Towards eXplainable Artificial Intelligence (XAI) in Tax Law: The Need for a Minimum Legal Standard

The case law does not give a clear answer on what degree of human involvement is required to avoid classification as automated decision-making and its consequences.

“the case does not provide a clear-cut answer as to what degree of human involvement is needed to avoid the label of automated decision-making and its implications.205”

AI Act transparency requirements apply only to high-risk systems, a classification that excludes tax-domain applications beyond those described in the paper.

“By design, they only apply to high-risk systems, 232 a classification that excludes applications in the tax domain beyond those described above.”

Many AI Act transparency requirements are directed at the user of the system, so taxpayers cannot invoke them to obtain information about systems used by tax authorities.

“This means taxpayersindividual or otherwise -cannot invoke the transparency requirements laid down in the AI Act to obtain information about the systems used by tax authorities.234”

The AI Act imposes no obligation to provide local post hoc explanations of specific decisions made with the involvement of an AI system.

“There is no obligation to provide local post hoc explanations of specific decisions made with the involvement of an AI system.”

Current data protection law does not establish a general duty of explanation for AI systems, even where those systems rely on personal data.

“From the analysis above, the authors conclude that current data protection law does not establish a general duty of explanation for AI systems, even when these systems depend on personal data.”

The broad set of transparency measures in the AI Act does not require explanations of high-risk AI systems or of their output.

“Yet, this broad set of transparency measures does not require explanations of high-risk AI systems or their output.”

Explanation under the AI Act is directed at the users who deploy AI systems for tax applications, not at the taxpayer.

“Explanation under the AI Act is directed towards the users who deploy AI systems for tax applications, not the taxpayer.”

The authors found no jurisdiction that extends the right to an explanation to corporations and other non-natural entities.

“Nevertheless, the authors were not able to find any jurisdiction that extends the right to an explanation to corporations and other non-natural entities.”

• AssumptionRests on an assumption: The authors' search across jurisdictions was comprehensive enough that not finding any jurisdiction extending the right to an explanation to legal persons indicates that no such jurisdiction exists.

Initiatives of individual countries and international organizations on XAI in tax law are close to nil.

“the initiatives of individual countries and international organizations in that field are close to nil.”

The OECD AI Recommendation and Council of Europe Recommendation CM/Rec(2020)1 require deployers and users of AI systems, in all settings including tax matters, to ensure that everyone affected by the systems' outcomes receives a relevant explanation and can effectively challenge a negative outcome.

“The OECD AI Recommendation and Council of Europe Recommendation CM/Rec (2020)1 266 both require deployers and users of AI systems -in all settings, including tax matters -to ensure that everybody affected by the outcomes produced by such systems receives a relevant explanation and has the right to effectively challenge a negative outcome.”

• cites OECD AI RecommendationNot checked yet

  Not checked in the free run.

• cites Council of Europe Recommendation CM/Rec (2020)1Not checked yet

  Not checked in the free run.

There is a glaring lack of regulations devoted to tax law, particularly regarding XAI.

“There is, however, a glaring lack of regulations devoted to the field of tax law, particularly in respect of XAI.”

The GDPR applies more broadly, covering all non-law-enforcement contexts in which the personal data of natural-person taxpayers is processed.

“while the GDPR is more broadly applicable, covering all non-law enforcement contexts in which the personal data of natural person taxpayers is processed.”

The European Union's main legal instruments on AI make no explicit mention of tax matters.

“The European Union is generally deemed to be at the forefront of global AI regulation, yet its main legal instruments in that field make no explicit mention of tax matters.”

A general right to explanation is found in the GDPR.

“a general right to explanation is to be found in the GDPR”

The OECD promotes the use of AI systems under the motto of seamless and frictionless tax administration.

“Instead, the use of AI systems is promoted by the OECD under the motto of “seamless and frictionless tax administration”.”

• cites OECD, supra n. 8Not checked yet

  Not checked in the free run.

In many cases, secrecy can ensue from the law just as easily as from technical complexity.

“Indeed, section 2.2.shows that, in many cases, secrecy can just as easily ensue from the law as from technical complexity.”

AI systems might be employed in ways that remove taxpayers' ability to challenge decisions that may be wrong.

“or they may be used in ways that deprive taxpayers of their right to contest potentially wrongful decisions.11”

• cites [11] sourceNot checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: Taxpayers have a right to contest potentially wrongful tax decisions.

Existing soft and hard law instruments on AI do not sufficiently address XAI in tax law because they fail to address the tax domain's particular information needs.

“Still, it is insufficiently addressed by soft and hard law instruments on AI, which do not address the particular information needs of the tax domain.”

After these effects were revealed, the backlash led Dutch Prime Minister Mark Rutte to resign.

“Once these effects were revealed, the backlash prompted Mark Rutte, Prime Minister of the Netherlands, to resign.”

• cites [60] sourceNot checked yet

  Not checked in the free run.

The Dutch childcare benefits scandal exposed systematic discrimination affecting tens of thousands of welfare recipients.

“A strong example comes from the Netherlands, in which the so-called childcare benefits scandal (toeslagenaffaire) has unveiled systematic discrimination against tens of thousands of welfare recipients.”

• cites [57]Not checked yet

  Not checked in the free run.

