Summary

AI in tax law offers a way to make dense public information easier to explore. This episode examines how large language models could support research, compare provisions, and explain the U.S. tax code while keeping legal interpretation and policy choices accountable to people. It connects access to information with economic inequality and the distribution of resources. The opportunity is substantial, but reliable research requires more than a fluent answer.

Key Takeaway

• AI for tax research can assist with summarizing, comparing documents, and locating material for review.
• Statutes, regulations, court decisions, and agency guidance have different roles and must be read in context.
• BillSum is a legislative summarization dataset, while compliance analytics and legal research tools perform different tasks.
• A useful explanation identifies its sources, relevant dates, uncertainties, and questions that still require expert judgment.

How Can AI Explain Tax Law?

A reader can use an LLM to turn a passage into a plain-language explanation, identify defined terms, or compare two supplied versions of a provision. Those tasks can make an unfamiliar document more approachable. They do not establish whether the resulting interpretation applies to a particular taxpayer or whether an omitted exception changes the answer.

The Internal Revenue Service’s guide to legal authorities distinguishes the Internal Revenue Code in Title 26 from Treasury regulations and other official guidance. It also stresses checking the tax year and effective dates. Research needs that context: a correct quotation from a newer provision can still be the wrong authority for an earlier year.

LLMs and Tax Code Research

The episode explores the possibility of connecting legislation with its history and consequences. A VLDB workshop paper by Andrea Colombo and Francesco Cambria describes an LLM-assisted process for organizing U.S. legislation and links between provisions. Mapping citations can help researchers navigate related material. It does not, by itself, determine which policy is fair or prove the economic effects of an amendment.

Another example requires a clear distinction. The Association for Computational Linguistics paper introducing BillSum describes a dataset of U.S. Congressional and California bills for summarization research. It is not a comprehensive tax-advice platform. Comparing these projects by their actual purpose makes their contributions easier to assess.

Legal AI and the Limits of Automation

A Stanford Law School study of AI legal research tools found that retrieval-based systems could still generate incorrect information. The practical lesson is to open the cited authority, check whether it supports the claim, and separate the original text from the model’s interpretation. Confidence in the wording is not a substitute for evidence.

Enforcement analytics is a separate application. The IRS’s 2024 operating-plan update described using AI and advanced analytics to help select complex partnerships for examination. Selecting cases for human review is different from establishing liability, rewriting legislation, or showing that an LLM understands the entire system.

Transparency and Public Accountability

Better access to information could help people ask more informed questions about revenue, spending, and distribution. Our discussion of participatory budgeting and public decisions examines a complementary approach to participation. Capitalism and the distribution of value and rethinking progress through a well-being economy explore the values behind economic choices.

A sound research process preserves the documents being analyzed, identifies assumptions, and makes disagreements visible. AI can help organize a debate; elected institutions, legal professionals, and the public still have to evaluate its conclusions.

Frequently Asked Questions

Q1: How can AI explain tax law?

It can summarize supplied provisions, explain terminology, and suggest questions for further research. Check each explanation against the relevant legal sources and dates.

Q2: Can an LLM replace professional tax advice?

A general explanation cannot establish how the law applies to someone’s full circumstances. Individual decisions require current authority and qualified review.

Q3: Is BillSum a tax compliance product?

No. It is a research dataset for legislative summarization, with bills and reference summaries used to evaluate methods.

Q4: Can legal AI produce incorrect citations?

Yes. Research has found errors even in systems that retrieve legal documents. Verify citations and the propositions attributed to them.

Q5: Does organizing the tax code automatically make it fairer?

No. Organization can support analysis, but fairness depends on policy objectives, distributional evidence, public scrutiny, and accountable decisions.

