Civil Law And Algorithmic Tax Enforcement Discrimination Claims In Europe

1. Introduction

Algorithmic tax enforcement discrimination occurs when a tax authority uses automated systems, risk scores, predictive analytics, profiling, data matching, or AI to identify taxpayers for audits, investigations, penalties, reassessments, collection measures, or fraud-risk treatment, and the system produces discriminatory or unjustified differences between persons or groups.

Examples include:

an AI system selecting taxpayers for audit disproportionately because of nationality or residence;

a tax-risk score using geographical or socio-economic proxies that disadvantage a protected ethnic group;

automated fraud detection producing more false positives for a particular group;

an algorithm treating cross-border taxpayers differently without objective justification;

automated tax reassessment based on inaccurate personal data;

an opaque algorithm making it impossible for a taxpayer to understand or challenge an adverse decision;

algorithmic profiling indirectly reproducing historical discrimination in tax administration.

There is an important legal qualification: European courts have not yet developed a large body of cases in which an AI tax-enforcement algorithm itself has been held liable for discrimination. The strongest legal framework therefore comes from combining EU tax-equality jurisprudence + GDPR automated-decision jurisprudence + general discrimination law + fundamental-rights principles.

2. Basic Legal Structure

An algorithmic tax-discrimination claim can be understood as:

TAX DATA → ALGORITHM → RISK SCORE → TAX ENFORCEMENT DECISION → DIFFERENTIAL IMPACT → DISCRIMINATION → LEGAL BREACH → DAMAGE → CAUSATION → REMEDY

The claimant normally has to establish several stages:

Existence of automated processing

Identification of the decision or enforcement measure

Different treatment or disproportionate impact

Protected ground or another EU-law equality principle

Comparable persons or situations

Lack of objective justification

Procedural/data-protection violation where applicable

Material or non-material damage

Causal connection

Appropriate remedy

3. Why Tax Algorithms Create a Special Discrimination Problem

Tax administrations increasingly have enormous quantities of information:

income records;

employment records;

property ownership;

bank information;

VAT transactions;

customs information;

cross-border financial data;

corporate ownership;

transaction histories;

previous audit results;

geographical information;

third-party reporting.

An algorithm may combine these variables and produce a risk classification.

For example:

Taxpayer A → low risk → no audit
Taxpayer B → high risk → automatic audit → additional assessment

The difficulty is that the algorithm may not expressly use a protected characteristic.

Instead, it may use proxy variables.

For example:

postcode → socio-economic characteristics → ethnic composition → higher risk score

or:

nationality → cross-border activity → higher fraud score → increased tax scrutiny

This creates the possibility of indirect discrimination.

4. EU Principle of Equality in Taxation

Direct taxation remains substantially within Member State competence. However, Member States must exercise that competence consistently with EU law.

The CJEU has repeatedly stated that national tax powers cannot be exercised in a manner contrary to applicable EU equality and free-movement principles. This is particularly important where algorithmic tax enforcement differentiates between residents and non-residents or between cross-border economic actors. (curia)

Therefore:

Fiscal autonomy ≠ freedom to discriminate.

An algorithm does not acquire legal immunity merely because it is used by a tax authority.

5. GDPR and Algorithmic Tax Enforcement

The GDPR is particularly important because tax authorities may process large quantities of personal data.

Article 22 GDPR

Article 22 provides protection against decisions based solely on automated processing, including profiling, where the decision produces legal effects or similarly significantly affects the individual. Exceptions exist, including where Union or Member State law authorises the automated decision and provides suitable safeguards. (EUR-Lex)

Importantly, GDPR Recital 71 expressly contemplates automated processing for fraud and tax-evasion monitoring and prevention, but says such processing should have appropriate safeguards, including information, human intervention, an opportunity to express one's view, an explanation, and the ability to challenge the decision. It also stresses minimising errors and preventing discriminatory effects. (EUR-Lex)

Thus:

Tax-fraud detection is not automatically unlawful merely because it is algorithmic.

