该情境显示人工智能在报税领域的采用率显著上升：美国计划使用AI报税的劳动者比例从去年的11%上升至约25%，增幅约14个百分点，接近2.3倍增长。个体案例中，用户通过一次性上传PDF文件并生成多标签电子表格，将原本需数小时人工整理的流程压缩至单次操作，体现效率提升。但该效率收益伴随高风险暴露，尤其在复杂税务结构（如抵押贷款、小企业收入、多来源咨询收入）下，错误放大的潜在成本更高。
错误率与系统局限呈现量化特征：研究显示大型语言模型在典型税务问题上的失败率约为2/3（≈66%），而即使达到99%准确率，在“每份报税包含大量字段”的条件下，统计上仍接近“每份申报必有1处错误”的水平。错误类型集中于数字识别偏差、过时法规引用及复杂规则分支处理失败，例如对K-1或1099表格解析不稳定，以及未识别最新税法调整。实际案例中，AI错误建议导致用户遗漏低于3000美元的加密货币收入申报，直接引发税务风险。
行为模式显示用户通过多模型交叉验证试图降低不确定性，但结果可能分歧扩大，例如3个模型给出不同税额判断，最终选择最高值以规避风险，反映决策偏向保守极值。使用场景逐渐分化：AI更适合用于文档整理、概念解释及辅助沟通，而非最终计算与合规判断。行业策略亦体现限制性部署，大型软件将AI功能限定在导航与问答层面，避免核心计算自动化。整体趋势表明，在高复杂度与低容错（接近0%容忍）的税务系统中，AI当前处于“效率提升显著但可靠性不足”的不对称阶段。
The context shows a sharp rise in AI adoption for tax preparation: the share of US workers planning to use AI increased from 11% last year to about 25%, a gain of roughly 14 percentage points, or about 2.3× growth. In individual cases, users compressed hours of manual organization into a single workflow by uploading PDFs and generating structured spreadsheets, demonstrating efficiency gains. However, these gains come with elevated risk exposure, especially for complex tax profiles involving mortgages, small businesses, and multiple income streams, where errors scale with complexity.
Error rates and system limits are quantifiable: studies show large language models fail about two-thirds (~66%) of the time on standard tax questions. Even at 99% accuracy, the high number of fields in tax returns statistically implies roughly one error per filing. Errors cluster in digit misreading, outdated rule application, and failure to navigate branching exceptions, including unreliable handling of K-1 and 1099 forms and missed recent law changes. In one case, AI incorrectly advised that cryptocurrency income below $3,000 was not reportable, leading to compliance issues.
User behavior shows attempts to reduce uncertainty via multi-model comparison, yet divergence can increase, as three models may produce conflicting outputs, prompting selection of the highest estimate as a risk-averse strategy. Use cases are bifurcating: AI is effective for document organization, explanation, and communication, but unreliable for final calculations and compliance decisions. Industry deployment reflects constraints, with major software limiting AI to guidance layers rather than core processing. The overall trend indicates a mismatch: significant efficiency gains alongside insufficient reliability in a near-zero-error tolerance system.