And the advances on this front, on the ability to take large amounts of data and process data quickly, the advances that we're seeing in the ability to do that is enormous. And that's where we're seeing most of these changes. Most of the theory, not all of the theory, but most of the theory that's being used has been around for quite some time, developed in the 1970s or 80s. What do we have now? We have the ability to — these very fast processors that can really process this data quickly. That's a big change. That's a big change. So yeah, computers are better at taking a very large amount of information and slicing and dicing it very quickly. But you know, it's not intelligence, nothing deep. So, and then to emphasize something that Jeff said, and you’re using it all the time, right? You're using it all the time. Did you use your smartphone today? You're using it all the time. Did you search for something? All of this stuff is coming into play in all of our lives all the time.

So that it’s sort of a caveat in to say that sort of, one side, both the caveat, but also that there's one side is this big data side. The other thing that computers are really good at doing is they're good at taking rules, and they're good at processing those rules quickly. Tax law is different than other kinds of law, most other kinds of law in the United States, because the questions like the questions that Ben mentioned, these questions that are really facts and circumstances questions that are developed really through, you know, you say common-law reasoning looking at sort of analogic reasoning. That's a large part of what we do, but as we know, a large part of what we do also is take rules — and that are complex rules — and apply a lot of rules. And that's something that computers are really good at.

So this is I think the powerful connection between rules, rule-based reasoning, and then combining the rule-based reasoning with enormous amounts of basically data processing, looking for those patterns. That together is extremely powerful, and because of the tax law and the way that tax law exists, it's a pragmatic statement. It's not something intrinsic about tax law. Tax law, just, it has a lot of rules. There's a lot of statutes. There's a lot of regulation. There's a lot of guidance. It's very susceptible to being analyzed, using rule-based legal reasoning, combined with massive amounts of data. And then when we think about the massive amounts of data, we want to think about the different kinds of data that we're looking at, right? So on the one hand, you know, Ben's company might be looking through large amounts of court cases along with statutes and rulings and looking for those facts and circumstances.

When we're talking about compliance, we're talking about enormous amounts of data that taxpayers provide, right? We can sort through enormous amounts of data provided by taxpayers and look for, as Jeff mentioned, for fraud or other, other patterns there. These are, and the statutes themselves are, probably a number of different groups have given it a shot. I've worked with one group as well to give it a shot. Statutes themselves are not really that susceptible to natural language processing. There's not enough of them. I did not — I know. You did not wake up this morning and think, “I'm going to go to a panel where someone's going to say there's not enough tax statute. There's not enough tax regulations,” but here I'm saying it to you because there's really not enough to do the kind of massive training that you need to do.

So this is a situation where I think if we can get some of that — the tax rules, statutes — formalized, then that will make the connection — made by hand; people can do it. It's, I mean, it's, you know, it's pretty big. It's not, not that big. We can do that. Then working together with the types of massive types of data analysis with the speed that that can happen now, it's a very, very powerful combination that I think when, when people talk about AI, so here's what I would say when people talk about AI: It's too much, ‘cause it's not really intelligence, but it's also too narrow because often people are only talking about the big data side of things when, in fact, when you take into account all of the things that computers can do, and we join together some formalization of the more complex rules and regulations that we have with the big data processing, that's potentially extremely powerful. With that, I turn it back to you, Cara. Will you maybe take some questions?

Griffith: So what I should take from that is that one, AI is not scary, even though I envisioned AI as the computer mainframe in the sci-fi movie where they're all on board the spacecraft, and now they will never get back to earth because the computer mainframe has gone on the road. It's not that, and we all need to kind of take a step back from that, but so a question that I'm going to pose to all three of you now is how do people get started? So when we were talking in our planning, and Jeff, I think it was you who said that 97 percent of companies have not started looking into AI, even though they may be using it on a daily basis, just in terms of the applications that they use on Google and things like that. How do these companies, practitioners, government agencies, how do people get started and walk down the path into AI that can be useful for them?

Saviano: Yeah, it is a great question. And it's one that we get a lot. And I think that it may still be nascent for accounting and finance and tax within companies or within governments, but it's not so young in other fields and other applications. And so it helps sometimes that we're following other fields. For example, you know, as a member of the human race, I'm glad that we're, that the really smart people have been using AI to solve medical problems and making great advances at that. I think that's wonderful, but it's time. And we're seeing finally that there's some application to tax. I think one way to get started would be to test certain systems that are available. I think we're all lucky that Ben is on this panel, and you'll hear more about Blue J Legal and opportunity to use tools from companies like Ben’s and to actually test the technology with real tax problems that you have today.