But:

Tax-fraud detection also does not automatically escape GDPR safeguards.

6. GDPR Article 22 and Tax Decisions

Suppose a tax authority's algorithm gives a taxpayer a "95% tax-evasion risk" score.

If that score automatically causes:

an audit,

freezing of a refund,

reassessment,

collection action,

penalty,

or another significant legal consequence,

the claimant should examine whether the decision was solely automated and whether Article 22 applies.

The crucial questions are:

A. Was there meaningful human involvement?

A nominal human approval may not necessarily resolve the problem if the official simply accepts the algorithmic output without genuinely examining it.

B. Was the decision authorised by law?

If Article 22(2)(b) is relied upon, the relevant Union or Member State law must also provide suitable safeguards.

C. Can the taxpayer challenge the decision?

The existence of meaningful review becomes especially important when an algorithm produces an adverse tax consequence.

7. AI Act and Tax Administration

The EU AI Act adds another layer, but its application to tax enforcement requires care.

The AI Act specifically states that AI systems intended for administrative proceedings by tax and customs authorities are not to be classified as high-risk law-enforcement systems merely because they are used by those authorities for administrative tax purposes. (EUR-Lex)

Therefore, it would be incorrect to say:

"Every AI tax-enforcement system is automatically a high-risk AI system under the AI Act."

That is not the correct position.

However, other EU law remains applicable, particularly:

GDPR;

Charter of Fundamental Rights;

EU equality principles;

applicable free-movement rules;

national administrative law;

national constitutional principles;

national tax procedure;

judicial-review requirements.

The AI Act therefore does not eliminate discrimination or due-process claims merely because a tax algorithm falls outside a particular high-risk category.

8. Case Law

Case 1 — Finanzamt Köln-Altstadt v Schumacker

C-279/93, CJEU, 14 February 1995

This is one of the foundational European tax-equality cases.

The case concerned the different tax treatment of residents and non-residents.

The CJEU recognised that residents and non-residents are not normally in objectively comparable situations for direct taxation because their overall economic circumstances may be different.

However, where a non-resident is in a situation objectively comparable to a resident—particularly where virtually all income is earned in the taxing State—different treatment may constitute unlawful discrimination. (Infocuria)

Relevance to algorithmic tax enforcement

Suppose an algorithm assigns a higher audit risk merely because:

"non-resident = higher tax risk."

Schumacker shows that the tax authority cannot simply assume that residence is sufficient justification for every difference in treatment.

The authority must consider:

objective comparability;

the purpose of the tax rule;

actual economic circumstances;

justification.

Principle

Residence-based differentiation must be legally justified when comparable situations are treated differently.

Relevance: Very high — direct tax-discrimination authority, although not algorithmic.

9. Case 2 — Asscher v Staatssecretaris van Financiën

C-107/94, CJEU, 27 June 1996

Asscher concerned different tax treatment connected with residence and the exercise of economic activity.

The Court examined whether a tax distinction affecting a non-resident constituted discrimination under the Treaty freedoms.

The case is important because European tax discrimination can arise even when a rule does not expressly say:

"foreign nationals receive worse treatment."

A residence criterion can indirectly disadvantage nationals of other Member States.

Algorithmic relevance

Imagine:

Residence status → automated risk multiplier → increased tax investigation.

Even though the algorithm does not expressly classify persons by nationality, residence may operate as a proxy.

Asscher therefore supports examination of the substantive effect of tax criteria rather than merely their formal wording.

Principle

A formally neutral tax criterion may produce prohibited unequal treatment where it disproportionately disadvantages persons exercising EU freedoms.

10. Case 3 — Talotta v État belge

C-383/05, CJEU, 22 March 2007

This is particularly important for the present topic.

