1. Meaning of Algorithmic Discrimination Claims

Algorithmic discrimination claims arise when an algorithm, automated decision-making system, artificial-intelligence model, scoring mechanism, or data-driven process produces unlawfully discriminatory treatment or discriminatory effects against an individual or group.

The discrimination may occur because an algorithm:

• directly uses a protected characteristic;
• uses a proxy for a protected characteristic;
• reproduces historical discrimination contained in training data;
• produces disparate outcomes for a protected group;
• relies upon biased or incomplete data;
• uses apparently neutral variables that have discriminatory effects;
• creates unequal access to employment, credit, housing, education, insurance, healthcare or public services.

The crucial point is:

An algorithm can discriminate even when its code does not expressly say "discriminate."

Algorithmic discrimination is therefore closely connected with equality law, employment law, consumer law, data protection, administrative law, human rights and constitutional law.

2. How Algorithmic Discrimination Happens

The typical chain is:

Historical data

↓

Training data

↓

Algorithm/model

↓

Prediction or classification

↓

Automated or human-assisted decision

↓

Different treatment or outcome

↓

Legal harm

For example:

A recruitment algorithm learns from ten years of historically male-dominated hiring decisions.

The model concludes that certain characteristics associated with previous successful applicants are desirable.

Those characteristics may indirectly disadvantage women.

The employer may argue:

"The algorithm never received gender as an input."

That does not necessarily answer the discrimination question.

The relevant inquiry may be whether the system produces a prohibited discriminatory effect.

3. Direct Algorithmic Discrimination

Direct discrimination occurs when a protected characteristic itself influences the decision.

Example:

An insurance algorithm charges women a higher premium solely because the customer is female.

Potentially relevant characteristics include:

• sex;
• race or ethnic origin;
• disability;
• age;
• religion;
• nationality;
• sexual orientation;

depending on the applicable jurisdiction and legal regime.

4. Indirect Algorithmic Discrimination

Indirect discrimination is particularly important in AI.

A seemingly neutral rule may disproportionately disadvantage a protected group.

Example:

An employer's algorithm gives additional points to applicants who have worked continuously for ten years.

That criterion appears neutral.

However, if applicants with caregiving responsibilities are disproportionately excluded, the criterion could potentially produce indirect discrimination depending on the applicable law and justification.

The central question becomes:

Does a neutral algorithmic criterion disproportionately disadvantage a protected group, and can the practice be legally justified?

5. Proxy Discrimination

A proxy variable is a variable that indirectly reveals or correlates strongly with a protected characteristic.

Examples include:

• postcode → ethnic composition;
• school attended → socioeconomic or ethnic background;
• purchasing behaviour → gender;
• language → nationality/ethnicity;
• employment history → gender-related disadvantage.

Therefore:

Removing the protected characteristic from the database does not necessarily remove discrimination.

An algorithm may reconstruct it indirectly.

6. Historical Bias

AI systems learn from existing data.

If the historical data reflects discrimination, the algorithm may reproduce it.

For example:

Historical hiring

80% men / 20% women

↓

Training data

↓

AI recruitment model

↓

Model predicts male-associated characteristics as "successful"

↓

Future applicants

Women receive lower scores.

The algorithm may therefore automate historical discrimination.

7. Disparate Impact

A system may produce discriminatory outcomes without discriminatory intent.

For example:

If the difference is statistically significant, investigators may examine whether the algorithm's criteria disproportionately affect Group B.

Statistical evidence can therefore become extremely important.

8. Algorithmic Discrimination in Employment

Employment is one of the most significant areas.

AI may be used for:

• recruitment;
• CV screening;
• candidate ranking;
• interviews;
• personality testing;
• promotion;
• termination;
• performance evaluation;
• attendance monitoring;
• wage determination.

Potential claims include:

• discriminatory recruitment;
• discriminatory dismissal;
• unequal pay;
• disability discrimination;
• age discrimination;
• gender discrimination.

9. Algorithmic Discrimination in Credit

Banks may use algorithms to determine:

• creditworthiness;
• loan eligibility;
• interest rates;
• fraud risk;
• repayment probability.

An algorithm may indirectly disadvantage a protected group through variables such as:

• location;
• employment history;
• transaction history;
• educational background.

This makes credit scoring a particularly important field for algorithmic discrimination litigation.

10. Algorithmic Discrimination in Housing

Algorithms may determine:

• which consumers see advertisements;
• rental eligibility;
• mortgage risk;
• housing recommendations;
• property rankings.

Discrimination may occur if an algorithm:

• excludes particular communities;
• restricts advertising;
• produces discriminatory risk classifications;
• uses geographic proxies.

11. Algorithmic Discrimination in Insurance

Algorithms can be used for:

• premium calculation;
• risk scoring;
• claims assessment;
• fraud detection.

The legal question is not merely whether the model is statistically accurate.

A model can be statistically predictive while still creating legally prohibited discrimination.

12. Algorithmic Discrimination in Education

AI can be used for:

• admissions;
• student-risk prediction;
• grading;
• scholarship allocation;
• disciplinary decisions;
• examination monitoring.

Potential problems include:

• disability bias;
• language bias;
• socioeconomic bias;
• racial/ethnic bias;
• gender bias.

13. Algorithmic Discrimination in Public Administration

Government authorities may use automated systems for:

• welfare eligibility;
• tax enforcement;
• immigration;
• policing;
• fraud detection;
• social benefits.

The legal stakes are especially high because public decisions may affect:

• liberty;
• livelihood;
• family life;
• social security;
• residence;
• public benefits.

Administrative law and constitutional equality principles may therefore operate alongside discrimination law.

