LLMs Mislead Tax Filers With Calculation Errors

Boston University professor Dokyun Lee warns as the April 15 filing deadline approaches that taxpayers should not rely on large language models to prepare final tax returns because models make calculation, eligibility, and context errors and risk leaking sensitive data. He cites TaxCalcBench (July 2025) finding models correctly computed fewer than one-third of federal-only returns, urging LLMs be used only for research and preparation.