Related Episodes

• Capitalism and Exploitation: The Engine Beneath the System
• Rethinking Progress: The Paradigm Shift in Societal Values and the Well-Being Economy
• Understanding Free Market Economics Through Milton Friedman’s Lens

Sources & Further Reading

• Association for Computational Linguistics — https://aclanthology.org/D19-5406/
• Bloomberg Law — https://pro.bloomberglaw.com/insights/technology/ai-for-legal-research/
• GitHub — https://github.com/VishalTheHuman/TaxEase.AI-Vertex-AI-Agent
• Internal Revenue Service — https://www.eitc.irs.gov/pub/irs-pdf/p3744b.pdf
• Internal Revenue Service — https://www.irs.gov/privacy-disclosure/tax-code-regulations-and-official-guidance
• K1x — https://k1x.io/leveraging-ai-for-tax-research-helping-small-firms-keep-up-with-changing-laws/
• Medium — https://medium.com/%40yauheniya.ai/building-an-ai-legal-agent-how-to-analyze-big-techs-tax-strategies-in-minutes-not-hours-1791dec1cfba
• Stanford Law School — https://law.stanford.edu/publications/hallucination-free-assessing-the-reliability-of-leading-ai-legal-research-tools/
• The Washington Post — https://www.washingtonpost.com/business/2025/07/26/doge-ai-tool-cut-regulations-trump/
• TIME — https://time.com/7277746/ai-deepfakes-take-it-down-act-2025/
• Times Union — https://www.timesunion.com/business/article/hearst-newspapers-uslege-partner-expanded-21197365.php
• VLDB Endowment — https://www.vldb.org/2025/Workshops/VLDB-Workshops-2025/LLM%2BGraph/LLMGraph-2.pdf
• Wikipedia — https://en.wikipedia.org/wiki/Large_language_model

1

Today we are diving deep into the US legislative and tax code.

2

The ultimate labyrinth.

3

It really is a system that is famously convoluted and definitely stressful.

4

And let's be honest, a primary source of political fog.

5

It is.

6

And for centuries, lawmakers have had to approach any kind of reform in really the

7

only way they could piecemeal one piece at a time.

8

Exactly.

9

Just think about it.

10

No single person or office could ever review the entire legal framework.

11

Holistically, it's just too big.

12

So every new law was added into this, this fog you're talking

13

a dense historical fog.

14

And that often created these hidden, you know, unintentional consequences,

15

somewhere else on the code.

16

A fog is the perfect word for it.

17

And our mission in this deep dive is to show how large language models,

18

LLMs are moving way beyond being simple chatbots.

19

Much further.

20

They're becoming, in a way, forensic auditors for the law itself.

21

The promise here is clarity, pulling that system out of the fog and giving you the

22

listener a full panoramic view of how the system actually functions, not just how

23

We think it does; that holistic view must be what changes everything.

24

It's the key.

25

I mean, the central challenge from a technical standpoint is that the law

26

exists as just massive amounts of unstructured text.

27

We're talking about statutes, court rulings, hundreds of thousands of

28

pages of statutory text, decades of court rulings that sometimes even

29

conflict with each other, and then millions of pages of IRS guidance.

30

It's a mountain of data.

31

And LLMs are the first tool that can not just read that mountain of data, but

32

actually correlate it, see the connections.

33

Yes.

34

Correlated to system-wide scale.

35

Historically, if you wanted to know the impact of one little amendment, you'd

36

hire a team of lawyers.

37

And it's been months, maybe years, tracing citations.

38

Right.

39

Slow, expensive, and always incomplete.

40

AI transforms that whole process from this, you know, bureaucratic obstacle

41

into a data-driven diagnostic tool.

42

Okay.

43

So it can take a proposed policy change and check it against everything else at

44

once simultaneously.

45

It checks it against the statutory text, the relevant court cases, and even the

46

economic impact data.

47

And that must be the aha moment for lawmakers.

48

Suddenly the AI is flagging things.

49

Exactly.

50

It surfaces contradictions, redundancies, loopholes, all the outdated provisions

51

that human review just misses because the scope is too wide.

52

You can finally see what works and what doesn't.

53

And that's the real focus here.

54

The goal isn't just automation.

55

It's about redesigning the system to be fair.

56

By looking at all this data and case law outcomes, you can ground reform in evidence.

57

So it's about getting away from political guesswork.

58

And towards a system that actually reflects transparency, coherence and, you

59

know, equity at scale, because it's finally based on verifiable data.

60

Okay.

61

Let's get under the hood.

62

That scale of analysis sounds almost impossible.

63

How do LLMs take this messy pile of documents and turn it into something

64

structured?

65

The answer is something called a knowledge graph or a KG.

66

The KG is the scaffold, then.

67

It is the essential scaffold.

68

But to build it, to build the US legislative graph, researchers first hit this,

69

This huge and kind of surprising challenge was document quality.