Global and supranational bodies, particularly the OECD, openly urge tax administrations worldwide to adopt AI systems for their tasks as they move from Tax Administration 2.0 to 3.0.

“International and supranational organizations, notably the OECD, explicitly encourage tax administrations from all over the world to implement AI systems in support of their tasks as part of the transformation from Tax Administration 2.0 to 3.0.”

• cites [46]Not checked yet

  Not checked in the free run.

No matter how efficient and accurate a tax law AI system is, it will still make occasional prediction errors that may seriously affect taxpayers.

“Even the most efficient and accurate tax law AI system will still be plagued by the occasional prediction error, which could trigger serious consequences for the taxpayer.”

A major rise in tax administrations' use of AI is strongly disrupting the tax profession.

“At the same time, the significant increase in the use of AI by tax administrations dramatically disrupts the tax profession.”

Two obstacles may hinder the uptake of XAI techniques for tax AI systems.

“Two challenges could hamper the adoption of XAI techniques for tax AI systems.26”

• cites [26] sourceNot checked yet

  Not checked in the free run.

Maximizing AI use by tax administrations worldwide appears to be a necessity rather than a choice because of personnel shortages relative to workload and the intertwined digital-physical reality of the tax world.

“The maximization of the use of AI systems by tax administrations around the world would seem to be a necessity rather than a choice due to personnel shortages in that public domain relative to the volume of work 48 and the intertwined digital-physical reality of the tax world.”

• cites [48]Not checked yet

  Not checked in the free run.

• cites [49]Not checked yet

  Not checked in the free run.

A decision based on statistically relevant correlations may be legally unacceptable despite the relevance of facts, because it lacks legal grounds.

“Although facts are relevant for the lawfulness of a decision, a decision based on statistically relevant correlations might nevertheless be unacceptable due to a lack of legal grounds.27”

• cites [27] sourceNot checked yet

  Not checked in the free run.

The precise causal relationship between the data an AI system takes in and the decision it produces is often unclear.

“Thus, it is not always obvious how the system's input affects its output -in other words, what is the exact causal link between the data ingested by the system and the decision it makes.”

Explainable AI can help make tax-sector AI systems accountable by ensuring their operations and results are understandable to people.

“XAI can contribute to accountability in the tax domain by providing explainability to the AI systems in the tax sector, that is, by ensuring that the functioning and the outputs of these systems can be explained in a way that a human observer can understand.”

It follows that tax administrations employ AI systems for many purposes.

“Consequently, it is not surprising that tax administrations use AI systems for various purposes.”

In 2021, reports showed that the Dutch tax authority's risk assessment system violated the fundamental rights of about 35,000 welfare recipients.

“In 2021, it came to light that Dutch tax authorities were using a risk assessment system that infringed the fundamental rights of an estimated 35,000 welfare recipients.”

• cites [58]Not checked yet

  Not checked in the free run.

To prevent harmful AI outcomes and offer redress when they happen, researchers and practitioners have created several methods for algorithmic accountability.

“To prevent AI from having such undesirable consequences -and to provide some redress should those consequences come to pass -scholar and practitioners have developed various approaches to ensure algorithmic accountability.”

• cites [53]Not checked yet

  Not checked in the free run.

Tax administrations have an advantage over private sector stakeholders because they can access huge volumes of tax-sensitive data when implementing AI systems.

“One of the big advantages tax administrations have over private sector stakeholders is that the former can rely on having access to huge volumes of tax-sensitive data as they implement their AI systems.29”

• cites [29] sourceNot checked yet

  Not checked in the free run.

Additional explainability requirements may apply where the decision-making process of tax authorities is automated.

“Furthermore, additional explainability requirements may apply whenever the decision-making process of tax authorities is automated.”

• cites FR: Code on the relationships between the public and the administration, art.R. 311-3-1-2.Not checked yet

  Not checked in the free run.

Taxpayers cannot determine when a case will be treated as high or low risk, either before or after assessment.

“As a result, taxpayers cannot ascertain the circumstances under which a case will be classified as high or low risk, whether before or after the tax assessment.”

• cites [88] sourceNot checked yet

  Not checked in the free run.

A tax decision cannot be justified if it is not explainable, for example where it follows the recommendations of a non-explainable AI system.

“Such justification is impossible if tax decisions are not explainable, e.g. when they follow the recommendations of non-explainable AI systems.”

The algorithms behind the STIR risk indicator are never revealed to the taxpayer or to courts, including during litigation.

“The AI algorithms used to determine such risk are never disclosed to the taxpayer or the courts, even during litigation between the taxpayer and the NRA”

• cites [84] sourceNot checked yet

  Not checked in the free run.

Tax secrecy for AI systems is justified by the need to keep tax inspections uniform and effective.

“Tax secrecy in connection with AI systems is justified by the need to ensure the uniformity and effectiveness of tax inspections.”

• cites [89] sourceNot checked yet

  Not checked in the free run.

• AssumptionThe authors state this as an assumption: Taxpayers who learn the risk-assessment criteria or the algorithms' modes of functioning could circumvent the filters and escape tax inspection despite being involved in tax fraud or non-compliance.

In simple cases decided on objective conditions, such as the participation exemption once the holding conditions are met, the required explanation is weaker: the tax authority need only state that the conditions set by the applicable legal text are fulfilled.

