General-purpose AI
The current iteration of commercialized Artificial Intelligence mostly surrounds the concept of Large Language Models (LLMs). Unlike indispensable tax software we rely on, which is programmed and dedicated to performing calculations, LLMs have fundamental design differences which currently limit its ability to replace bespoke tax software.
Without getting too technical, the major differences are:
LLMs are trained on large text datasets, to predict the most probable next word in a sequence. LLMs, at their core, are statistical operators, rather than arithmetical or logical operators. This predictive nature is not conducive to accurately handling complex calculations.
LLMs generally lack the memory and multi-step reasoning needed to handle the nuances of tax calculations. While performing a complex calculation, the LLM may "forget" prior details. Take a look at the 10-key on your desk: you have a tape to look at, and memory keys to utilize - LLMs don't necessarily have those tools, or at least the skillset to use them reliably.
Most importantly, LLMs don't have an actual comprehension of what math is - even as a concept. They can recognize patterns, interpolate, extrapolate, and make inferences; however, tasks involving calculus, linear algebra, and - you guessed it - taxes, all require specialized algorithms which are not native to the current generation of LLM architecture.
TaxGPT vs. Calculations
With the prior points in mind, lets dive into what TaxGPT can and cannot do (yet). For starters, no current "AI" offerings - even those offered by the biggest tax software providers - actually have the architecture to overcome the limitations I've already mentioned.
That said, TaxGPT can perform calculations based on information that is provided to it - it just needs to be provided enough data to perform the entire calculation from start to finish. For example:
Uploading a tax return summary for an individual client, you can ask if the taxpayer(s) qualified to make IRA contributions. Based on the provided data points of earned income, modified adjusted gross income threshold, etc., TaxGPT can determine whether or not the client was qualified to make IRA contributions.
Asking for an analysis of payroll tax vs. self-employment tax in a single-member LLC vs. S corporation scenario.
Uploading a multi-year profit & loss statement and asking for an analysis of ratios or trends.
What AI can't do (on its own - yet):
Provide definitive guidance on Section 965 for a 40-year multinational corporation with dozens of subsidiaries.
perform a multi-year calculation and analysis of Alternative Minimum Tax.
Which is probably a good thing - tax professionals are just as relevant and necessary as ever. As always, AI is here to augment our abilities, not replace us. Always use your professional judgment as you interact with any tax tools!