There's other third-party applications that are available today. There are workbenches available at some of the large technology companies like Microsoft and others that are available to most of, many of the corporates that are participating today and other large companies. And so there are many opportunities now, Cara, to get started by leveraging systems that already exist today. And if you're a little bit more adventurous, there are ways to find a bespoke solution to some problems that you have. And there's lots of places to go for that. So I do think that there's more opportunities now than ever to find tools that exist in the world and and actually solve some problems that you have as a tax professional.

Lawsky: I would just build on that a little bit. I think Jeff's last point there, that to solve problems that you have, right? So they're your problems. Just like, OK, I would say that I don't do this, but I mean, let's be honest. We kind of probably all do. I say you should not go to the hardware store and pick out a tool that looks great and come home and look around your house and see where you can use the tool. And I mean, you might do that. It's OK if you do that; I sometimes have been known to do that. But probably if you're going to be doing this for your work and spending a lot of money, that's not the way forward. I sort of see this as an iterative process which is these tools are here to solve real problems that you have or that your clients have. Now, how do you identify those problems? You have to know a little bit about what the tools can do.

So you have to learn a little bit about what the tools can do and then go away and think about what are my problems. You don't have to — someone's coming to sell you something. You don't, no. What is your actual problems? What are your client's actual problems? And then what problems maybe are susceptible to the kinds of tools that are available. So, you know, you can say what's something where I'm dealing with a lot of information and it would be easier if the information could be processed, you know, by some — OK, oh, here's something. Now, let me see what is there that exists that can process this particular type of information. But not getting sold on something just because it's like a, you know, it's a shiny tool, OK, it's a shiny tool, but what is your problem? What is your actual problem? What is your client's actual problem? How can someone help you with that actual problem? How can these tools help you and your client?

Griffith: Ben, you’ve got this. I need tools, so talk to us about Blue J and about how that could be useful to a multitude of different types of taxpayers and practitioners.

Alarie: Sure. I think I might start by saying that I agree with what Jeff said, with what Sarah said. I think the onus is really on the providers of these kinds of tools to make them relevant to the real problems that people are experiencing in the same way. That, you know, we've seen this technology, and the point was made earlier on this panel that we've seen machine learning, technology, artificial intelligence, insinuate itself into our lives in a very subtle and gradual way. So far in the way it's done that is by solving real problems that we have. We wake up, we have a question, we type it into a search field and we look at the results. That's a real problem, I was craving some information, and that's how I'm finding my answer.

And so I think a lot of the onus is on the companies, the other providers of these tools, to make them really, really, user-friendly and make them a real solution to real problems that folks are facing. And so, with respect to the Blue J Tax solution that you asked about, this is a solution to the kinds of facts and circumstances problems that crop up routinely in tax practice, whether it's low-income taxpayer clinics or making Blue J Tax available to low-income taxpayer clinics right across the country. A number of large firms are piloting the software already and are using the software up in Canada. It's in widespread use among the top firms and within government to solve these kinds of problems that come up in tax administration and in tax practice, both in accounting and in law firms, and big corporations are adopting it too to get clarity into the various tax positions that they're looking to analyze in their tax function.

So one of the ways that I would suggest getting started is taking one of these tools for a free trial. Providers of these tools want you to try them on, try them out on real problems that you have in your tax practice. And so one of the things that we'll do in the email that goes out following the webinar, for everyone who's registered, is make available a pilot of Blue J Tax for 90 days on a complimentary basis so that you can actually go through that exercise of analyzing some of your own tax problems using the software. And I think the onus really is on the providers of the software to make sure that it's a solution to the tax problems, those tax questions that you're coming up against in your practice, in your professional life.

Griffith: Ben, I'm going to follow up with you because we had a question that came in from a [inaudible]. Research has shown that natural language processing models fail to extract anything relevant from statutes. And Sarah, you mentioned this in your comments. How do you expect to overcome these types of limitations?

Alarie: Yeah, it's a great question. I think the technology is improving all the time. So many folks will, if you've been paying attention to some of the headlines — maybe it's just my curated social media feed, but what I see are a bunch of references to GPT-3 and the work of open AI and improving natural language processing models and generative models to work with tax. And I think it's already doing really cool stuff. I think, again, this is the state-of-the-art now; these things are all improving at pace. And increasingly there are legions of very sophisticated people spending a lot of time improving the underlying technology. So I share Sarah's reservations around not having unlimited statutory texts to train these models on. But I'm also not prepared to say that we'll never get there because I think there's a good chance that if we, as humans, can make sense out of these materials, then you can train an algorithm to extract the similar information from these systems.