“The tax authority only has to explain in its decision that the conditions required by the applicable legal text are fulfilled.”

• cites Art. 3 Council Directive 2011/96/EUNot checked yet

  Not checked in the free run.

Tax secrecy therefore affects taxpayers unequally.

“This shows that tax secrecy does not affect all taxpayers equally.”

Using AI in government-to-citizen (tax authority-to-taxpayer) relations generates stronger legal explainability obligations than using it in B2C or B2B relations.

“In consequence, using AI in G2C relations creates stronger legal obligations of explainability than in B2C and B2B relationships.”

• cites Bibal et al., supra n. 17, at sec.2.2.Not checked yet

  Not checked in the free run.

To satisfy the principle of formal motivation, tax authorities must mention and explain all factual and legal grounds on which the decision is based.

“To comply with this principle, all factual and legal grounds on which the decision is based should be mentioned and explained by the tax authorities.”

• cites Id., at sec.2.2.1.Not checked yet

  Not checked in the free run.

Decisions of tax authorities must comply with the principle of formal motivation.

“In particular, the decisions of tax authorities must comply with the principle of formal motivation.”

• cites Bibal et al., supra n. 17, at sec.2.2.Not checked yet

  Not checked in the free run.

Poland's STIR system computes a risk indicator for each entrepreneur using undisclosed AI algorithms based on financial-sector criteria for combating VAT fraud.

“In Poland, the Clearing House System (System Teleinformatyczny Izby Rozliczeniowej, STIR) establishes a risk indicator for each entrepreneur, which is calculated using secret AI algorithms based on criteria used by the financial sector to combat VAT fraud.”

• cites [83]Not checked yet

  Not checked in the free run.

The level of explanation required rises in proportion to the degree of discretionary power held by the tax authorities.

“The bar for explanation is raised proportionally to the level of discretionary power enjoyed by the tax authorities.”

• cites J.-L.Autin, La motivation des actes administratifs unilatérauxNot checked yet

  Not checked in the free run.

The limits of constitutionally compliant use of AI systems in tax law are set by the constitutional requirement that taxes be levied only under a statutory law in a precise and predictable manner (the principle of no taxation without representation), which follows from the rule of law in taxation.

“the limits for constitutionally compliant use of AI systems in tax law are set by the constitutional requirement that taxes only be levied employing a statutory law in a precise and predictable manner (the principle of”

• cites J. Hattingh, The Multilateral Instrument from a Legal Perspective: What May Be the Challenges?Not checked yet

  Not checked in the free run.

Without an explainable AI system, whether because of tax or trade secrecy or technological complexity, it would not be possible to determine whether the use of AI systems conflicts with constitutional principles, for instance through arbitrary or discriminatory treatment.

“However, without an explainable AI system -either due to tax or trade secrecy and/or technological complexity -it would not be possible to determine whether the use of AI systems contradicts constitutional principles, e.g.due to its arbitrary or discriminatory character.”

Such problems will grow more common unless those who design, deploy, apply, and oversee AI systems act quickly to make them explainable to stakeholders.

“Problems of this sort will only become more frequent unless proper actions are taken promptly by entities involved in designing, deploying, applying and supervising AI systems to ensure their explainability in ways relevant to their stakeholders.”

• cites [61]Not checked yet

  Not checked in the free run.

Taxpayers receive no meaningful description of how the AI algorithms basically function.

“Nor are they afforded an even marginal description of the basic functioning of the AI algorithms.”

• cites [88] sourceNot checked yet

  Not checked in the free run.

Several Latin American tax administrations (Chile, Peru, Brazil, Colombia, Ecuador, Argentina, Mexico) lack procedures letting taxpayers obtain information about how AI systems used in tax audits work.

“Scholars from South America have identified a lack of procedures for taxpayers to obtain information about the functioning of AI systems in tax audits to detect tax fraud and measure the risks of compliance with tax obligations by Latin American tax administrations, such as those of Chile, Peru, Brazil, Colombia, Ecuador, Argentina and Mexico.”

• cites [82]Not checked yet

  Not checked in the free run.

Where AI tax-oriented systems are used only in business-to-consumer or business-to-business relations, the applicable explainability requirements remain weak and narrow.

“As long as AI tax-oriented systems are used only in B2C or B2B relations, explainability requirements will remain weak and narrow.”

• cites Directive 2011/83Not checked yet

  Not checked in the free run.

• cites Regulation 2019/1150Not checked yet

  Not checked in the free run.

• cites Directive 2014/65Not checked yet

  Not checked in the free run.

A sufficient level of explainability of an AI system is a building block towards making its use compatible with the constitutional principles of legal certainty and statutory taxation, which embody legal precision and predictability of taxation.

“level of AI system explainability constitutes a building block towards the compatibility of its use with the constitutional principles of legal certainty and statutory taxation that embodies the principle of legal precision and predictability of taxation.”

Explanation is always required when AI systems are used to render decisions in tax authority-to-taxpayer relationships.

“This shows that explanation is always required when AI systems are used to render decisions in TA2T relationships.”

• AssumptionRests on an assumption: The great majority of relationships in the tax domain are between tax authorities and taxpayers.

Tax secrecy does not stop large tax consultancy firms from systematically reviewing their cases to gain insight into how the system works, in exchange for payment.

