Faris Adani
Tax season, that annual ritual fraught with complexity and often dread, once again descends upon us. In an era increasingly defined by the rapid ascent of artificial intelligence, a natural inclination might be to turn to these sophisticated newfangled tools, hoping they could demystify the laborious paperwork, streamline the process, and perhaps even unearth elusive savings. However, a recent, comprehensive investigation by the New York Times delivers a stark warning: employing leading AI chatbots for tax preparation is, for now, a perilous gamble likely to backfire spectacularly, potentially costing taxpayers thousands of dollars and untold headaches.
The New York Times put four prominent AI models through their paces: OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, and xAI’s Grok. Their mission was to solve a series of real-world tax scenarios, meticulously designed from training materials provided by the reputable tax service TaxSlayer. The results were unequivocally dismal. All tested chatbots struggled immensely, failing to correctly identify and complete the appropriate forms, and consistently fumbling key calculations. On average, these AI assistants miscalculated the tax money owed to the IRS by a staggering sum exceeding $2,000. This isn’t merely a rounding error; it’s a significant financial discrepancy that could lead to penalties, audits, and a costly, time-consuming back-and-forth with tax authorities.
Benedict Evans, a respected technology analyst, succinctly captured the core issue for the NYT: "The problem with taxes is all those very small little details matter, and it’s not going to get every single little detail right." He acknowledged the rapid advancements in AI, noting, "These models get dramatically better over the course of every six months. But they still give you what is roughly the right answer, and that’s not what you want." For taxes, "roughly right" is catastrophically wrong. The precision demanded by the tax code leaves no room for approximation, inference, or the occasional "creative interpretation" that LLMs are prone to.
The conundrum lies in the fundamental nature of current large language models (LLMs). While incredibly adept at processing, summarizing, and generating vast amounts of information, their strength lies in pattern recognition and prediction, not in absolute factual accuracy or logical deduction. This inherent characteristic explains why chatbots frequently "hallucinate," fabricating false factual claims even when tasked with summarizing a single, clear document. It’s why AI programming assistants can subtly slip errors into their code, and image generators produce strange visual artifacts and inconsistencies. Arithmetic, particularly when entwined with the byzantine and ever-shifting landscape of tax laws and their highly specific, interconnected forms, exposes this weakness in its most financially impactful form. It’s a recipe for not just inconvenience, but for significant financial penalties and protracted disputes with the Internal Revenue Service.
Erik Brynjolfsson, a senior fellow at the Stanford Institute for Human-Centered AI, drew a crucial distinction for the NYT. He explained that established tax software like TurboTax or TaxAct operates on "procedural, following ‘if-then’ logic built for mathematical precision." These systems are meticulously coded with the exact rules and calculations of the tax code, designed to guide users through a structured process that ensures accuracy. In contrast, large language models are fundamentally "prediction engines." They generate responses by predicting the most probable sequence of words based on their vast training data. While this enables them to be "superhuman at many tasks," it also means they can "fail at some that seem simpler to humans" – especially those demanding absolute, unyielding precision and an understanding of complex, rule-based systems like tax law.
This distinction highlights why the NYT‘s testing methodology, and its outcome, are so telling. The AIs only began to fare better when provided with "highly specific instructions," detailing precisely where each piece of information should be placed within each IRS document. This, as the article rightly points out, fundamentally "defeats the point of using an automated tool in the first place." The average taxpayer turns to automated software precisely because they don’t possess the granular knowledge of tax forms, line items, and obscure regulations. They seek a system that guides them, interprets their data, and accurately populates the necessary documents without requiring them to already be an expert. Asking an AI to do taxes currently demands that the user is the expert, essentially transforming the AI into a glorified, error-prone data entry assistant rather than a true problem-solver.
The risks associated with relying on AI for taxes extend far beyond mere inconvenience. Incorrect filings can trigger IRS audits, lead to hefty fines, accrue interest on underpaid taxes, and even result in legal repercussions in severe cases of misrepresentation. The emotional toll and time drain of navigating an audit or disputing penalties with the IRS are significant. Furthermore, the question of accountability remains largely unanswered: if an AI makes a costly mistake, who bears the legal and financial responsibility? Is it the user, who ostensibly reviewed the output, or the developer of the AI tool? This is a nascent but critical area of legal and ethical debate that complicates the adoption of AI for high-stakes tasks.
The perils aren’t merely theoretical. TurboTax, a giant in the tax preparation industry, conducted its own experiments with AI, deploying its "Intuit Assist" chatbot to answer tax questions. The company’s findings mirrored the NYT‘s: the chatbot frequently spun off irrelevant answers, and when its responses were on topic, they were often demonstrably wrong. This internal validation from a company deeply invested in accurate tax solutions further underscores the current limitations of general-purpose LLMs in this domain.
Despite these current shortcomings, the future isn’t entirely bleak for AI in the tax landscape. As Benedict Evans suggests, AI models are improving rapidly. Their potential utility, however, will likely lie in very specific, carefully controlled applications rather than standalone, unsupervised tax preparation. Future AI tools might excel at summarizing recent changes in tax legislation for tax professionals, identifying potential deductions or credits based on user-provided financial data (as a prompt for human review, not a definitive answer), or automating certain aspects of data entry when seamlessly integrated with established financial software. Personalized guidance, too, could emerge, but always with the critical "human in the loop" oversight. The key will be the development of highly specialized AI models, meticulously fine-tuned on verified legal and tax texts, rather than relying on general-purpose LLMs which are designed for broad utility, not pinpoint accuracy in a niche, complex domain. Integration of AI as a feature within robust, rule-based tax software seems a far more probable and responsible path forward than replacing the entire process with an AI chatbot.
In conclusion, while the allure of an AI-powered solution to the annual tax headache is undeniable, the current state of artificial intelligence, particularly general-purpose LLMs, renders it an unreliable and financially risky proposition. The precision demanded by the intricate, ever-evolving tax code is simply not a strength of these predictive engines, which are prone to factual errors and approximations. Until AI technology matures to a point where it can consistently demonstrate absolute accuracy, explain its reasoning, and operate with the same logical rigor as traditional, rule-based software, taxpayers are strongly advised to stick with established, reliable methods for preparing their returns. The cost of a spectacular backfire, in terms of financial penalties, audits, and undue stress, far outweighs any perceived convenience offered by today’s nascent AI tax assistants. For now, the safest bet remains human expertise and thoroughly vetted, purpose-built tax software.