Mr Talotta was a non-resident taxpayer operating a restaurant in Belgium. Belgian legislation used minimum taxable bases for certain non-residents where adequate evidence was unavailable.

The CJEU held that EU law precluded income-tax rules imposing such minimum tax bases only on non-resident taxpayers. The justification based on effective fiscal supervision was insufficient. (Infocuria)

Why Talotta is highly relevant to algorithms

Imagine a tax algorithm operates:

insufficient information + non-resident → automatically assumed higher taxable income.

That resembles the structural problem in Talotta.

The algorithm cannot transform an otherwise problematic discriminatory rule into a lawful one merely because the differentiation is hidden inside software.

Key principle

Administrative difficulty or fiscal-supervision objectives do not automatically justify discriminatory tax treatment.

Algorithmic formula

Incomplete data → algorithmic assumption → higher tax burden → nationality/residence disadvantage → justification test

11. Case 4 — Gielen v Staatssecretaris van Financiën

C-440/08, CJEU, 18 March 2010

Gielen concerned a tax deduction available to entrepreneurs under an hours-based requirement and the position of non-resident taxpayers.

The CJEU examined whether the discriminatory effect could be neutralised by an optional alternative tax regime.

The Court's reasoning demonstrates an important point:

An unlawful discriminatory treatment cannot necessarily be cured merely by offering the taxpayer another theoretical route to tax treatment.

(Infocuria)

Algorithmic relevance

Suppose:

Algorithm A → automatically classifies cross-border taxpayer as high risk.

The administration cannot necessarily respond:

"The taxpayer could have entered another administrative procedure."

The actual legal effect of the discriminatory system must be examined.

Principle

The existence of an alternative legal mechanism does not automatically eliminate discriminatory treatment.

12. Case 5 — CHEZ Razpredelenie Bulgaria

C-83/14, CJEU Grand Chamber, 16 July 2015

CHEZ is an important authority on indirect discrimination and apparently neutral criteria.

Electricity meters were placed unusually high in districts predominantly inhabited by persons of Roma origin because of alleged concerns about tampering.

The CJEU examined:

direct discrimination;

indirect discrimination;

apparently neutral practices;

group disadvantage;

objective justification;

proportionality.

(Infocuria)

Algorithmic tax relevance

This case provides a strong conceptual framework for proxy discrimination.

A tax algorithm may never contain:

"ethnicity = Roma"

but may use:

postcode;

neighbourhood;

income pattern;

language;

migration history;

household characteristics.

If these variables disproportionately burden a protected group, the legal analysis cannot stop at:

"The algorithm does not contain an ethnicity field."

The court can examine the effect of the apparently neutral criterion.

Principle

Neutral-looking criteria can generate indirect discrimination when they place a protected group at a particular disadvantage.

13. Case 6 — SCHUFA Holding (Scoring)

C-634/21, CJEU, 7 December 2023

This is one of the most important algorithmic-decision cases for the present topic, although it is not a tax case.

SCHUFA generated automated credit scores.

The CJEU examined Article 22 GDPR and held that automated scoring can fall within the GDPR's protection where the score plays a decisive role in a subsequent decision producing legal or similarly significant effects. (curia)

Application to tax enforcement

Consider:

Taxpayer data → AI risk score → tax authority relies on score → audit/reassessment.

The important question becomes:

Is the algorithmic score merely an internal administrative tool, or does it effectively determine the taxpayer's treatment?

If the latter, Article 22 and related safeguards become considerably more important.

Principle

An algorithmic score can be legally significant even where another institution or official formally makes the final decision.

14. Case 7 — CK v Magistrat der Stadt Wien / Dun & Bradstreet Austria

C-203/22, CJEU, 27 February 2025

This is a major recent authority on algorithmic transparency.

The CJEU held that a person affected by automated decision-making is entitled to meaningful information about the logic involved, sufficiently detailed to enable the person to understand and challenge the decision. (Infocuria)

The case concerned automated credit scoring rather than taxation.