“Tax secrecy does not prevent large tax consultancy firms, before such reverse engineering takes place, from systematically evaluating cases in which they have been involved to provide insights into the system's functioning in exchange for monetary remuneration.”

• cites [90]Not checked yet

  Not checked in the free run.

The constitutional principle of no taxation without representation can be regarded as a constitutional requirement for explainable AI in tax law.

“The constitutional principle of”

The OECD, backed politically by the G20, has pushed nearly 100 countries and jurisdictions to adopt a minimum standard against tax treaty abuse.

“the OECD -with political support from the G20 -has pushed nearly 100 countries and jurisdictions to implement a minimum standard to prevent tax treaty abuse.”

• cites [123]Not checked yet

  Not checked in the free run.

The principle of legal certainty, as recognized by EU law in ECJ case law, imposes similar requirements.

“The principle of legal certainty, as recognized by EU law in Court of Justice of the European Union (ECJ) case law, sets similar requirements.133”

• cites [133]Not checked yet

  Not checked in the free run.

In democratic countries, constitutions require that the calculation and levying of any tax be precisely and comprehensively regulated by statutory law.

“In principle, constitutions in democratic countries ensure that the calculation and levying of any tax must be precisely and comprehensively regulated by statutory law.129”

• cites [129] sourceNot checked yet

  Not checked in the free run.

XAI technologies may therefore be necessary in practice to make constitutional rights effective when AI is used.

“Consequently, XAI technologies might be de facto necessary to ensure the effectiveness of constitutional rights when AI is used.”

Direct interpretability is not always feasible for tax AI.

“However, direct interpretability is not always feasible for tax AI.”

Explainability makes AI systems function more easily, cheaply, quickly and accurately by helping to find and remove software bugs.

“explainability contributes to the easier, cheaper, faster and more accurate functionality of such systems by enhancing the identification and elimination of bugs in the software used to run these systems.”

• cites [104]Not checked yet

  Not checked in the free run.

Explaining only some rather than all vital elements of an AI system may mislead judges and taxpayers.

“Focusing on only some, rather than all, vital elements of an AI system may lead both judges and taxpayers astray.”

• cites [100]Not checked yet

  Not checked in the free run.

AI applications in tax, which must handle large volumes of taxpayer data and achieve high accuracy in tasks like fraud detection, tend to use machine learning models that are not transparent by design.

“applications of AI to the tax domain -which need to deal with large volumes of taxpayer data and, in applications such as tax fraud detection, achieve high levels of accuracy -tend to rely on machine learning models that are not transparent by design.”

• cites [111] sourceNot checked yet

  Not checked in the free run.

Any decision by a tax authority that affects taxpayers must be issued in a sufficiently clear, precise and predictable manner.

“For that purpose, any decisions by a tax authority that have an impact on taxpayers must be issued in a sufficiently clear, precise and predictable manner.132”

• cites [132] sourceNot checked yet

  Not checked in the free run.

Using opaque AI technologies in tax administration may create obstacles to complying with the constitutional requirements of legal certainty and accountability.

“Consequently, the use of opaque AI technologies in a tax context might introduce obstacles to compliance with the constitutional requirements of legal certainty and accountability surrounding the activities of tax authorities.135”

• cites [135]Not checked yet

  Not checked in the free run.

From a normative perspective, merely introducing AI technologies does not remove the constraints to which tax authorities are subject.

“As these are general provisions of tax law, the mere introduction of AI technologies into the fold is not enough -at least from a normative perspective -to void the constraints to which tax authorities are subject.134”

• cites [134]Not checked yet

  Not checked in the free run.

Failing to identify bugs at late stages typically causes severe and sometimes irreversible harm to stakeholders, including society.

“The lack of such identification at the mentioned late stages typically causes severe and sometimes even irreversible consequences to various stakeholders, including society as a whole.”

• cites [106]Not checked yet

  Not checked in the free run.

Several provisions of the ECHR could be seen as requirements for the explainability of AI systems.

“Several provisions of the ECHR could be seen as requirements for the explainability of AI systems.”

In these cases, the right to a fair trial ensures that taxpayers have an effective review of the information on which the tax authorities base their decisions.

“In these cases, the right to a fair trial ensures taxpayers have an effective review of the information on which the tax authorities base their decisions.154”

• cites [154]Not checked yet

  Not checked in the free run.

Non-explainable AI systems cannot be effectively supervised in the sense specified by the Slovak Constitutional Court.

“In light of these requirements, non-explainable AI systems cannot be effectively supervised in the sense specified by the Slovak Constitutional Court.140”

• cites [140]Not checked yet

  Not checked in the free run.

The court made an implicit connection with the right to a fair trial under article 6 of the ECHR, because the inability to lodge an effective defence stems from the same features of the SyRI legislation, namely its lack of transparency, which rendered the system inexplainable by law.

“This reveals that the court made an implicit connection with the right to a fair trial in article 6 of the ECHR, since the inability to lodge an effective defence stems from the same features of the SyRI legislation, i.e. its lack of transparency in respect of the functioning of SyRI, thereby rendering that system inexplainable by law.159”

• cites [159]Not checked yet

  Not checked in the free run.

The proportionality requirement also requires determining the patterns used for automated assessment and the models or other databases that lead to a particular decision.

