AI Drafted the Tax Argument. The Practitioner Still Owns It.
In tax law, being fast only matters after you make sure everything is accurate.
Generative AI changes how we work, but it does not change legal responsibility. A tool that quickly creates a polished memo can also make up citations, invent safe harbors, or suggest transaction structures that look right but miss the real requirements. In tax, this is not just a drafting issue—it is a compliance issue.
People often say AI is risky because it sometimes makes things up. That is true, but it is not the whole story. AI works by guessing what is likely, while tax law relies on clear authority, rules, and checking facts. The system does not care if an answer sounds correct—it cares if you can trace it to a real law, regulation, ruling, case, or record.
The Event
Generative AI is now used to read statutes, draft memos, summarize legal sources, and help with tax analysis. It really does save time and makes complicated material easier to handle.
But this same ability also changes the risks. Large language models do not perform legal reasoning as tax professionals do. They predict what text should come next. When asked about rules like IRC §351 or §199A, the model might give an answer that looks finished but actually makes up a safe harbor, gets a Treasury Regulation wrong, or invents a precedent. The real risk is not messy output—it is a polished answer that is actually wrong.
Courts have already seen what can go wrong. There are examples of AI-generated legal filings that cited fake cases, leading to sanctions, thrown-out memos, and disciplinary actions. In tax, these mistakes are even more serious because the system relies on people following the rules themselves. Fake authority does not just mislead a Court—it also adds misinformation to a system that is already very complex.
This issue became even more important after Loper Bright. Now that Chevron has been overruled, courts must interpret ambiguous laws on their own, without relying on agency explanations. This means the actual text, structure, legislative history, and real records matter even more. AI can quickly find and summarize lots of material, but finding information is not the same as interpreting it. If AI invents a Senate report or Treasury explanation, it is not just harmless background—it can actually distort the record.
The Real Driver
The real reason behind these changes is not technology itself—it is the incentives people face.
Tax professionals are under pressure to work faster, cut costs, and get more done in less time. AI fits these needs well. It makes it seem like you have more help. You can ask for a memo, a transaction outline, or a list of citations and get something that looks professional. Because the output looks smart, people are more likely to trust it without checking.
That is the trap. AI is not a junior associate or a paralegal. It is not someone you can supervise like a regular team member. You give it prompts, but it is not trained in the same way. It lacks loyalty, cannot exercise judgment, and does not understand what happens if it is wrong. Treating AI like a legal assistant is misleading and reduces accountability.
The better model is instrumentality. It is better to think of AI as a tool that helps people do more. But unlike a calculator or a word processor, AI can create legal claims. If a calculator makes a mistake, you can usually see it. If AI makes a mistake, it might look convincing even though it is wrong. Because of this, the human practitioner must check every legal point before it goes out. As drawn from United States v. Boyle, reliance on another actor does not excuse failure to perform a non-delegable duty. The source material applies that principle to AI verification. Checking whether a case exists, whether a quotation is accurate, and whether a cited authority supports the proposition is not the kind of judgment that can be outsourced to a machine. It is ordinary professional diligence.
Circular 230 points in the same direction. Section 10.22 requires due diligence in determining the correctness of representations made to Treasury. A large language model is not a person who can be supervised within the meaning of that framework. The current rules were built to account for human error and misconduct. AI changes scale. One practitioner can make an isolated mistake. One AI workflow can generate thousands of plausible errors.
The Pattern
The recurring pattern is institutional lag.
A new tool is used in professional practice before rules and controls are in place. People start using it because it saves time. Firms allow it because it cuts costs. Regulators and courts react only after problems surface in filings, client advice, or public records. By that point, the issue has grown from a single mistake to a bigger process risk.
Tax is especially at risk because technical details matter a lot. AI is good at getting the form right. For example, it can describe a §351 exchange with property transferred, stock received, and control met. But tax law also considers whether the deal has real economic substance, a business purpose, and an actual economic impact. These questions depend on intent, context, and real facts. AI can replicate the language of these rules, but it cannot determine whether the facts actually fit.
That produces synthetic compliance. A transaction can appear properly assembled yet fail the doctrine that matters. Under IRC §7701(o), economic substance requires meaningful economic change and substantial non-tax purpose. A model can draft business justifications from common patterns, but those justifications are not facts. They are generated language. The same problem appears in step transaction analysis. AI can sequence steps neatly. A Court may collapse them if they were part of a fixed plan. This pattern also shows up in jurisdictional analysis. AI often misses important differences because it learns from the most common patterns in its training data. Federal tax rules might be overrepresented, while state, local, and international rules are less visible. This can cause mistakes, such as using the wrong Court circuit, treating OECD guidance as U.S. law, or mixing up VAT with U.S. sales tax. Tax law needs specific answers, but AI often turns specifics into general responses.
Implication
The responsibility stays with the practitioner, not the tool.
If AI creates a fake authority, the Court does not punish the software. If client data is put into a public model, the privilege problem is not the model’s fault. If an AI-generated structure fails the economic substance test, the machine is not penalized under §6662 or subject to promoter issues under §6700. The person is responsible for the submission, the advice, and any breach of confidentiality.
The privilege issue is especially clear. Most generative AI tools run in the cloud. Putting client facts, return positions, or tax strategies into these tools can send sensitive information to a third-party provider. If that provider logs, stores, or reuses the data, confidentiality can be lost. The Kovel analogy does not fix this. Kovel protects some non-lawyer helpers who work under a confidential relationship. Public AI tools do not fit this model because they lack agency, loyalty, or a clear confidentiality framework.
IRC §7216 brings up a similar concern when tax return information is shared without permission. The source material says this is more than just an ethics issue—it is a structural problem caused by data being stored, hidden, or reused. The main lesson is simple: a public AI prompt box is not a private conference room, no matter how friendly it seems.
Lessons for Practitioners
When using AI, the real work is in verifying the results. The draft might come from the tool, but responsibility for accuracy lies with the practitioner.
Just because something sounds fluent does not mean it is reliable. A polished list of citations is risky because it makes it look like someone has already checked it.
Anti-abuse rules show the limits of drafting based only on form. AI can put steps together, but economic substance and step transaction analysis rely on facts, purpose, and context.
Confidentiality risks begin as soon as you enter information. Putting client facts into a public or poorly controlled AI system can cause privilege and disclosure problems before you even draft a memo.
Governance is more than just a policy memo in a folder. Good controls include keeping records of source checks, usage logs, and version histories; using restricted tools; having zero-retention contracts; and ensuring a human reviews everything before it goes to a client, Court, or agency.
Human Element
When people are under pressure, they are more likely to take shortcuts if those shortcuts look professional. AI makes this easier by hiding the rough parts of early drafts. The draft comes out looking finished and confident, which makes people trust it before they have checked it.
Forward View
The continuing pattern is not prohibition. AI will remain in tax practice because the efficiency gains are too large to ignore. The profession will not stop using tools that can summarize material, organize research, and accelerate drafting. The question is whether those tools are placed within a system that makes errors harder to detect. That system must make human review the main checkpoint. Courts might require proof that AI-assisted filings have been checked. Treasury could update Circular 230 to cover generative AI. Firms might limit public tools and require verified databases for legal work. All these steps lead to the same point: AI can help with tax work, but it cannot take on legal responsibility.
The tax system can handle faster tools, but it cannot handle unverified information that appears to be professional judgment.