Application to tax algorithms

Suppose the tax authority says:

"Your tax-risk score was 87."

That alone may not provide meaningful transparency.

A taxpayer may need information enabling them to understand:

what categories of personal data were used;

which factors influenced the result;

how those factors affected the result;

whether inaccurate data were used;

whether the algorithm disproportionately affected certain groups.

The court must still respect legitimate interests such as trade secrets and third-party rights.

Principle

Algorithmic opacity cannot automatically defeat an individual's ability to understand and challenge an adverse automated decision.

Importance

This is particularly useful when a taxpayer alleges:

"I cannot prove discrimination because the tax authority will not disclose how its algorithm classified me."

15. Case 8 — Österreichische Post

C-300/21, CJEU, 4 May 2023

Österreichische Post involved algorithmic profiling to predict political affinity.

Although not a tax case, it is particularly relevant to algorithmic profiling and damages.

The CJEU held that a GDPR infringement by itself does not automatically create a right to compensation. There must be:

an infringement;

damage; and

a causal link between the infringement and damage.

At the same time, EU law does not impose a minimum seriousness threshold for non-material damage. (Infocuria)

Tax relevance

Suppose a discriminatory tax algorithm:

profiles taxpayer → wrongly identifies taxpayer as high risk → causes additional scrutiny → causes financial loss and distress.

The claimant would need to establish the relevant GDPR infringement and the resulting damage and causal connection.

Principle

Algorithmic unlawfulness and compensation are related but separate questions.

16. Case 9 — Gielen / Talotta / Schumacker Combined Principle

These tax cases should be read together.

They establish an important proposition:

Tax administration has discretion, but not unlimited discretion.

The State may:

design its tax system;

determine taxable bases;

combat tax evasion;

protect fiscal supervision;

distinguish objectively different taxpayers.

But it must still respect applicable EU law.

Therefore:

Tax efficiency cannot automatically justify algorithmic discrimination.

This is especially important where an administration argues:

"The algorithm is necessary to detect tax fraud."

The correct legal question is not merely whether fraud detection is legitimate.

The questions are:

Is the distinction genuinely relevant?

Is it based on reliable data?

Is the criterion objectively justified?

Is the system proportionate?

Are less discriminatory alternatives available?

Is there human review?

Can the taxpayer challenge the outcome?

17. What Counts as Algorithmic Tax Discrimination?

A. Direct discrimination

Example:

Algorithm assigns higher audit probability to taxpayers because they possess a particular nationality.

This is the easiest form conceptually.

B. Indirect discrimination

Example:

Algorithm uses residence as a risk factor.

Residence may sometimes be legally relevant, but if the criterion disproportionately disadvantages persons protected by EU law, the authority must justify the differentiation.

Talotta + Schumacker + Gielen become relevant.

C. Proxy discrimination

Example:

postcode → socioeconomic data → ethnic composition → high-risk score.

This is particularly important under CHEZ.

D. Data-quality discrimination

Suppose historical audit data contain discriminatory enforcement patterns.

The algorithm learns:

"Previously audited group = high risk."

The algorithm then reproduces the historical bias.

This produces a feedback loop:

Historical bias → training data → algorithm → new enforcement → new data → stronger historical bias

18. Algorithmic Error Versus Discrimination

These should not be confused.

Algorithmic error

The system incorrectly identifies a taxpayer as high risk.

Algorithmic discrimination

The error or differential treatment disproportionately affects a protected group or violates an applicable equality rule.

For example:

10% false-positive rate for Group A
40% false-positive rate for Group B

This disparity does not automatically prove unlawful discrimination.

Further analysis is required concerning:

the protected characteristic;

statistical significance;

causal mechanism;

comparability;

justification;

proportionality;

accuracy;

alternative explanations.

19. Causation

Causation is often the hardest part of the civil claim.