I think in the interim between now and when the technology is powerful enough to do it in a fully automated way, you can overcome a lot of those shortcomings by brute force: introducing human labor, smart human labor, into data processing, structuring, labeling, and you can create a system that provides those affordances, like does those things that you want the system to do. It doesn't have to be a purely technology-driven solution. Sometimes it's 90 percent technology and 10 percent sweat of the brow to produce a really compelling resolve. And this is about marrying, you know, facts and circumstances, standards-based tests, which certainly play a role in tax matters, with the rules-based analysis, that really you can extract the underlying logic from the Internal Revenue Code if you have dedication and perseverance and you're able to plow through the many hundreds of thousands of words that are there for inspection and structuring.

Lawsky: If I could just build on that as well. There are — the code is pretty big, but a lot of times the issue is the legal issues that are being looked at don't draw on the entire code or regulations. So for example, if, say, Ben is looking at independent contractor versus employee, the amount of law in the statute and regulations that is relevant is limited. And you can go ahead and hand-formalize that though. There's some interesting issues that come up with hand-formalizing the statues. I don't consider it to be a brute forcing issue. I think it actually is something that requires a lot of thought, requires a lot of care.

But we don't have to formalize the entire code and regulations to be able to have the statute and regulations that we need to address any particular question. This is the idea of joining together the two types of technology. If you're looking at issues about independent contractor versus employee, there's a stack of cases; there's a stack of rulings. There's not that much in the statute and regulations that you can formalize yourself. The issue of how to formalize it is an interesting question. I'm working with a group of computer scientists on programming language that's designed — the structure of the language is based on the structure of the statute. The way the statute is written in the United States, and in France it turns out, is general rules followed by exceptions. So it's a logic that is a logic that is of general rules with exceptions. And we're working on creating a programming language that is designed to formalize statutes. And then you have people, you do pair programming, you sit down with a lawyer and a programmer and you formalize the statute. You maybe in five years or in 10 years, we'll have natural language processing that can do it. Guess what? In five or 10 years, maybe we're done doing it, right? We don't have to wait. And so it's a mix. It's a mix of the different; it's iterative. It's a mix of the different kinds of technologies.

Saviano: It’s breakthroughs. It is such an important point, Sarah, and you look at just what we've been talking about: the breakthroughs in statutory analysis with using artificial intelligence. But if you look across the breadth of what we do as tax people, you know, statutory analysis is but one piece of it at the core. Look at how AI can help all of us as tax people. It's making better decisions. We want to make better decisions. And it could be as simple as a decision as to what's the number that goes on line 15 of the tax return. And that's where we're finding, especially with structured data and back to this question about how to get started. With the tremendous amounts of data that's available to us as tax people, much of that is using the structured data to help make better decisions in compliance or in other aspects of tax controversies that we're finding lots of application there. This data sourcing issue, though, is so critical that we've mentioned. I think all of us have mentioned the need for identifying the right data streams to train models. And it is a limitation. It can be a limitation. It has pushed us to develop synthetic data generation capabilities.

And if you can't get real data to drive your models, or if you don't have legal authority to use the data that you'd like to use, then we're seeing opportunities to create your own data called synthetic data to build models based off of those synthetic data streams. And I think it's an important area for all of us to think about because we're only going to see greater regulatory and legal restrictions on data that we may want to use to develop our machine-learning models. And I think that the opportunity to create and have a synthetic data generation ability in the AI world is going to be important too.

Lawsky: Cara, can I make one comment about data since we’re mentioning data? This is just something that always — I think this is being focused on more and more throughout machine learning. And I think it's worth mentioning in this context, too, is that if the data builds in bias, then bias comes out the other end and the question of the ethics around big data. I just want to mention it. We don’t have to focus on it now, but I think we would be remiss if we talked about these big data models and didn't remember that there is a significant field of study. That's a growing field that studies the ethics of big data and bias that is within the data. And that then will produce at the other end results that are potentially problematic and build in that exact same bias, and it's just important to flag that. And when you're thinking about your data sources, thinking about where did the data itself come from and what kind of bias is in there, there's way more in this field than we can discuss in 20 minutes, but I just wanted to mention that as something that people, you know, as you're working on your data and bringing together that data, and where are you getting the data, what is the source of that data, that large field of ethics around that it's important to keep in mind, I think.

