Context

KPMG International hosted a roundtable discussion in June 2025, as part of the Global Responsible Tax Program, to explore the ethical implications of artificial intelligence (AI) in tax and revenue administration.

As governments around the world integrate AI technologies into their tax systems, it's essential to examine how these tools can be developed and deployed responsibly.

The session sought to understand the practical consequences of algorithmic decision-making in tax, including the risk of bias, the problem of black-box models and the regulatory and institutional challenges of maintaining trust, transparency and human oversight. Participants also reflected on the varying maturity of AI use across jurisdictions, the changing role of tax professionals, and the need for global cooperation on governance.

Held under the Chatham House Rule, the roundtable brought together a diverse group of experts (see list below). This write-up summarizes the personal views of participants and does not necessarily reflect the views of any particular organization, including KPMG.

Executive summary

  • High potential, high risk: Participants acknowledged the promise of AI in tax administration (from automating queries to targeting non-compliance) but stressed the dangers of insufficient safeguards, especially in high-stakes decisions.
  • Black box decision-making erodes trust: Even simple AI models can produce outcomes that taxpayers and auditors cannot explain. When decisions can't be understood or challenged, institutional credibility suffers.
  • Fairness requires clarity, not ignorance: Excluding sensitive variables like gender or ethnicity from models does not guarantee fairness. Without collecting and reviewing disaggregated data, administrations risk inadvertently reinforcing existing biases.
  • Institutional capacity and training gaps: As AI tools reshape workflows, tax authorities must ensure officials are trained to interpret, challenge and contextualize AI outputs. Without this, the benefits of innovation could be undermined by poor human oversight.
  • Fragmented global regulation is a growing risk: National and regional approaches to AI governance vary widely, from the European Union's (EU) risk-based AI Act to the US moratorium-style provisions. This makes international coordination increasingly urgent, though some participants note the importance of diversity even here for both innovation and avoidance of regulatory capture
  • Explainability and responsibility must align: Participants stressed the need for clear frameworks assigning responsibility and enabling explanation. Where decisions impact taxpayers, they must be reviewable by humans with appropriate authority.
  • The human remains essential: AI can enhance tax administration, but can't replace judgment, discretion or legitimacy. Building systems that augment, rather than obscure, human agency will be critical to ethical deployment.

Black box decision-making

When explainability is absent, accountability vanishes

Taxpayer trust cannot survive AI decisions that no one can explain or challenge. Participants agreed that a major ethical risk in tax AI systems lies not in their complexity but in their potential opacity. Even relatively straightforward models (those using 20 to 100 data points) can become “black boxes” if tax officials and taxpayers can't understand or audit the basis of decisions.

Examples like the Dutch childcare benefits scandal and the UK Post Office case illustrated the dangers: flawed systems, assumed to be completely reliable and shielded from scrutiny, created devastating real-world harms.

Participants called for mandatory transparency standards for AI models used in enforcement or assessment. Whether a decision is generated by a statistical model or a more advanced learning system, taxpayers must be able to ask, and receive a meaningful answer to, “why?”

Statistical fairness

Ignoring sensitive variables won’t eliminate bias

Fairness by ‘not looking’ can be as dangerous as unfair targeting. There was strong concern that many tax administrations do not adequately monitor whether AI systems produce biased results. Models that omit sensitive variables like, gender or ethnicity, can still yield skewed outcomes — they may discover and rely on proxy indicators that correlate with protected characteristics.

Participants noted that in some cases, excluding such data actually increases the number of false positives for disadvantaged groups. For example, if male taxpayers are more likely to commit fraud, a model that doesn't know gender may overestimate risk for women.

Collecting and analyzing disaggregated compliance data (something institutions like the International Monetary Fund (IMF) have advocated) is essential to detecting and correcting systemic bias. But this remains patchy in practice.

Human capacity and training

AI changes the role of the tax administrator

New technologies demand new skills and inclusive governance structures. AI tools are altering how tax professionals work. Risk detection, assessments and procedural decisions are increasingly assisted by AI, requiring staff to interpret and contextualize model outputs. This, participants noted, raises the bar for training and procedural clarity.

Tax administrators will need to understand not only tax law, but the structure, logic and limitations of the models they rely on. The blending of statistical reasoning with legal judgment creates a new kind of complexity — one many institutions are not yet ready for.

There were also concerns about who has access to these tools. In some administrations, AI development is confined to small internal groups, which may exacerbate issues of representation, particularly around gender and neurodiversity. Broader inclusion in governance and design processes was encouraged.

Regulatory fragmentation: Global technology, local laws

National AI strategies diverge — but tax data moves across borders

Participants emphasized the growing divergence in regulatory approaches. The EU’s AI Act takes a risk-based stance, while the US has adopted more deregulatory postures in some domains. Meanwhile, countries like Japan, Australia and Chile are piloting lighter-touch or sector-specific approaches.

This patchwork increases the risk of regulatory arbitrage (particularly problematic in taxation, where digital data and corporate structures often span multiple jurisdictions). It also complicates the prospects for coordinated standards on transparency, fairness and accountability.

Participants agreed on the need for functional regulation governing specific uses of AI, such as risk scoring or case selection, rather than abstract definitions of AI itself. They called for internationally coordinated, principles-based frameworks that emphasize human oversight and due process.

Practicality and realism

Gradualism and simplicity may be the most responsible path forward

Rather than ambitious end-to-end automation, participants encouraged tax administrations to focus on modest, high-impact use cases such as chatbots for taxpayer support or pre-filled returns using existing datasets. These applications offer clear efficiency gains and carry fewer ethical or legal risks.

Several speakers emphasized that AI should augment existing human-led processes, not bypass them. Hybrid systems that pair AI insights with human review and responsibility can build resilience while improving service delivery.

Caution was also urged against overreliance on AI-generated outputs for legal claims. Examples were cited where ‘hallucinated’ (e.g. factually incorrect or entirely fabricated) legal precedents appeared in taxpayer submissions, undermining both compliance and credibility.

Institutional culture and accountability

Responsibility must be built in. Innovation without liability creates systems that evade correction

Beyond the technology itself, participants stressed the importance of assigning clear responsibility within institutions. When decisions go wrong, there must be accountability — for design, oversight and outcomes.

Too often, organizational silos and unclear chains of command mean that no one takes responsibility when AI-driven decisions harm taxpayers. Without institutional structures that prioritize transparency and answerability, public trust will erode — regardless of technical performance. With them, performance is far easier to improve.

Participants pointed to the need for tax administrations to treat AI implementation as a governance issue, not just a technical upgrade. That includes embedding ethical review processes, maintaining documentation of model changes, and ensuring that all decisions are ultimately defensible by human staff. Well-designed AI enhancement could improve not only efficiency, but also organizational accountability, and perceptions of transparency and legitimacy by the taxpayers.

Contributors:

  • Joanna Bryson, Professor of Ethics and Technology, Hertie School, Berlin
  • David Hadwick, PhD researcher, DigiTax Centre of Excellence, University of Antwerp
  • Becky Holloway, Programme Director, Jericho
  • Neal Lawson, Partner, Jericho
  • Benita Mathew, Lecturer in AI and Fintech, University of Surrey
  • Dominic Mathon, Head of Tax & Treasury, RELX
  • Chris Morgan, Adviser to the Global Responsible Tax Program at KPMG International
  • Tim Sarson, Partner, UK Head of Tax Policy, KPMG in the UK
  • Grant Wardell-Johnson, Global Tax Policy Leader and Chair of the Global Tax Policy Leadership Group, KPMG International