“Also, the mentioned proportionality requires the determination of patterns used for automated assessment and models or other databases that lead to a particular decision (paragraphs 137-138).”

• cites [137] sourceNot checked yet

  Not checked in the free run.

The Slovak Constitutional Court found that legislation permitting AI systems without effective supervision is incompatible with the constitution because it fails to ensure proportional application of the technology.

“The Slovak Constitutional Court was also vocal about the incompatibility with the constitution of legislation that permits the use of AI systems without effective supervision, thereby failing to ensure the proportional application of the technology (paragraph 129).139”

• cites [137] sourceNot checked yet

  Not checked in the free run.

• cites [139]Not checked yet

  Not checked in the free run.

As confirmed by the ECJ in Fransson, the EU Charter does not cover purely domestic situations or situations not caused by an application of EU law by the EU Member States.

“Consequently, as confirmed by the ECJ in Fransson (C-617/10), the EU Charter does not cover purely domestic situations or situations that are not caused by an application of EU law by the EU Member States.152”

• cites [152]Not checked yet

  Not checked in the free run.

Any legislation that provides for the use of an AI system without granting its addressees, such as taxpayers, the right to receive a sufficient explanation of its functioning would not appear to pass the muster of legality in a democratic country.

“Indeed, any legislation that provides for the use of an AI system without granting to its addressees (e.g.taxpayers) the right to receive a sufficient explanation of its functioning would not appear to pass the muster of legality in a democratic country.160”

• cites [160]Not checked yet

  Not checked in the free run.

Accordingly, article 6 of the ECHR is a relevant legal source ensuring the right to a fair trial in disputes triggered by applying AI systems to tax law only when they arise out of the imposition of tax surcharges on taxpayers.

“Accordingly, article 6 of the ECHR constitutes a relevant legal source that ensures the right to a fair trial in disputes triggered by applying AI systems to tax law only when they arise out of the imposition of tax surcharges on taxpayers.”

Hence, the right to a fair trial under the ECHR provides limited assistance in ensuring the explainability of AI systems in tax law, even within its geographical scope of application.

“Hence, the right to a fair trial under the ECHR is of limited assistance in ensuring the explainability of AI systems in tax law, even within its geographical scope of application.”

In eKasa, the Slovak Constitutional Court implicitly underscored the need for explainability of AI systems in light of constitutional principles.

“In eKasa (17 Dec. 2021), 137 the Slovak Constitutional Court implicitly underscored the need for the explainability of AI systems in light of constitutional principles.138”

• cites [137] sourceNot checked yet

  Not checked in the free run.

• cites [138]Not checked yet

  Not checked in the free run.

The reason for this limited assistance is not that the right to a fair trial lacks merit for XAI in tax law, but its limited personal, substantive and territorial scope of application.

“The reason for this does not lie in the inadequate merits of that right to XAI in tax law but in its limited personal, substantive and territorial scope of application.”

Sufficient explainability appears to be a pivotal feature of all constitutionally compliant AI systems in the tax domain.

“Accordingly, sufficient explainability appears to be a pivotal feature of all constitutionally compliant AI systems in the tax domain.141”

• cites [141] sourceNot checked yet

  Not checked in the free run.

The right to respect for private and family life under article 8 of the ECHR is connected to explanation requirements in the ECHR.

“Another ECHR right connected to explanation requirements in the ECHR is the right to respect for private and family life, enshrined in article 8 of the Convention.”

Article 8 of the ECHR has become one of the key fundamental rights affecting the legality of using AI systems in tax law due to the SyRI judgment of the Hague District Court.

“In fact, it has become one of the key fundamental rights affecting the legality of the use of AI systems in tax law due to the seminal and widely debated 156 judgment of the Hague District Court in the Netherlands in SyRI (Systeem Risico Indicatie) (5 Feb. 2020).157”

• cites [156]Not checked yet

  Not checked in the free run.

• cites [157] sourceNot checked yet

  Not checked in the free run.

Failure to provide adequate explanations regarding the processing of this information and the uses of the outputs might violate the positive obligation to protect private and family life.

“Failure to provide adequate explanations regarding the processing of this information and the uses of the outputs of this processing might therefore violate the positive obligation of protection of private and family life.161”

• cites [161]Not checked yet

  Not checked in the free run.

Article 8(1) of the ECHR also establishes positive obligations on the state.

“In addition to the non-interference requirement found by the Dutch court in this case, article 8(1) of the ECHR also establishes positive obligations to the state.”

• cites [161]Not checked yet

  Not checked in the free run.

In Ferrazzini v. Italy, the ECtHR Grand Chamber significantly restricted the applicability of article 6 of the ECHR to tax law cases.

“Regarding the right to a fair trial in the ECHR, its applicability to tax law cases was significantly restricted by the ECtHR (Grand Chamber) in Ferrazzini v. Italy (12 July 2001).144”

• cites [144]Not checked yet

  Not checked in the free run.

Parties to the ECHR retain the power to extend their understanding of the right to a fair trial beyond the minimum required by the ECHR.

“However, parties to the convention still retain the power to extend their understanding of the right to a fair trial beyond the minimum required by the ECHR.”

• cites [155]Not checked yet

  Not checked in the free run.

Many uses of AI in the tax domain fall outside the explanation requirements contained in the GDPR.