Alarie: Yeah. And maybe if I can jump in on that and add, I think this is why it's so critical that a lot of these solutions build in the ability for humans to make the final call. So as Jeff was saying, you know, you want to make a better decision as a professional. If you're thinking about it from the perspective of a member of the judiciary, if you have in front of you a prediction based on a thousand past precedents on a particular issue and a model trained on those 1,000 past precedents says, well, you know, it's really an 80 percent chance that based on all of these past judgments, a representative judge drawn out of this pool would have said that, you know, party X wins and not party Y wins. And if you feel as the judge that that seems like the wrong side of history to find in favor of X rather than Y, it's important that that judge has the opportunity to say, “Actually, I think X is not the right answer anymore. I recognize that past precedent suggests that it is the better answer on these facts, but let me put together the normative case for finding Y, and I'm going to find Y.” And it's important that the law then is able to evolve and change and counter some of those biases that can be surfaced through the use of machine learning. I think that's extraordinarily important to bear in mind.

And so one of the key things that, that we've always done at Blue J is have that in mind as the use case. And so we're making a prediction. We're cognizant that it's based on past cases and the past is not always a perfect predictor of the future. And sometimes you want things to change when you recognize how things have been done, and you actually prefer for the system to evolve and change. And so I think Sarah is exactly right in making that point. And it's why you want to have a lot of these systems open for further learning and further evolution through the instrumentality of humans mixing their good judgment with the synthetic abilities of these machine-learning models.

Griffith: I have to tell you, it's fascinating. I had a bunch of questions on data, ‘cause there's a large amount of data that has to be acquired and where do you get it? And then how do you protect the privacy of it? And, you know, there's a whole host of concerns. I hadn't even gone down the bias concept, but absolutely. You know, if you have garbage in, you get garbage out, and you determine what that garbage is. So I'm going to try and quit to a question from a viewer. Ben, I think that this one is aimed at you, but everyone should feel free to respond. The question is, “Are there any applications for AI with estate planning, wills, and trusts, drafting administration, both in probate court and outside of probate court?”

Alarie: So that's a good question. What I can say is Blue J Tax doesn't have those capabilities at the moment. But I'm not an expert on all the different solutions that are out there. It's quite possible that there are solutions out there. I don't know of any off the top of my head that would address that specific context, wills and estates and probate issues. But I expect it's only a matter of time. As this technology really picks up, we should expect a doubling of the coverage every year of these tools. And it's the power of doubling that, you know, really over not a long period of time will actually result in really, really significant improvements in coverage. So if you double every year, year-on-year for 10 years, you have over a thousand times more coverage. And so I'd expect, with doubling year-over-year, those kinds of solutions are going to be in place by the end of the 2020s. I don't think they're there now, but it's just a matter of time.

Saviano: Yeah, I think you're right, Ben. And I'm not aware of any specific tools for that purpose, but especially when looking at how AI can be used applicable to legal contracts, I think that, you know, AI applied to the law has been more advanced and, you know, that disruption has been happening before disruption of tax through AI. I think that we've seen massive disruption in the legal field, and there are a number of firms that have developed natural language processing tools to extract information from contracts. What's really exciting about some of these tools now, and I know, Ben, that this has become a focus of Blue J, is the user interface. And so you don't need to have a PhD in computer science to do this stuff. Then you can actually, as a lawyer reading contracts, you can work with some of these tools. If you want to build a natural language processing tool to extract information from will, it's much more conceivable to do that today as a lawyer, as a non-PhD scientist, to be able to apply the technology that exists today. It's only going to become more advanced. That's going to be an important development in this field. So, you know, for the organizations that are watching today, you don't have to have an army of PhD scientists to add to your tax practice that the technology is developing such that, you know, tax practitioners can use this without those advanced degrees.

Griffith: So I want to move now, Jeff, you've talked about global governance in your opening remarks, and that really kind of speaks to my interest. So I want to start with a question on the use of technology by taxing authorities. And what do you all see as the future for taxing authorities using AI going forward? And then I kind of wanted to jump separately, I'll jump into some of the global governance things, which I was quite interested in.

