Artificial intelligence tools such as ChatGPT and Grok are becoming household fixtures with Americans increasingly turning to them. These tools are being used to find answers on everything from cooking recipes to complicated tax questions. How reliable is AI for taxpayers seeking tax advice, particularly on U.S. international tax issues?
While AI has laudable capabilities and provides great speed in its responses, the reality is that currently in the U.S. international tax area (indeed, in the tax area generally), AI often delivers inaccurate or incomplete information. Since U.S. international taxation is such a highly specialized field, AI can leave users at risk of costly mistakes. For now, a thorough and proper tax analysis requires the expertise of a seasoned professional who can understand the interplay of the tax law, Treasury regulations, IRS notices and rulings, court cases, and international tax treaties.
AI Tax Advice: The Temptation and Limitations
There is an undeniable appeal of AI in the complicated U.S. tax space that often overwhelms even tax professionals, let alone the lay person. With its speedy answers that are generated so easily and for free, the AI tool seems a godsend in the complicated tax world. For the more straightforward and typically domestic (as opposed to international) tax issues, these tools might be sufficient to point users in the right direction.
The cracks in AI’s armor start to show when tackling U.S. international tax matters. This is not only because international tax is an area rife with complexity but also because it is highly dynamic. Aside from the technical aspects involved, U.S. international tax is shaped by frequent regulatory updates, IRS interpretations, and judicial decisions. AI struggles to keep pace.
Take for example, the proliferation of IRS Practice Units, many of which focus on international tax matters. Practice Units are internal IRS training materials for examiners. The IRS just started issuing these in 2020. As international tax issues are becoming more and more common in today’s global economy the IRS has had to be proactive training its agents to understand them and catch noncompliance.
The Units are publicly available on IRS.gov but are not binding and cannot be used as precedent. From January to the first week of May 2025 the IRS has already issued 35 such Units. Of these, 22 cover complicated international tax topics such as “French Foreign Tax Credits”, “Base Erosion Anti-Abuse Tax”, “Foreign Tax Credit Computations”, “Foreign Tax Credit Limitations”, “General Deductions of a Foreign Corporation Engaged in U.S. Trade or Business (Non-Treaty)”, “Reduced Foreign Taxes Under Treaty Provisions” and many more subjects delving into technical international taxation. The Practice Unit resource is just one of many forms of tax guidance that is constantly evolving. The tax laws, regulations, treaties and updates may not be picked up by AI models resulting in incorrect or incomplete answers.
The Risk of AI Inaccuracy
The problem of AI’s questionable accuracy has its roots in how AI generates its responses. AI tools such as ChatGPT and Grok rely on so-called “large language models” trained on a huge and diverse amount of text that includes online forums, books, articles, websites and public records. This training results in an end-product that sounds like highly intelligent answers cloaked in perfect conversational text. Critically, this alone is not sufficient for a tax analysis, and it is all too easy to forget that AI cannot think independently. Professional judgement remains critical.
AI tools do not guarantee accuracy. This is particularly so in nuanced areas such as international tax which demands precise, up-to-date expertise. There’s the significant problem of “simplexity”. The tax law is highly complex. If AI presents the law as if it is simple, this can result in presenting the law as something different from what the law actually is, as has been encountered with the IRS chatbot “Interactive Tax Assistant”.
Other problems involve misinterpretation, relying on outdated sources, entirely missing out on important information, conflating similar but distinct international tax concepts (e.g., it might confuse the Foreign Tax Credit with the Foreign Earned Income Exclusion) and even “hallucinating”.
My own experiences with an AI “hallucination” occurred recently. The first, when I asked whether there was an exception in the FBAR filing rules for non-U.S. accounts owned by a foreign government entity. I was examining whether Pope Leo XIV, a U.S. citizen, might have a possible FBAR obligation with regard to accounts at the Vatican Bank. AI incorrectly reported that such an exception existed (the exception applies only to foreign financial accounts of any governmental entity of the United States). Another incident occurred when I asked if the U.S. had negotiated totalization agreements with the countries of Ecuador or Costa Rica and received an (incorrect) answer in the affirmative. No such totalization agreements exist.
Other tax professionals testing AI have found incorrect answers, in one instance 100% of the time. Even though answers were on the right track, AI failed to catch the critical nuances, of which the U.S. tax laws are replete.
Any of these missteps could lead to incorrect answers or calculations, potentially triggering an IRS audit and penalties. The problem, especially for the layperson, is that the AI generated responses sound so good, that they are taken at face value.
AI Errors Have Real-World Consequences
Getting things wrong, especially in international tax compliance which remains a top IRS enforcement priority, can lead to serious consequences. Errors in reporting foreign income or assets involves stiff penalties. For example, there is a $10,000 penalty for an untimely FBAR or foreign information return such as Form 8938 or 5471; a 40% accuracy-related penalty on unpaid tax for errors involving foreign assets. Reliance on AI and AI chatbots will not suffice as “reasonable cause” to avoid imposition of IRS penalties.
Last year, the U.S. Taxpayer Advocate Service cited a Washington Post review finding that AI chatbots from leading tax return preparation companies gave inaccurate tax advice up to 50% of the time on complex tax questions. Beyond the significant financial costs that can result from inaccurate AI tax advice, taxpayers face the stress of audits and the burden of correcting mistakes.
Why Human Tax Professionals Matter A Lot
Currently, AI has serious shortcomings when it comes to U.S. international taxation. Experienced international tax professionals occupy an irreplaceable role and unlike algorithms, they specialize in deciphering the many nuances involved in international tax matters.
The top professionals stay current on IRS guidance, monitor treaty updates, and analyze case law on a regular basis. They tailor tax advice to a client’s unique situation. Doing this often requires using professional judgment because a one-size-fits-all approach does not work in the international tax world. Tax professionals strategize and provide all-important context that can optimize the tax outcome for each particular case. AI has yet to master any of these skills; it has a long way to go.
This isn’t to say AI has no place in the world of international tax planning. It can certainly be a useful time-saving tool and starting point for a general explanation of unfamiliar tax concepts. Relying on AI alone, however, is a gamble not worth taking. In time, as AI inevitably evolves, its accuracy and abilities will develop, but at the moment AI is not a substitute for a qualified tax professional.
Stay on top of tax matters around the globe.
Reach me at vljeker@us-taxes.org
Visit my U.S. tax blog www.us-tax.org
NO ATTORNEY-CLIENT RELATIONSHIP OR LEGAL ADVICE
This communication is for general informational purposes only. It is not intended to constitute tax advice or a recommended course of action. Professional tax advice should be sought as the information here is not intended to be, and should not be, relied upon by the reader in making a decision.
Read the full article here