“This movement is necessary because many uses of AI in the tax domain are not covered by the explanation requirements present in the GDPR.”

Current international and national law does not establish legally binding requirements for XAI in tax.

“Current international and national law does not establish legally binding requirements for XAI in tax, 173 but some soft law instruments engage with matters of AI and explanation, which are scrutinized in section 4.2.”

• cites [173] sourceNot checked yet

  Not checked in the free run.

AI systems that process only anonymized data or aggregate statistical data are not subject to GDPR-related explanation requirements.

“systems that only process anonymized data 193 or deal with statistical data at an aggregate level are not dealing with information that can be associated with a specific natural person.194 Therefore, they are not subject to any GDPR-related explanation requirements.”

Some soft law instruments engage with matters of AI and explanation.

“Current international and national law does not establish legally binding requirements for XAI in tax, 173 but some soft law instruments engage with matters of AI and explanation, which are scrutinized in section 4.2.”

• cites [173] sourceNot checked yet

  Not checked in the free run.

The GDPR locates explanation at the level of the individual decision rather than at the level of the system.

“As a result, the GDPR situates explanation at the level of the decision and not at the level of the system.199”

The prohibition of discrimination in article 14 of the ECHR does not seem to provide a solid legal basis for ensuring the explainability of AI systems in tax law.

“-the prohibition of discrimination in article 14 of the ECHR does not seem to provide a solid legal basis for ensuring the explainability of AI systems in tax law.169”

• cites [169]Not checked yet

  Not checked in the free run.

The proposed AI Act regulates AI systems as products placed on the European single market, regardless of whether they rely on personal data.

“The second approach, which has its centre in the proposed AI Act, regulates AI systems as products put on the European single market, regardless of their reliance on personal data.”

• cites AI Act (Commission Proposal) (2021)Not checked yet

  Not checked in the free run.

The EU data protection law approach to explainability for AI is limited to processing data associated with natural persons.

“The first approach to such regulation within the EU data protection law framework is restricted to situations involving the processing of data that can be associated with natural persons”

• cites Regulation (EU) 2016/679Not checked yet

  Not checked in the free run.

The European Union exercises considerable influence over the regulation of digital technologies even beyond its borders.

“Since the European Union exercises considerable influence over the regulation of digital technologies even beyond its borders, 172 an analysis of the European approach to XAI regulation provides insights on the limited extent to which existing law creates requirements for explanation in tax AI systems.”

• cites [172]Not checked yet

  Not checked in the free run.

Where data about natural persons is used in decision-making, those natural persons hold the right to request an explanation of the decision.

“To the extent that data about natural persons is used in decision-making, the natural persons to which this data refers have the right to request an explanation of the decision.”

• cites articles 13-15 of the GDPRNot checked yet

  Not checked in the free run.

Some sources of hard and soft law could have important implications for adopting XAI in a given jurisdiction, not least by sketching the contours of a prospective minimum legal standard for XAI in tax law.

“Some of these sources of hard and soft law could have important implications for adopting XAI in a given jurisdiction, not least by sketching the contours of a prospective minimum legal standard for XAI in tax law.”

Many tax-domain AI applications, especially those involving corporate taxpayers, fall outside the scope of EU data protection law.

“meaning that many applications in the tax domain do not fall within its scope, notably those related to corporate taxpayers.”

In many other tax-related cases, the ECtHR either declared the applications inadmissible or found no violation of article 14 in conjunction with other ECHR provisions or Protocols.

“In many other tax-related cases, the ECtHR either declared that the applications were inadmissible or found no violation of the invoked article 14 of the ECHR in conjunction with other provisions of the ECHR or the Protocols.167”

• cites [167]Not checked yet

  Not checked in the free run.

A significant number of tax-law AI uses are not covered by either the EU data protection rules or the proposed AI Act.

“Although we will see that many applications of AI in the tax domain fall outside the scope of either of these regulations, their examination provides useful insights for designing explanation systems in tax contexts.”

AI regulation is not exhausted by data governance issues.

“Yet, data governance matters do not exhaust the role of AI regulation, 180 which is why the European Commission recently proposed a Regulation for a European Approach to AI. 181”

• cites Lehr & OhmNot checked yet

  Not checked in the free run.

• cites AI Act (Commission Proposal) (2021)Not checked yet

  Not checked in the free run.

Any processing of personal data for tax purposes is covered by the GDPR, even where the actual outcome does not amount to personal data.

“any processing of personal data for tax purposes, even if the actual outcome does not amount to personal data, is covered by the GDPR.192”

• cites LV: ECJ (Fifth Chamber), 24 Feb. 2022, Case C-175/20, SIA "SS" v. Valsts ieņēmumu dienestsNot checked yet

  Not checked in the free run.

Because data processing must comply with all applicable legal requirements, the design of data processing must account for requirements arising from the tax domain as well as those from data protection law.

“Since data processing must comply with all applicable legal requirements, the design of data processing must account not just for those requirements stemming from data protection law but also for those derived from the tax domain.207”

Any explanation drawn from AI Act technical documentation is global, addressing the AI system as a whole rather than specific decisions.

“Still, such an explanation will be global, as it focuses on the AI system in its entirety and not any specific decisions.”

The AI Act does not provide grounds for a general right to an explanation of tax decisions concerning a taxpayer.