Saviano: Sure, sure. I think that, you know, as we sit here in December of 2020, that if we look out over the next year as we emerge from the pandemic and the bright spotlight that has been shining on institutions around the world, there have been obviously failures of institutions, and we're seeing heightened interest from governments and tax authorities to apply these advanced technologies to solve their problems. We spent some time, just for an example, last week with the Asian Development Bank — had a get together of 15 to 20 tax authorities across Asia-Pacific region. And what was so interesting to me, Cara, is participating in that session you could feel the sense of urgency that you've got tremendous budget deficits around the world, economic crisis, and governments are looking for every edge possible. How can they not only automate, how can they do it better, faster, cheaper, but how do they better serve the citizenry?

And that's, of course, what government is focused on it. So in our work with governments, we're seeing tremendous interest in these multi-stakeholder platforms and using advanced tools like AI to perhaps skip a generation of technology. And that's also really exciting to us, that for some, perhaps they can leapfrog technology developments that they have not adopted but use the crisis that we have been in and are still in, use that crisis in order to promote the change in government to affect positively citizens, taxpayers, and how that citizen-government experience using AI systems and other technology to improve the experience, make better decisions. One last point I want to make, we talk a lot about automation and, you know, today and even using AI to read statutes. One of the areas that we think is very, very young that we're starting to see some solutions built is around using AI tools to develop better policies. So it's not just at the administration side in government, but we're seeing now for the first time AI being used to get real-time insights from how taxes have been impacting taxpayers and designing new policies and improving tax policymaking through AI system. That's really exciting to me too.

Alarie: Maybe I can jump in here. I think one of the things that tax administrations understand is that one of the most costly things that that can happen are mistakes. Mistakes are extremely costly for a tax administration, and mistakes, of course, come in two different kinds. One kind is where you're pursuing a taxpayer for taxes that you think are owing where you're actually mistaken. So this could cost many hundreds of thousands of dollars, millions of dollars in tax litigation, to pursue a particular taxpayer for tax that if, you know, you would have been better advised or done more effective research upfront, if you'd leveraged some of the machine-learning tools that might've helped you reach a better decision, you wouldn't have chased those revenues because you would have realized that those taxes just aren't due.

On the flip side, they're also mistakes of not pursuing taxpayers. And so you may be making false-negative decisions and errors on that front. And so you're not actually pursuing those sources of revenue that you should be pursuing. And this gets into the anomaly detection and other things that Jeff was talking about earlier, and even just better understanding the law. And so those mistakes are very costly. And sometimes they're really small mistakes that, but the costs kind of add up if you make a mistake and then it leads to several interactions with the taxpayer and several phone calls and faxes going back and forth. This is just costly in terms of human time and resources. If you can make better, more consistent, more reliable, more accurate decisions as a tax administration from the get-go, you're in a much stronger position. And I think business has understood this for a very long time — the whole movement towards highly reliable manufacturing, Six Sigma, all of these initiatives that have really taken manufacturing by storm. There's a recognition that mistakes are extremely costly. And so I think tax administrations are seeing that as well.

Lawsky: And one other complication is the politics itself that prevents tax administrations from developing that even when the technology exists. So go way, way back. The IRS, you know, developed the [discriminant function (DIF)] score to figure out who to audit. And that was pretty smart. And I don't say that just ‘cause my mom was a statistician who worked on the DIF. Hi, Mom, if you're watching. But what they did there is they got that through doing random audits, which is exactly what you should do, do random samples. They were skewed samples, so it was good. They got all this information, really in-depth audits. And they built this function that let them take the information from a return and figure out — combination of red flags, fancy regression analysis, and so forth — and figure out who to audit, and Congress said, stop harassing people and wouldn't let them do the in-depth random audits anymore.

Similarly, concerns about — and so, you know, that it's not as good a function as it could be — private people, tax preparers, lawyers have a lot of resources to try to develop these tools. If the IRS doesn't have the funding that it needs to develop the tools from its end to avoid the types of errors that Ben discussed, both the false positives and false negatives, the IRS doesn't have the resources to develop those tools. It's not going to be able to keep up with the technology happening among private lawyers. So it's in addition to developing good policies; the politics itself can prevent the kind of development that needs to happen. It's just an extra layer of complication.

Griffith: Yeah. Without a doubt. I, you know, when you look at budget constraints and things like that, it would be a big challenge. Jeff, you mentioned in your opening remarks, and this speaks to my interest on global governance, on the need for multi-stakeholder platforms, which then leads to needing new governance models. And that anytime you're talking about having something that is multinational, multi-state, a lot of the state and local tax world multi-state organizations are always challenged because you have a variety of interests that you're trying to get aligned. Do you think it's feasible when we're looking at meeting sort of global governance to allow for an AI application? Is this feasible? Is this possible? Is it going to happen? How difficult do you think it will be?