70

The source documents for US laws are all over the place, format-wise, depending

71

on when they were published.

72

Let me guess, the older the document, the worse the quality.

73

You got it.

74

The oldest chunk from 1951 to 1992 is about 11,425 laws that only exist as image-based PDFs.

75

Oh, wow.

76

So just scans of old paper, ancient scanned images, which immediately brings up a

77

huge problem with optical character recognition.

78

Right.

79

OCR.

80

That's the software that tries to turn an image of text into actual text.

81

You can edit.

82

And if the scans are bad, the OCR is going to be full of errors.

83

We're talking typos, words split in half, punctuation just gone.

84

So your foundation is shaky from the start.

85

Very shaky.

86

Yeah.

87

So to solve this, they came up with an LLM-assisted strategy.

88

They basically turned the AI into a highly specialized editor.

89

And they used a specific model for this.

90

They used a powerful open-source model.

91

Llama 370B was instructed to act as that correction agent.

92

It reads the messy OCR text and just cleans it up.

93

It fixes the typos and keeps the structure.

94

All while trying his best not to hallucinate or make things up, that one step

95

turns that low-quality junk into high-quality text.

96

That's your node in the graph.

97

Okay.

98

So now we have clean text, clean nodes.

99

Now you have to connect them all.

100

That's the next step: edge extraction, finding all the relationships between those

101

thousands of laws, which must be hard because they don't just use clean hyperlinks, do they?

102

Not at all.

103

The references are often buried in unstructured text, like the act of 1974 or some short

104

title.

105

It's not consistent.

106

So you need an AI that can understand the legal intent of a sentence, not just match keywords.

108

Precisely for that, they fine-tuned the different smaller models, Mr.

109

All seven B, just for this one job: relationship identification.

110

And it classifies the type of relationship, the edge, you could call it.

111

That's right.

112

It's not just law A mentions law B.

113

It's how it mentions it.

114

So what are the categories?

115

There are four main ones.

116

Does one act amend another?

117

Does it abrogate it?

118

Which means repeal it.

119

Okay.

120

Is it the legal basis of another law?

121

Or does it just cite it for reference?

122

And running this process over 60 years of history.

123

That's what builds the graph.

124

That's what built it.

125

Yeah.

126

A massive, queryable graph with over 31,000 unique functional relationships: the edges.

127

That's incredible.

128

But what happens if the AI gets it wrong?

129

If Mistral misclassifies an edge.

130

That's a huge risk.

131

I mean, if it says a law merely cites another, but it actually

132

obligates it; that's a catastrophic error for a lawyer.

133

It would be.

134

And that's why these systems have to be transparent.

135

A misclassified edge could lead to an indefensible legal position.

136

The beauty of the graph is that a human can then query that edge

137

and check the original statutory language.

138

So the AI is the discovery tool, not the final word.

139

Exactly.

140

It finds the connection in seconds.

141

Then the human verifies it.

142

Got it.

143

Okay.

144

Let's shift from the theory to the real world.

145

How does this impact big finance?

146

Let's talk big tech and global taxes.

147

Right.

148

This is where that structured data gets applied to these incredibly dense financial documents.

149

I saw an AI legal agent was used to analyze three massive 10K filings,

150

Alphabet, Microsoft, and Nvidia.

151

It was the goal was to check their tax positions against the OECD's

152

Pillar 2 initiative.

153

That's the global minimum tax, the 15% threshold for the effective tax rate or ETR.

154

And these 10 key filings are hundreds of pages long.

155

The AI agent using this graph-based knowledge was able to run these really complex analyses.

156

And the results were specific.

157

I mean, very specific.

158

Incredibly, for example, it found that Nvidia's effective tax rate was

159

calculated to be 12%.

160

12%.

161

So that's three points below the global minimum.

162

Correct.

163

And that potentially exposes Nvidia to what are called top-up taxes.

164

They have to pay the difference that 3% to the jurisdictions where they operate.

165

And the AI to point to the exact page in the document to prove it.

166

With precise page citations.

167

That's the verifiability we're talking about.

168

The analysis also found these really different tax splits, right?

169

Domestic versus foreign.

170

It did.

171

Alphabet and Nvidia were heavily US-centric.

