2026-09-26
In a prior blog post, I discussed how I downloaded the Tax Court's opinions and converted the PDF files to text files. I have kept that collection up to date, and it now includes 14,848 opinions.
For this post, I used the text files to look at collection due process (CDP) opinions. I wanted to answer two questions:
CDP cases arise under sections 6320 (liens) and 6330 (levies), which were added by the IRS Restructuring and Reform Act of 1998. In a CDP case, the Tax Court generally reviews the determination made by the IRS Independent Office of Appeals after a CDP hearing.
CDP cases are easy to identify because their docket numbers end with the letter "L" (for example, 16440-99L). Small tax case CDP dockets end with "SL." There were a total of 2,174 CDP opinions. The first CDP opinion was filed in March 2000. CDP opinions make up 17.5% of the 12,441 Tax Court opinions filed since 2000 (14.6% of all 14,848 opinions on the Tax Court website).
Rather than reading all 2,174 opinions, I used an open-source large language model, Google's Gemma 4 (the 26-billion-parameter version, gemma4:26b), running locally on my computer with Ollama.
I used a technique called single-shot classification. The model is given the full text of an opinion, followed by a multiple-choice question, and is told to answer with a single letter. Instead of letting the model write an answer, I read the probability the model assigned to each letter. The letter with the highest probability is the classification, and the probability serves as a confidence score.
Each opinion went through two steps:
Step 1: Does the opinion have one issue?
A procedural disposition on a single question (e.g., a dismissal for lack of jurisdiction, mootness, or a ruling on a motion for summary judgment) counted as one issue.
Step 2: For single-issue opinions only, who won?
I spot-checked 5 opinions from each category (1.B multiple issues, 2.A single-issue IRS won, 2.B single-issue taxpayer won, and 2.C single-issue mixed result) to confirm the classifications were reasonably accurate. I did not see any problems with the breakout between single issue vs. multiple issues. But two of the other classifications were incorrect. For example, in Peterson v. Commr., T.C. Memo. 2016-17, the taxpayer had asserted that he was not sent a notice of deficiency. The IRS appeals officer had mistakenly verified the mailing of the notice of intent to levy instead of verifying the mailing of the notice of deficiency. The court was unable to conclude whether the notice of deficiency had been mailed. Thus, the case was remanded to appeals. The AI model treated this as category 2.B, the taxpayer winning, when it would be better classified in category 2.C, remand to Appeals.
Another opinion (docket number 17789-16SL) was classified in category 2.C (mixed result), when it should have been in category 2.A (IRS won). I only spot-checked 20 out of 2,174 opinions, with 2 mistakes out of 20.
| Opinions | Percent | |
|---|---|---|
| CDP opinions | 2,174 | 17.5% of all opinions since 2000 |
| Single-issue CDP opinions | 652 | 30.0% of CDP opinions |
| IRS won | 573 | 87.9% of single-issue CDP opinions |
| Taxpayer won | 69 | 10.6% of single-issue CDP opinions |
| Mixed result, remand, or unclear | 10 | 1.5% of single-issue CDP opinions |
CDP opinions seem to be among the most narrowly focused types of Tax Court opinions, so I had expected most of them to have a single issue. I was surprised that 70% of the CDP opinions had more than one issue. I was not surprised that the IRS won the large majority (88%) of the single-issue CDP opinions. Much of the time one of the issues is dealing with whether the IRS abused its discretion (1,681 of the CDP opinions include the phrase "abuse of discretion" or something similar). Abuse of discretion is a pretty high bar for taxpayers to meet.
The classifications were made by an AI model; not by me. For 93.4% of the opinions, the model assigned a probability of at least 80% to each of its answers. For the other 144 opinions (6.6%), the model was less certain. I have only spot-checked 20 of the opinions, and two of those opinions that had a high probability for its answers were wrong.
[It is amazing how easy the classifications were with Claude. The hardest part was manually spot-checking to verify that the classifications generally looked good.]