Consider:

AI risk score
↓
audit
↓
reassessment
↓
taxpayer appeals
↓
court confirms tax liability

The taxpayer may have difficulty arguing that the algorithm itself caused the ultimate tax liability if the underlying tax debt was independently established.

Conversely:

AI risk score
↓
erroneous investigation
↓
frozen refund
↓
business cash-flow loss

creates a potentially stronger causal argument, subject to national law.

Therefore:

Algorithmic discrimination ≠ automatic damages.

20. Material Damage

Possible material losses include:

wrongly imposed tax;

unlawful penalty;

interest;

additional professional fees;

accountant costs;

legal expenses where recoverable;

business interruption;

financing costs;

loss caused by delayed refunds.

However, whether each category is recoverable depends on the applicable EU and national legal basis.

21. Non-Material Damage

Possible claims may concern:

distress;

loss of privacy;

reputational harm;

anxiety caused by unlawful profiling;

loss of control over personal data;

humiliation arising from discriminatory treatment.

Under Österreichische Post, a GDPR compensation claim requires infringement, damage, and causation, but non-material damage does not have to cross an EU-imposed minimum seriousness threshold. (Infocuria)

22. Right to Explanation

An algorithmic tax-discrimination claimant may seek information concerning:

Input data

What information was used?

Variables

Which characteristics affected the risk assessment?

Weighting

How heavily did each factor influence the score?

Output

Why was the taxpayer placed into a particular risk category?

Human intervention

Did a tax official genuinely review the result?

Error correction

Could incorrect data be corrected?

Group effects

Was the system tested for discriminatory outcomes?

The Dun & Bradstreet judgment is especially relevant to meaningful information about automated decision-making. (Infocuria)

23. Human Review

Human review is particularly important where the algorithm triggers serious consequences.

A genuine review should ideally allow the official to:

examine the underlying facts;

reject the algorithmic recommendation;

correct erroneous data;

consider the taxpayer's explanation;

investigate unusual circumstances;

document reasons.

A purely formal process:

"AI says high risk → officer clicks approve"

is much weaker as a safeguard than genuine independent review.

24. Fundamental Rights

Algorithmic tax enforcement can implicate several Charter rights, depending on the facts:

The precise Charter route depends on whether the matter falls within the scope of EU law.

25. Important Distinction: Tax Enforcement vs Tax Policy

This distinction is essential.

Tax policy

Example:

Government chooses a 30% corporate tax rate.

This is primarily a question of fiscal policy.

Algorithmic enforcement

Example:

Tax authority's AI selects certain businesses for investigation.

This raises questions of:

administrative legality;

equality;

data protection;

procedural fairness;

accuracy;

proportionality;

judicial review.

Therefore, an algorithmic enforcement claim is not necessarily an attack on the tax rate itself.

26. Evidence in an Algorithmic Tax Discrimination Case

A claimant should seek evidence such as:

Algorithmic evidence

model version;

model documentation;

risk-scoring methodology;

variables;

thresholds;

training data;

validation reports;

bias testing;

error rates.

Administrative evidence

audit-selection records;

reasons for investigation;

decision logs;

human-review records;

internal guidance;

override records.

Personal-data evidence

taxpayer profile;

data sources;

inaccurate information;

inferred characteristics;

data-sharing records.

Statistical evidence

group selection rates;

false-positive rates;

false-negative rates;

audit frequency;

penalty frequency;

reassessment rates.

Financial evidence

additional tax;

penalties;

interest;

professional expenses;

business losses.

27. Burden of Proof

In discrimination litigation, the claimant generally needs to establish sufficient facts from which discrimination may be inferred, after which the evidentiary burden may shift depending on the applicable EU/national equality regime.

This makes statistical evidence particularly important for algorithmic systems.

For example:

Group A = 5% audit rate
Group B = 25% audit rate

This is evidence of a disparity, but it is not automatically proof of unlawful discrimination.