172

Over 90% of their tax burden was domestic.

173

But Microsoft was much more global, more like a 60 40 split.

174

That kind of insight used to take teams of people weeks to figure out.

175

And now it's minutes.

176

It's a huge shift, which brings us back down to Earth.

177

This isn't just for huge corporations.

178

It's becoming well essential for smaller law and accounting firms too.

179

Absolutely essential.

180

I mean, think about the pace of change.

181

Small firms get buried under this constant stream of new regulations,

182

interpretations and court decisions.

183

It just can't keep up.

184

They don't have the research staff.

185

But AI acts as an equalizer.

186

The efficiency gains are just massive.

187

How massive are we talking?

188

We're seeing an 80% reduction in time spent finding relevant tax guidance.

189

80%.

190

That's not just an improvement.

191

That's a total game changer for a small practice.

192

It really is.

193

That's time they get back to actually advising clients to use their professional

194

judgment.

195

It's a huge business advantage and better for the clients who get faster,

196

more accurate advice.

197

And now we're even seeing that clarity filtered down to the public with tools

198

like taxis.ai.

199

Right.

200

These are the user-friendly assistance programs that help individual taxpayers.

201

Yeah.

202

They use a technique called retrieval-augmented generation, or RAG.

203

A.A.G.

204

They draw on this knowledge base of up-to-date forms and instructions to give

205

you personalized guidance.

206

So it's demystifying a process that, frankly, a lot of people feel is

207

intentionally complicated.

208

It's a huge step toward equity, but this is really important, but we have to add a major caveat here.

210

Okay.

211

This is a powerful tool.

212

And like any powerful tool, it needs human judgment and validation.

213

It's an assistant, not a replacement.

214

So we have the good use case where agencies like the IRS use it for fraud

215

detection or bill analysis.

216

Right.

217

That's AI enhancing the existing process.

218

But then there's the other side, the push for pure efficiency, which brings us

219

to the, let's say, controversial use of the DOGE AI deregulation decision tool.

220

The name alone is something and its goal is huge, isn't it?

221

It's staggering.

222

The tool is supposed to analyze 200,000 federal regulations to identify and

223

eliminate a hundred thousand of them, a hundred thousand and the projected

224

savings are 3.3 trillion dollars annually, a very powerful incentive to move fast.

225

But what went wrong when they tested it?

226

You mentioned a HUD employee's experience.

227

Well, the employee who tested it reported that the AI's analysis just came to

228

the wrong legal conclusions.

229

How so it would sometimes misread the language.

230

It would claim that a regulation was outside the scope of the underlying

231

statute when, in fact, the AI had just misinterpreted the law I was based on.

232

So it's like the AI decided a key paragraph was irrelevant when that

233

paragraph was the entire legal foundation for the rule.

234

Precisely, which brings us back to this idea of a legal grade standard.

235

The best tools, the ones you can trust, are transparent and traceable.

236

You have to be able to follow the data trail back to the source text.

237

It has to be an enhancement to a professional's judgment, not a replacement for it.

238

What an incredible arc we followed from messy scanned PDFs to a precise,

239

queryable map of the entire US legal system.

240

LMS aren't just simplifying taxes.

241

They are making the underlying structure of the law visible for the very first time.

242

And that structural view gives us these profound insights.

243

Researchers analyzed the graph to see which subjects have the most influence.

244

They're centrality.

245

And what did they find about modern lawmaking?

246

They found that the centrality ranking of major subjects, things like economics, public finance, it's remained almost entirely unchanged since 2015.

248

Unchanged for nearly a decade.

249

Almost static.

250

Might they?

251

Which suggests a stabilization or maybe more provocatively, an

252

ossification of thematic focus in US lawmaking.

253

So the core plumbing of the system is stuck on the same things, even as the

254

world outside is changing rapidly.

255

That's what the data implies.

256

The core pillars of influence haven't really shifted.

257

So here is the provocative thought we want to leave you with.

258

If AI can now diagnose exactly which parts of our legal system are the most

259

complex, the most redundant or outdated.

260

Mm hmm.

261

If it can give us that perfect clarity, then the important question is no longer

262

can we achieve simplicity and fairness in our laws.

263

It's whether the political will exist to actually use this newfound evidence

264

to modernize the rules.

265

A profound question to watch unfold.