“Consequently, the AI Act does not provide grounds for a general right to an explanation of tax decisions concerning a taxpayer.235”

Identifying whether meaningful human participation exists in a decision-making process can be difficult given the practical conditions under which the decision occurs.

“Identifying the existence or absence of meaningful human participation in a decision-making process can also be difficult in light of the conditions under which a decision happens in practice.203”

• cites Almada, supra n. 188Not checked yet

  Not checked in the free run.

• cites B. Wagner, Liable, but Not in Control?Ensuring Meaningful Human Agency in Automated Decision-Making SystemsNot checked yet

  Not checked in the free run.

Under article 11 of the AI Act, a system's technical documentation must provide its design specifications, including the general logic of the AI system and of the algorithms.

“Under article 11 of the AI Act, a system's technical documentation must provide its design specifications, including but not limited to "the general logic of the AI system and of the algorithms".230”

Article 14(3)(a) of the AI Act requires the identification and construction of technically feasible measures ensuring effective human oversight before a high-risk system is placed on the market or put into service.

“Likewise, article 14(3)(a) requires the identification and construction of technically feasible measures to ensure effective human oversight before a high-risk system is placed on the market or put into service.”

The GDPR's focus on individual decisions may be insufficient to capture systemic patterns in decision-making, such as those leading to algorithmic discrimination.

“This focus on individual decisions might, however, be insufficient to capture systemic patterns in decision-making, such as those leading to algorithmic discrimination.198”

• cites Pałka, supra n. 174Not checked yet

  Not checked in the free run.

Adopting local post hoc explanations can be useful for complying with the AI Act's monitoring requirements.

“Nevertheless, the adoption of local post hoc explanations can be useful for complying with monitoring requirements.”

Post hoc explanation measures are required only where no other way exists to address risk in a particular context of application.

“Therefore, post hoc explanation measures are only required if there is no other way to address risk in a particular context of application.”

Prior use of soft law on AI in interpretation has been demonstrated by the Slovak Constitutional Court's eKasa judgment, which referred to Recommendation CM/Rec(2020)1.

“At the level of interpretation, soft law regarding AI has already proved to be useful, as follows, for example, from the Slovak Constitutional Court's judgment in eKasa, in which the Court broadly referred to non-legally binding Recommendation CM/Rec(2020)1 of the Committee of Ministers to Member States on the human rights impacts of algorithmic systems.”

Austria's Data Protection Act extends a narrow set of data protection legal rights to legal persons.

“indeed Austria's Data Protection Act (DSG) (BGBI.I 165/99) extends a narrow set of data protection legal rights to legal persons.”

• cites Austria's Data Protection Act (DSG) (BGBI.I 165/99)Not checked yet

  Not checked in the free run.

From an instrumental perspective, data protection law gives natural persons in the tax domain, i.e. individual taxpayers, mechanisms to assert their rights only to the extent that decisions are based on their personal data.

“an instrumental perspective, it provides natural persons in the tax domain (individual taxpayers) with mechanisms they can use to assert their rights, but only to the extent that decisions are based on their personal data.”

A sufficiently detailed description of the system's logic might amount to an explanation.

“A sufficiently detailed description of this logic might amount to an explanation, as discussed in the context of the GDPR above.”

Data controllers have considerable leeway to avoid explaining, provided they offer alternative mechanisms that adequately protect data subject rights.

“Data controllers therefore have considerable leeway to avoid explaining so long as they provide alternative mechanisms that ensure the adequate protection of data subject rights.”

Explainable AI techniques are particularly useful for challenging decisions involving AI systems, because contesting a decision depends on suitable information about what was decided.

“They are particularly useful when it comes to challenging decisions involving AI systems -through the courts, administrative authorities, or other channels for complaint -as exercising the right to contest a decision depends on the existence of suitable information about what has been decided in the first place.208”

Member States are not prevented from establishing norms governing the processing of data referring to identified or identifiable legal persons.

“However, Member States are not blocked from establishing norms to govern the processing of data referring to identified or identifiable legal persons”

• cites Recital 14Not checked yet

  Not checked in the free run.

Because article 14(4)(c) requires human overseers to be able to interpret AI output correctly, post hoc explanations can be used to satisfy that requirement.

“Since article 14(4)(c) states that human overseers must be able to correctly interpret AI output, post hoc explanations can be used to meet this requirement.”

Under article 9(2) of the AI Act, high-risk AI systems must be accompanied by an effective risk management system allowing identification and analysis of known and foreseeable risks and estimation and evaluation of risks arising during operation.

“Under article 9(2) of the AI Act, high-risk AI systems must be accompanied by an effective risk management system that allows for the identification and analysis of known and foreseeable risks, as well as for the estimation and evaluation of risks that arise during operation.”

The exclusion of legal persons from the GDPR is not a legislative oversight but a constitutive property of the EU data protection regime.

“As Recital 14 states, this decision is not an oversight by the legislator, but a constitutive property of the EU data protection regime.”

• cites Recital 14Not checked yet

  Not checked in the free run.

International legal standards have not yet caught up with the constitutional diagnosis that AI in tax matters threatens taxpayer rights.

“However, international legal standards have yet to catch up with the constitutional diagnosis presented above.”