The authority may argue:

different economic behaviour;

different transaction patterns;

different legal obligations;

different risk exposure.

The claimant can then challenge whether those explanations are genuine and proportionate.

28. Proxy Variables

Proxy discrimination is one of the biggest risks.

Suppose the algorithm does not use ethnicity.

Instead it uses:

postcode;

language;

country of birth;

migration history;

family structure;

business location.

If these variables function as substitutes for a protected characteristic, the analysis may move beyond formal neutrality.

CHEZ provides an important conceptual foundation for examining apparently neutral measures that produce group disadvantage. (Infocuria)

29. Historical Data Bias

Tax algorithms may learn from historical enforcement.

Suppose historically:

Group X was audited more frequently.

The training dataset then contains:

Group X → more audits → high risk.

The new AI learns that correlation.

It subsequently produces:

Group X → high risk → more audits.

The system has therefore transformed a historical enforcement pattern into an apparently objective algorithmic prediction.

This is a major feedback-loop problem.

30. Fiscal Supervision as a Justification

Tax authorities will commonly have legitimate objectives such as:

preventing tax evasion;

detecting fraud;

protecting public revenue;

improving compliance;

targeting limited audit resources.

These are legitimate governmental interests.

But Talotta demonstrates that fiscal-supervision objectives do not automatically justify discriminatory treatment. (Infocuria)

The proportionality inquiry therefore becomes:

LEGITIMATE AIM → SUITABILITY → NECESSITY → BALANCING

31. Proportionality

A tax authority should ideally be able to demonstrate:

1. Legitimate objective

Example: detecting VAT fraud.

2. Rational connection

The algorithm must actually assist in detecting VAT fraud.

3. Necessity

A less discriminatory method should not be reasonably available.

4. Proportionality

The burden imposed on taxpayers should not be excessive compared with the enforcement objective.

32. Civil Liability Structure

A civil claim can potentially be structured as:

Step 1 — Legal duty

Equality, GDPR, administrative-law, or another applicable legal duty.

Step 2 — Algorithmic breach

The system uses unlawful criteria, inaccurate data, inadequate safeguards, or discriminatory profiling.

Step 3 — Enforcement consequence

Audit, reassessment, penalty, collection action, refund delay, etc.

Step 4 — Damage

Financial or non-material damage.

Step 5 — Causation

The claimant establishes that the unlawful algorithmic processing/treatment caused the damage.

Step 6 — Remedy

Possible remedies depend on the legal basis and national procedure:

annulment;

reassessment;

correction of personal data;

cessation of unlawful processing;

human review;

disclosure/explanation;

compensation;

administrative remedies;

judicial review.

33. Case Law Table

The first five are primarily tax/equality or discrimination authorities; the last three are primarily algorithmic/data-protection authorities. There is currently no need to pretend that they are all cases directly concerning an AI tax-enforcement system.

34. Direct vs Analogical Authority

Relatively direct tax-discrimination authorities

Schumacker

Asscher

Talotta

Gielen

These establish principles concerning discrimination in taxation.

Direct algorithmic authorities, but outside taxation

SCHUFA

Dun & Bradstreet

Österreichische Post

These establish principles concerning automated processing, profiling, transparency and compensation.

General discrimination authority

CHEZ

This is particularly useful for indirect and proxy discrimination.

Thus, a strong European legal argument combines:

TAX EQUALITY CASES + ALGORITHMIC DECISION CASES + DISCRIMINATION CASES.

35. Hypothetical Example

Assume a Member State creates an AI system called TaxRisk-AI.

It evaluates:

income;

transactions;

residence;

postcode;

cross-border payments;

previous tax history.

The algorithm assigns:

Group A → 8% audit probability
Group B → 32% audit probability

A taxpayer from Group B is audited, reassessed and fined.

The taxpayer claims discrimination.

Legal analysis

1. Data

What information did TaxRisk-AI use?

