Article on Using Machine Learning for Predicting Tax Ownership (9/27/26)

I spent more time than I should have reading Young Ran (Christine) Kim and Dmitry Erokhin, Algorithmic Tax Ownership, 51 BYU L. Rev. 1437 (2026), here. Don’t get me wrong, I think the article’s discussion of its premises is quite good. What threw me off was the discussion of the “Algorithmic” part of the article. The authors deploy math and computer modeling to predict tax ownership by digesting a large dataset of cases that have dealt with tax ownership.

Of course, the authors mention prominently the iconic case from my time at DOJ Tax: Frank Lyon Co. v. United States, 435 U.S. 561, 582–83 (1978), GS here. I had an indirect role in Frank Lyon (which I’ll tell in the digression below).

Frank Lyon is the classic Supreme Court screw-up in tax law. For further discussions of the screw-up, see the citations in the article. (p. 1446 n. 22 and 1447 n. 32). One of my favorite articles is Charles I. Kingson, The Confusion over Tax Ownership, 93 TAX NOTES 409, 410 (2001), SSRN here (criticizing Frank Lyon as “nobody knows what it means”). Shortly after leaving DOJ Tax for private practice in 1997, I advised a bank wanting to do for a lead shareholder essentially the same deal as Worthen Bank did for Frank Lyon through his corporation. My advice was essentially that it’d be a crap shoot because nobody knew what Frank Lyon meant, but that we could clone the documents in Frank Lyon to recreate the exactly same structure and to be even more sure the bank should change the bank's name to Worthen Bank and his corporation's name to Frank Lyon Corp. Of course, that was hyperbole, but not much because the multi-factor test of tax ownership that Frank Lyon deployed is less than clear in their application in other cases.

Back to the article, it concludes as follows (p. 1494):

To refine tax ownership analysis, this Article applies machine learning, specifically focusing on repo transactions. It identifies economic benefits and burdens  Factor 1) as the most decisive factor in classifying transactions as either sales or loans, with 94.6% accuracy. Additional factors, while sometimes helpful, do not significantly impact classification outcomes.

Earlier (p. 1484):

Nevertheless, the proposed model undeniably reshapes the current tax ownership analysis by showing the decisive factor of the analysis is who bears the economic benefits and burdens of the underlying asset. The clarity this model brings to tax ownership analysis is applicable in a wide array of contexts including repo transactions, sales and leasebacks, litigation financing, and more.

True, as modeled at the edges by considering and weighting other factors, the basic test of benefits and burdens nails the ownership question. I am not sure in the real world that the mélange of other factors (such as those touted in the majority opinion in Frank Lyon) and the additional computer modeling touted in the article mean anything material in the real world. So, I suppose, there’s that.

But if “that” is really something for the future, I probably will not have to encounter it because I likely will not be worrying about this stuff much longer. Some of you younger guys may get into it.

JAT Digression:

Now, for the reason for my particular interest in Frank Lyon.

This is a longer digression which I would normally present in numbered paragraphs. But, as I present below, I think unnumbered paragraphs are the proper presentation.

While with DOJ Tax Appellate, I had a case like Frank Lyon that preceded Frank Lyon. The case was American Realty Trust v. United States, 498 F.2d 1194 (5th Cir. 1974), GS here. There is a lot of drama behind that case that I won’t go into now except that the Government trial attorney told the trial judge that the tax ownership issue was a fact issue for which she had the right to a jury trial. (Usually, in a tax refund suit, it is the taxpayer demanding the jury, but in this case, it was the Government by the Government attorney.) The judge said something to the effect: “OK, little lady, you’ll get your jury.” The jury returned a special verdict that “the ART-Helmsley transaction” was a good faith purchase and lease-back by ART, and not merely a financial arrangement.” (498 F.2d, at 1196-1197; cleaned up). (And, just in case you are wondering, the Helmsley was that Helmsley (Wikipedia here), husband of Leona Helmsley (Wikipedia here), later tried for tax crimes where it came out that she was the “Queen of Mean” and said only little people pay taxes.) (BTW, a classmate from UVA Law ’67, defending Leona argued in opening “I don't believe Mrs. Helmsley is charged in the indictment with being a tough bitch.” )

In American Realty Trust, the Government trial attorney’s insistence on a jury was a problem on appeal that we tried to finesse by arguing that the issue was legal rather than factual. But the panel was not impressed, saying: “In oral argument before this Court, government counsel suggested that the proper characterization of the ‘substance’ of the transaction was essentially a factual matter.” (Note that suggestion, by my reviewer, Jonathan Cohen, on the case because I had transferred to the trial section, was inconsistent with the brief; if I had argued the case, I would never have made that concession.)

After emphasizing that the submission to the jury and form of its instructions and special verdict was the Government’s, the panel said it could just hold for the taxpayer on the standard review of factual issues determined by juries. (See 498 F.2d, at 1198-1199.) But the panel then purported to address our real argument that Helmsley retained the burdens and benefits but twists the argument to say that we were arguing sham, not bona fide, and substance. There is a difference between burdens and benefits and notions such as sham, not bona fide, and substance. The first is just an economic inquiry without any suggestions of impropriety that may be in words like sham, not bona fide, and even substance.

The Government lost American Realty Trust.

So, when I got to the Trial Section, Frank Lyon was in the pipeline. The Government lost Frank Lyon at the trial level. I did not handle the case, but because of my interest in the issue from American Realty Trust, I urged to the DOJ Appellate person writing up the Section’s position as to whether to appeal Frank Lyon that the appeal should be taken. I don’t know whether my “urging” influenced the final decision, but DOJ Tax Appellate recommended appeal and the Solicitor General approved the appeal. (Dean Griswold, discussed below, was the SG until March 1973; I don’t think he was the SG when the SG approved appeal in Frank Lyon; if he had been SG, I am sure he would have not taken Frank Lyon as a client to petition for certiorari.)

The 8^th Circuit decided for the Government. Frank Lyon Co. v. United States, 536 F.2d 746 (8th Cir. 1976), here. Judge Bright wrote a masterful opinion based on the benefits and burdens.

Frank Lyons then engaged Dean Griswold to seek certiorari. Dean Griswold (as we called informally him at DOJ because of his former deanship at Harvard Law) as SG was cautious about the Government seeking certiorari and had an excellent reputation with the Court for candor. Frank Lyon was an early, perhaps earliest, request for certiorari after leaving the SG position. I infer (but cannot prove) that his reputation and the Court wanting to give him one was the reason the Court granted cert. After all, his name on the petition for cert was his statement to the Court that the issue was cert-worthy and his name on a petition meant something to the Court. And, then when the Court took the case, I think they realized that they should not have taken the case (meaning they should have DIGGED it), but chose to deliver wrote an opinion that was pretty much nonsense in terms of what Court opinions are supposed to do—give guidance in the application of the law. Frank Lyon is poor guidance for future cases. Frank Lyon was a mess. (See the articles I cite above.)

Finally, I asked ChatGPT to hold forth on the difference in the 8^th Circuit’s opinion in Frank Lyon and the Supreme Court’s opinion. I went through several prompts and ChatGPT does not like the Supreme Court’s majority opinion, preferring Justice Stevens’ dissent and Judge Bright’s opinion in the 8^th Circuit. ChatGPT's responses are pretty good. The prompts and responses may be read here.