Council of Europe Recommendation CM/Rec(2020)1 demands high explainability standards when algorithmic systems make decisions that pose significant human rights risks.

“As far as contestability is concerned, Council of Europe Recommendation CM/REC (2020)1 requires that the explainability of processes and output be held to high standards where algorithmic systems are used in decision-making processes that carry high human rights risks.261”

• cites [261]Not checked yet

  Not checked in the free run.

XAI in tax law lacks a minimum legal standard.

“No Minimum Legal Standard for XAI in Tax Law and Fragmentation of the AI Legal Environment”

Human rights frameworks can contribute to shaping a minimum legal standard for the use of XAI in tax matters.

“even the well-established reticence to apply ECHR provisions to tax law matters is not enough to dispel the relevance of XAI for the lawful use of AI in the tax domain, which suggests that human rights frameworks can play a role in the development of a minimum legal standard for the use of XAI in tax matters.”

Tax secrecy will continue to be a problem when AI systems are used for tax purposes.

“Moreover, where AI systems serve tax purposes, the problem of tax secrecy will remain.”

• AssumptionRests on an assumption: AI systems are used in tax matters.

Ferrazzini indicates the ECtHR hesitates to apply ECHR fair trial guarantees in tax cases.

“While the Ferrazzini ruling shows the ECtHR is reluctant to extend the fair trial provisions of the ECHR to tax matters,”

Constitutional and human rights frameworks are strained by inexplainable AI in tax matters, since missing explanations can undermine the precision, transparency, and predictability that tax law and its administration require.

“As section 3. shows, current constitutional and human rights frameworks come under strain when inexplainable AI systems are used in tax matters, as the lack of explanation potentially erodes the precision, transparency and predictability expected of tax law and its execution.”

The AI legal environment is fragmented.

“No Minimum Legal Standard for XAI in Tax Law and Fragmentation of the AI Legal Environment”

Because inexplainable AI in tax matters affects taxpayers' ability to protect themselves from adverse tax authority decisions, adopting technical explanation measures appears necessary for the ECHR to be effective.

“deemed to impact the rights of taxpayers to effectively protect themselves from the adverse effects of tax authority decisions, and the adoption of technical explanation measures would therefore appear to be a requirement for ensuring the effectiveness of the ECHR itself.”

The same problem extends to all recommendations in national instruments that, to varying degrees, address the need for explainability of AI systems at the technological level.

“This problem also extends to all recommendations found in national instruments that, to varying extents, deal with the need for the explainability of AI systems at the technological level.”

XAI guidelines are pertinent to tax contexts.

“Therefore, the guidelines established for XAI can be considered relevant in tax matters, meaning that AI systems used by tax authorities should be capable of technical explainability as well as legal explainability.”

• AssumptionRests on an assumption: AI systems are used in tax matters.

EU law imposes only a limited obligation to adopt explanation in tax contexts.

“Consequently, these laws impose only a limited obligation to adopt explanation in tax contexts”

The Dutch first-instance SyRI ruling links breaches of private and family life with fair trial breaches, citing a common feature of AI legislation: lack of explainability.

“the SyRI judgment by a Dutch court of first instance finds an implicit connection between the violation of the right to respect for private and family life and the violation of the right to a fair trial because of a feature common to all legislation governing the use of AI systemstheir lack of explainability.265”

• cites [265]Not checked yet

  Not checked in the free run.

Until that Recommendation provision becomes hard law, displacing trade secrecy to enable AI explainability will likely remain hard.

“That is indeed very important for the effective explainability of AI systems, but as long as this point of the Recommendation is not implemented as hard law, it will prove difficult to displace trade secrecy for purposes of the explainability of AI systems.263”

• cites [263]Not checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: Council of Europe Recommendation CM/Rec(2020)1 is not implemented as hard law.

Trade secrecy cannot stand in the way of ensuring that affected persons receive explanations, as acknowledged in Council of Europe Recommendation CM/Rec(2020)1.

“Trade secrecy cannot stay in the way of achieving this end, as acknowledged in Council of Europe Recommendation CM/Rec (2020)1.”

• cites Council of Europe Recommendation CM/Rec (2020)1Not checked yet

  Not checked in the free run.

The OECD and Council of Europe recommendations could create tensions with tax and trade secrecy.

“The aforementioned recommendations could trigger tensions in terms of tax and trade secrecy, as discussed in section 2.2.”

ECHR case law gives reason to hold that legislators must ensure AI use does not violate fair trial, private and family life, or non-discrimination principles.

“Indeed, various judicial decisions within the framework of the ECHR provide grounds to maintain that legislators are obliged by case law to ensure that the use of AI does not interfere with principles such as the right to a fair trial, respect for private and family life and non-discrimination.264”

• cites [264]Not checked yet

  Not checked in the free run.

XAI will not solve every issue introduced by the opacity of AI in tax applications.

“This is not to say that XAI will solve every issue introduced by the opacity of AI in tax applications.”

The safeguards of Section III of the AI Act apply to judicial activity in tax-related cases.

“which means the safeguards of Section III of the AI Act apply to judicial activity in tax-related cases.”

Because the AI Act rests on the single-market competence rather than data protection law, it applies to contexts that fall outside the scope of data protection law.

“This change in legal grounds ensures that the AI Act applies to contexts that fall outside the scope of data protection law.”