↓

2. Algorithm

Which variables caused the higher risk score?

↓

3. Differential treatment

Why was Group B audited more frequently?

↓

4. Protected ground/proxy

Does any variable operate as a proxy for nationality, ethnic origin or another protected characteristic?

↓

5. Comparability

Are Group A and Group B objectively comparable?

↓

6. Justification

Does the tax authority have a legitimate fiscal reason?

↓

7. Proportionality

Was the discriminatory effect necessary?

↓

8. GDPR

Was automated profiling involved?

↓

9. Transparency

Can the taxpayer understand and challenge the score?

↓

10. Damage

What financial or non-material harm resulted?

↓

11. Causation

Did the algorithm cause the adverse treatment and damage?

↓

12. Remedy

What relief is available under the relevant EU and national law?

36. Most Important Legal Principle

The central principle can be expressed as:

A tax authority cannot automatically convert an unlawful discriminatory practice into a lawful one merely by placing the decision-making process inside an algorithm.

Technology changes the method of administration, not necessarily the underlying legal standards.

37. Key Challenges for Claimants

The most difficult problems are usually:

1. Black-box problem

The taxpayer may not know why the algorithm produced the result.

2. Proxy problem

The algorithm may not explicitly use a protected characteristic.

3. Causation problem

The tax authority may argue that the final decision was made independently by an official.

4. Statistical problem

A disparity does not automatically prove unlawful discrimination.

5. Fiscal-autonomy problem

Member States retain substantial competence over direct taxation.

6. Damage problem

An unlawful data-processing practice does not automatically establish compensable damage under GDPR.

7. Trade-secret problem

The administration or vendor may resist disclosure of technical information.

The Dun & Bradstreet judgment is particularly important where transparency and meaningful explanation are disputed. (Infocuria)

38. Ultra-Simple Exam Explanation

Algorithmic tax enforcement discrimination means using AI, profiling or automated risk systems in tax administration in a way that unfairly disadvantages a taxpayer or group.

European law examines:

Equal treatment

Indirect discrimination

Proxy discrimination

GDPR profiling

Automated decision-making

Transparency

Human review

Proportionality

Damage

Causation

Effective remedy

Important cases:

Schumacker → tax equality

Asscher → residence discrimination

Talotta → discriminatory tax assessment

Gielen → alternative treatment does not automatically cure discrimination

CHEZ → indirect/proxy discrimination

SCHUFA → automated scoring

Dun & Bradstreet → explanation of algorithmic decisions

Österreichische Post → GDPR damages and causation

39. Exam Formula

Main formula

TAX DATA → AI RISK SCORE → ENFORCEMENT DECISION → DIFFERENTIAL IMPACT → DISCRIMINATION → JUSTIFICATION → PROPORTIONALITY → DAMAGE → CAUSATION → REMEDY

GDPR formula

PERSONAL DATA → PROFILING → AUTOMATED DECISION → LEGAL/SIGNIFICANT EFFECT → ARTICLE 22 → SAFEGUARDS → EXPLANATION → HUMAN REVIEW → CHALLENGE

Discrimination formula

PROTECTED GROUP → NEUTRAL VARIABLE/PROXY → DISPROPORTIONATE EFFECT → OBJECTIVE JUSTIFICATION → PROPORTIONALITY → REMEDY

Final keyword bank

Algorithmic Tax Enforcement – Tax Risk Scoring – Automated Tax Audit – Tax Profiling – Fiscal Supervision – Direct Taxation – Equal Treatment – Indirect Discrimination – Proxy Discrimination – Nationality – Residence – GDPR – Article 22 – Profiling – Automated Decision-Making – Human Review – Explainability – Algorithmic Transparency – Data Accuracy – Bias – Statistical Disparity – Proportionality – Fiscal Autonomy – Causation – Material Damage – Non-Material Damage – Compensation – Judicial Review – Effective Remedy.