Saviano: Well, I think it will be very difficult. I think that we don't have the institutions today that have the mandate. I think that in order to really have the global scale and to actually have a multi-stakeholder platform that's multinational and is multi-dimensional, multiple taxes that are affected in different taxpayer sectors, I think it does; it gets very complex when you look at the great multinational institutions that exist today, whether it's the World Bank or the OECD, or the U.N. and others, that they're not really configured to take on something like this. I think we need to change at the institution level. I think we need an unprecedented level of public-private partnerships in order to promote that change.

But what is most interesting, Cara, to us is, you know, looking at where we sit in 2020 and what has been happening for the past nine months and the great feats of companies and individuals coming together to solve some real, real difficult issues. I think that we have an opportunity to adjust the mandates of some of the multilaterals to form new partnerships and actually have multi-stakeholder platforms, but not to sugarcoat it, I think it would be a profound step from where we are today. And I think that's really what's needed to effect change like this around the world. Just to give you one example, though, of a country that's done well: Estonia, the little country of Estonia, and how they have been a leader in adopting digital technologies. And as a result, their tax authority, they don't have as many workers in their tax authority. They compare to neighboring countries where, you know, the issues, Ben, that you talk about of cutting mistakes and having greater accuracy, the [key performance indicators] look quite good. The access to information, they vote on blockchain, they store medical records on it. This is AI-enabled systems. We have these microcosms in the world to see what good looks like. And it's up to us now, groups like this with governments and multilaterals, to come together and take those success stories and create a new layer of digital public goods in the world for everybody to benefit. And I do think we can do it. And I think that, ironically, that the pandemic can be an accelerant to achieving some of those great results.

Griffith: That's really interesting. I had written down as a question whether the pandemic had actually accelerated development, and it sounds like, you know, sometimes you find the silver lining and . . .

Saviano: I believe that. I wonder if others do, but I definitely believe that we have an opportunity for it to be an accelerant of digitization around the world. Absolutely.

Griffith: Yeah. It's fascinating. Ben, in your business, have you seen more adoption of Blue J in Europe and Canada, but in the U.S. and in Europe and developing countries, have you seen any place that stands out to you as really sort of being on the forefront, at least in terms of adopting your product?

Alarie: I think, no question, North America is the leader, but we are fielding requests all the time from different jurisdictions around the world that have an interest. We've had conversations with South Africa, with the U.K., with some countries in South America. We've had conversations with Australia. So there is a lot of interest in these kinds of AI solutions, machine learning solutions and tax, no question. I think what's also extremely common are the ethical considerations that arise. When you start thinking about deploying this technology, everyone's interested in the efficiency, benefits, the reliability, the accuracy, but they want to make sure it's safe for their tax systems. And so I think that's quite legitimate, and we see a civil society reacting in different ways so that we see open AI, really well-funded, out of California dedicated to making AI safe and available. And so I think organizations like open AI are also part of that global conversation that Jeff was talking about. And there, you know, if you turn your mind to reasons for pessimism, we can come up with a long list of reasons for pessimism. If you turn your mind to reasons for optimism, I think you can also come up with a bunch of reasons for optimism. And so I tend to be more optimistic by nature. And so I lean towards the optimistic side, but I think it's right; we shouldn't be sugar coating some of the challenges.

Griffith: Yeah. Without a doubt, there's a ton of challenges to be had. It is nice to see all of the, you know, the open-source development, and so that there will be something that is available to everyone should they want to use it. And so it's fascinating, but we have one sole minute left. I have to tell you, you guys have been just fantastic. This has been a fun conversation. I hope that maybe we can find some follow-ups and you guys will all join in again. It was a ton of fun, and I very much appreciate it. And before we close out, I wanted to say that for anyone that has viewed, we will be sending around a link. We can also provide contact information for the speakers, as well as information on Blue J and any papers and things that they might want to pass along. We can do that as well. So I want to thank you guys very, very much for participating today. I hope it was as much fun for you. I know I learned a lot. I have a paper full of questions, so expect the follow-up email to ask for another webinar in the near future.

Thanks to everyone who watched. I hope you all have a great day.

Alarie: Thanks.

Saviano: Thank you so much.

Lawsky: Thanks, Cara.




Copy RID