Artificial-intelligence evangelists and alarmists share one conviction. Both believe, despite scant evidence so far, that AI is on the verge of replacing human knowledge workers. Today's large language models—ChatGPT, Claude, Gemini, Grok—can write code and polish prose. But would you trust them to do your taxes?
In principle, tax filing should be the ideal use case for an AI system. It is a highly routinized but complex process. It requires attention to detail, an understanding of ever-changing regulations and the ability to interpret ambiguous instructions.
But it's also a task that has little to no room for error. The Internal Revenue Service won't accept “The robot did it” as an excuse for misfilings, which can carry life-altering penalties for the taxpayer.
Given that specialized AI systems can produce gold-medal answers to International Math Olympiad questions, calculating taxes should be no big deal. But AI systems also fail to count how often the letter r appears in “strawberry.” This variation in capabilities is sometimes referred to as the “jagged frontier.”
Before they replace human workers, AI systems need to be able to do the required tasks—and consumers and employers need to trust that AI will do them properly.
Before they replace human workers, AI systems need to be able to do the required tasks—and consumers and employers need to trust that AI will do them properly. There have been hundreds of examples of lawyers relying on AI systems that “hallucinated” case law or misrepresented exhibits and filings. The consequences of this happening on your tax return could be severe.
Predictions are tricky, but as a mathematician working on economic and technology policy, I guess that Americans won't use large language models for their taxes in the year ahead. I know I don't trust them to complete my own return without error.
In the future, that will likely change. Models are already pretty good at extracting information from structured text, and AI systems are becoming more reliable with each update. By 2027 I anticipate using a large language model to help draft my filings, perhaps as a first pass before reviewing and submitting them myself. By 2028 I predict enough people will trust general-purpose AI to handle their individual taxes to have a material effect on companies like H&R Block and Intuit.
My timelines might be too conservative if there is a push to make reliable consumer AI products. Alternatively, regulatory or liability hurdles could slow progress, as governments grapple with privacy and other implications of AI-driven automation in sensitive areas like tax compliance.
Still, the ubiquity, complexity, and accountability tax filing requires make it a good litmus test of the AI zealots and doomsayers' beliefs. My research leads me to worry about the effect that AI may have on the labor market: Policymakers should be preparing for major economic disruptions.
TurboTax will be the canary in the coal mine. Until people can trust AI to do their taxes, they shouldn't trust it to do their jobs. That gives us a little time to prepare—but not much.
