Artificial intelligence (AI) is improving our lives by utilizing technology and machine learning to accomplish tasks that require considerable human labor. It can deliver similar and often better outcomes in a cost-effective manner. I have written here that calculative actions do not require much creativity thus are the most obvious fields in which machines are superior to humans. It is very fitting to ask, then, how can AI improve tax enforcement and compliance? Valuation is one of the most calculative and arduous areas in tax administration that automation can greatly improve.
Valuation often requires many efforts determining fair market value (FMV) when there is no willing buyer and seller that negotiate the asset’s purchase price. Because of the essential role that asset valuations play in determining tax liabilities, there is a high sensitivity by the IRS and taxpayers to their accuracy. Transactions between related parties or not at arm’s length such as transfer of bequests, nonfungible real estate, or closely held business interests present complex valuation issues as there is no clear FMV. Congress uses a traditional carrot-and-stick approach to valuation by encouraging taxpayers to comply through clarifying and simplifying reporting obligations, along with imposing penalties for misstating the value of assets.
Yet, the harder it is to value the asset, the more complicated the appraisal method becomes. Taxpayers frequently take advantage of this intricacy in an attempt to overvalue their assets to claim a higher tax deduction or in other instances undervalue their assets to lower their income or property tax due. Expensive professional expert opinions and appraisals become central to estimate the value of assets. Skilled appraisers utilize various methodologies, assumptions, and models that lead to wildly disparate results. With taxpayers lacking established sources of valuations for tax reporting and the IRS’s inadequate resources and expanded responsibilities, the agency becomes even more overburdened with oversight and controversies involving valuation. Moreover, a common practice of judges that lack valuation expertise is to split the difference between the litigating parties’ valuation in the middle, which further incentivizes the parties ex ante to assume extreme valuation positions. This inaccurate and meaningless middle-ground adjudication is reached after the parties wasted time, energy, and resources on appraisals and litigation. It is far from optimal tax practice as it results in high compliance costs, massive tax adjudication, and much revenue loss to Treasury. With recent discussions on expanding wealth tax, valuation issues will become even more pronounced in our tax system. The current formulaic or forced-sale valuation approaches are neither efficient nor accurate. While appraisers do make use of certain technology to compare assets it is still a costly and lengthy process that yields much litigation.
In this Article, Soled and Thomas suggest valuation reform utilizing AI. They demonstrate that there exist innovative ways by which machine learning can provide more accurate and uniform results and improve tax compliance. They survey studies and advances in AI that already proven to be more effective in valuing certain nonfungible assets compared to its human counterparts. Through machine learning the computer can analyze large volumes of data (millions of entries) of asset features (price, medium, size, etc.) from existing sales and learn to make predictions about prices for similar types of assets. For example, in valuation of private business interests, Goldman Sachs, JP Morgan, Morgan Stanley, and Bloomberg developed a machine learning tool that relies on publicly available financial data, industry classification, and other measures as a learning data set. Studies conducted on auctioned paintings were found to provide accurate price prediction superior to current standard economic pricing models. Other studies proved AI delivered more precise results in valuation of multifamily real estate compared to traditional methods of real estate appraisals that the passage of time makes them inherently inaccurate. The recent pandemic is another case in point. It became widely apparent that old appraisal models have been lacking and have caused buyers to walk away from a sale because their lenders undervalued their prospective homes and did not take into account extreme market supply-and-demand conditions and acutely limited inventory.
AI’s ability to go over millions of combinations of variables creates a more accurate valuation in a cost-efficient manner. Nevertheless, implementation of an automated approach to valuation is not without flaws. Designing, modeling, and testing complex algorisms involve high sunk costs. Mistakes can be detrimental. The Zillow episode is a ripe example. The popular real estate website used machine learning to create algorithms to value residential real estate prices to the general public of sellers and buyers. Thereafter, the AI platform provided offers at the spot price to homeowners without physically examining the property. Alas, the Zillow algorithm did not take into account factors such as the condition of the house, market demand, etc. and the company lost millions of dollars over inflated prices of houses it bought. This is an example where human supervision and testing of automated processes are crucial to successful implementation of AI.
Yet, designed properly, where data critical to valuation is readily available, these technological tools that utilize large information flow and machine learning tools can solve concrete practical problems in a more cost-effective, quicker, and simpler way. The more data is collected and examined to compare the assets at issue with other similar assets, the more accurate the valuation is. Moreover, because machine learning can produce valuations faster, it can better account for changing conditions thus produce more precise valuations than traditional methods that take time to update.
Soled and Thomas call on Congress to either develop independent valuation AI platforms or delegate this task to Treasury to contract with the private sector as it did with tax return preparation services. Other foreign and local government have already begun to capitalize on machine learning in asset valuation. It is no longer a private market gadget but a method the public sector is learning to appreciate as it offers important benefits such as faster and more accurate results compared to man-made traditional processes susceptible to bias and errors.
Alas, the authors assume there exists a neutral algorithm that is based on the most current and comprehensive data. Having been married to an AI developer I quickly learned that not all algorithms are created equal. The elephant in the room remains high net-worth taxpayers with deep pockets and large incentives to aggressively game the system by building their own AI models and proving their valuation schemes turn to their favor. The battle of appraisers could be then replaced with a different battle—the accuracy of algorithms. Courts would need to hire expert software programmers to decide which AI is more precise, which may make the process even less efficient or cost-effective. Accordingly, Soled and Thomas’ incremental implementation proposal is wise. They propose to focus on real estate and expand it gradually to other types of assets. Yet, I suggest this gradual implementation approach will relate to phasing out of old valuation models as well. It is certainly more politically feasible to utilize new AI tools parallel to current methods making it an accessible safe harbor to the public. A concurrent approach of choosing to relay on AI OR man-made appraisals will allow Treasury to gain enough experience and confidence (as well as withstand litigation) to abandon outdated appraisal methods. Automation of asset valuation is inevitable; nevertheless, it should be done with prudence.