The many technological developments of recent decades have created a new world of tools for tax administrations. Now, the use of artificial intelligence (AI) is generating both excitement and concern.

The growing use of AI by tax administrations raises questions that go beyond efficiency. Tax systems rely not only on enforcement but also on legitimacy, trust, and perceptions of fairness. If AI can help identify potential non-compliance, automate processes, and improve taxpayer services and experiences, it may contribute to stronger and more equitable tax systems. However, if poorly designed or insufficiently governed, AI could also reinforce existing inequalities, reduce transparency, and weaken public trust, leading to lower levels of voluntary compliance.

The question, therefore, is not whether tax administrations should adopt AI—its capacity to handle large and ever-increasing volumes of data is undeniable—but how they can do so in a manner that strengthens fairness and accountability while maintaining confidence in the tax system and the tax administration.

We can broadly classify the uses of AI in tax administration into two categories: inward-facing applications within tax administration processes and outward-facing applications—those that more directly affect the taxpayer’s experience.

Inward-facing AI applications include risk assessment, audit case selection, fraud detection, network analysis, predictive analytics, document analysis, and document drafting. In our opinion, AI should not replace humans in making important and difficult decisions. Rather, it should assist tax administration officials by processing the vast quantities of information they receive, in both structured and unstructured formats, to identify patterns that could otherwise remain hidden, detect anomalies, perform network analysis, identify high-risk tax returns, and improve compliance strategies.

Outward-facing, or taxpayer-facing, AI serves a different purpose: the objective is not enforcement but service and compliance facilitation. Chatbots, virtual assistants, automated response systems, and personalized guidance tools can help taxpayers navigate increasingly complex obligations and changing tax rules. By providing timely information, for instance when taxpayers file a tax return, these technologies may reduce compliance costs and improve the taxpayer experience. AI can also help ensure that consistent criteria are applied regardless of the channel used or the person providing assistance.

Taxpayers may also use AI to prepare tax-related documents and returns and, in some cases, even to decide whether to contest certain tax adjustments based on an analysis of outcomes for taxpayers in similar circumstances.

Together, inward- and outward-facing applications illustrate AI’s potential on both sides of tax administration: improving the identification and treatment of compliance risks while making interactions with taxpayers faster and easier. For administrations operating with limited human and financial resources, these gains can also allow scarce capacity to be directed toward cases requiring greater human judgment. Although AI may also be used to improve data quality, we must keep in mind that all this potential is limited by the availability, completeness, and accuracy of data. Viewed in this context, AI should be considered part of a wider process of digital transformation.

There are impressive examples of potential gains in both productivity and process standardization, with curated generative AI assistants being used by tax officials to respond to taxpayers, as covered in these posts on this blog: Sometimes Sofia does not know. And that is good and Optimus Project: “The Revolution of Artificial Intelligence in São Paulo’s Tax Processes”. Another example is the recently launched Kar Saathi initiative in India, presented by the tax administration: an AI assistant in the e-filing portal that helps taxpayers by answering questions such as how to claim a deduction.

We must keep one thing in mind: the promises of AI should not obscure the many challenges and potential risks associated with its use, such as algorithmic bias, a lack of transparency, and disparities in access to technology.

One of the most important risks associated with AI is algorithmic bias. Because AI systems learn from historical data, they may reproduce or amplify disparities already present in those data. In tax administration, this can affect taxpayer profiling, risk assessment, and audit selection even when protected characteristics are not explicitly included in a model. Outcomes also depend on the objectives embedded in the model: a system optimized only for enforcement returns may produce a different distribution of scrutiny than one that also incorporates fairness considerations.

The relevance of these concerns has been demonstrated in research showing that algorithms designed to maximize different enforcement outcomes do not necessarily produce the same distribution of audits. This analysis found that machine-learning models optimized solely to classify potential non-compliance can shift the focus of audit activity toward lower- or middle-income taxpayers, while a regression model optimized to predict the potential amount of tax adjustments would shift the focus toward higher-income taxpayers. The study further demonstrates that alternative model designs that take social parameters into account can produce different distributions of audit coverage across income levels, illustrating that algorithmic outcomes depend not only on the data used but also on the objectives embedded in the model itself[1]. Moreover, discrepancies might not result from explicit targeting but rather arise from the interaction between case-selection algorithms and characteristics associated with benefits, such as certain tax credits or subsidies directed toward particular groups of individuals. Another study illustrates how seemingly neutral risk models can generate unequal outcomes across demographic groups, even when the characteristics associated with those groups are not directly included in the algorithms.

The lesson for tax administrations is that AI should support, not substitute for, human judgment. Oversight should extend beyond the final decision to the design of risk indicators, selection of training data, testing for disparate impacts, and periodic review of models as new evidence emerges.

Two related principles are essential: explainability and accountability. Officials should be able to understand and challenge an AI-generated recommendation, while responsibility for decisions affecting taxpayers must remain with the tax administration and its public officials.

These principles are reflected in international AI governance frameworks, including the OECD AI Principles, the European Union AI Act, and the Council of Europe’s Framework Convention on Artificial Intelligence, which emphasize human oversight, transparency, accountability, and fundamental rights.

Beyond concerns about biased training data, tax administrations should also consider the integrity of the data and documents processed by AI systems. Recent research has demonstrated that generative AI can be vulnerable to prompt injection and other forms of adversarial manipulation, whereby hidden instructions embedded within digital documents influence the model’s outputs without the user’s knowledge.

For tax administrations, comparable risks could arise if AI systems are used to analyze taxpayer submissions, contracts, invoices, transfer pricing documentation, or other digital records. This underscores the need for robust controls to detect hidden instructions, validate source documents, and ensure that AI-generated recommendations are independently reviewed before they influence administrative decisions. Protecting the integrity of the data processed by AI is therefore just as important as ensuring the quality and fairness of the data used to train the models.

AI can make administrative reasoning harder to interpret, particularly when complex models influence risk profiling or enforcement decisions. Transparency therefore needs to operate at two levels: taxpayers should know when automated systems have contributed to a decision-making process, and officials should be able to understand, document, and even challenge the role played by the model. It is essential to maintain human accountability, address privacy concerns while documenting interactions, and use accurate, official data to reduce the risk of hallucinations. We must remember that identifying sources may not be enough.

The digital divide also deserves attention. Not all taxpayers possess the same level of digital literacy, access to technology, or ability to engage with digital services. These differences could leave some taxpayers at a disadvantage, or even worse make them a target of social engineering attacks that could affect their tax positions or abuse their potential benefits.

It is a good time to remember that technology is neither good nor bad; its effects depend on how we use it. AI can bring efficiency gains, but it is neither inherently a tool for equity nor a driver of inequality—and it is certainly not neutral.

Good luck!

[1] See Black, Elzayn, Chouldechova, Goldin, & Ho, 2022.

73 total views, 24 views today