14. Leading Case Laws

1. Griggs v Duke Power Co.

401 U.S. 424 (1971), U.S. Supreme Court

Facts

The employer introduced educational and testing requirements that appeared neutral but disproportionately excluded Black workers.

Decision

The Supreme Court recognised the principle of disparate impact.

Principle

An employment practice can be legally problematic even where discriminatory intent is not established if a facially neutral requirement disproportionately disadvantages a protected group and is not sufficiently justified.

Algorithmic relevance

This is one of the most important conceptual precedents for algorithmic discrimination.

An AI hiring model can similarly produce:

neutral criterion → disproportionate group disadvantage.

The absence of discriminatory intent does not necessarily end the inquiry.

15. International Brotherhood of Teamsters v United States

431 U.S. 324 (1977)

Principle

The U.S. Supreme Court addressed systemic employment discrimination and statistical evidence.

Importance

The case demonstrates how statistical patterns can provide evidence of discrimination.

Algorithmic relevance

Algorithmic discrimination frequently requires statistical analysis.

Evidence may include:

• selection rates;
• rejection rates;
• prediction scores;
• false-positive rates;
• false-negative rates;
• outcomes across demographic groups.

The Teamsters approach is therefore highly relevant to proving systemic algorithmic discrimination.

16. McDonnell Douglas Corp. v Green

411 U.S. 792 (1973)

Principle

The case established an important burden-shifting framework for employment discrimination.

The claimant can establish an initial case, after which the employer may provide a legitimate non-discriminatory explanation.

Algorithmic relevance

Suppose:

An employer rejects an applicant using an AI system.

The employer says:

"The algorithm selected the candidate based upon qualifications."

The claimant may challenge whether the stated criterion was genuinely legitimate or merely a technological explanation for a discriminatory result.

The case illustrates how discrimination litigation can operate even where decision-making is partially automated.

17. CHEZ Razpredelenie Bulgaria AD v Komisia za zashtita ot diskriminatsia

C-83/14, CJEU

This is one of the most important European authorities for algorithmic discrimination analysis.

Facts

Electricity meters were installed at a height that made them difficult for residents to inspect in a particular neighbourhood.

Principle

A measure that appears neutral can constitute indirect discrimination where it disproportionately disadvantages a protected group.

Algorithmic significance

This principle is directly relevant to AI.

An algorithm does not need to contain an explicit discriminatory instruction.

For example:

"postcode" → algorithmic risk score

may produce discriminatory consequences even though "ethnicity" is not included.

Key lesson

Neutral code does not necessarily produce neutral outcomes.

18. HK Danmark (Ring and Skouboe Werge)

Joined Cases C-335/11 and C-337/11

Principle

The CJEU examined disability discrimination and reasonable accommodation in employment.

Algorithmic relevance

AI employment systems can disadvantage disabled workers through:

• rigid productivity targets;
• automated attendance systems;
• facial-recognition systems;
• voice analysis;
• physical-performance assessments;
• inflexible scheduling.

An algorithm may classify a disability-related characteristic as poor performance without accounting for reasonable accommodation.

Key principle

Automated systems must not be treated as inherently neutral where their criteria fail to accommodate legally protected differences.

19. Glor v Switzerland

Application No. 13444/04, ECtHR, 2009

Principle

The European Court of Human Rights addressed disability discrimination and proportionality.

Algorithmic relevance

An automated classification system may impose apparently neutral requirements that disproportionately burden people with disabilities.

The case supports the broader proposition that equality analysis must consider the actual circumstances of the individual, rather than merely the formal neutrality of a rule.

20. Guberina v Croatia

Application No. 23682/13, ECtHR, 2016

Principle

The ECtHR considered disability-related discrimination and the need to consider individual circumstances.

Algorithmic significance

AI systems frequently simplify people into categories.

For example:

"Household type = standard"

may fail to account for disability-related circumstances.

The case is therefore useful by analogy in showing why rigid classifications can create discriminatory outcomes.

21. SCHUFA Holding AG

C-634/21, CJEU

Subject

Automated credit scoring.

Principle

The CJEU examined automated scoring under GDPR provisions concerning automated individual decision-making.

Algorithmic discrimination relevance

Credit scoring systems may affect:

• loans;
• housing;
• telecommunications;
• insurance;
• financial services.

Where algorithmic scoring effectively determines access to opportunities, the legal system may scrutinise:

• automated processing;
• human involvement;
• data quality;
• decision-making consequences.

Importance

This is a major modern authority connecting algorithmic decision-making with individual legal rights.

22. Ligue des droits humains v Conseil des ministres

C-817/19, CJEU

Principle

The CJEU examined large-scale automated processing of passenger information and emphasised:

• necessity;
• proportionality;
• fundamental rights;
• safeguards.

Algorithmic discrimination relevance

Large datasets and automated classification can produce:

• false positives;
• discriminatory patterns;
• disproportionate surveillance.

The case supports careful judicial scrutiny of high-impact automated systems.

23. Bărbulescu v Romania

Application No. 61496/08, ECtHR Grand Chamber, 2017

Principle

The ECtHR considered workplace monitoring and privacy.

The Court emphasised safeguards and proportionality when employers monitor workers.

Algorithmic relevance

AI workplace monitoring may evaluate:

• productivity;
• communications;
• facial expressions;
• keystrokes;
• location;
• behavioural patterns.

The case demonstrates that technological monitoring must be assessed against fundamental rights rather than being treated as legally neutral merely because technology is used.

24. López Ribalda and Others v Spain

Applications Nos. 1874/13 and 8567/13, ECtHR Grand Chamber, 2019

Principle

The Court considered covert workplace video surveillance and proportionality.

Algorithmic relevance

Modern employers may use AI-powered surveillance rather than ordinary cameras.

For example:

Camera → facial recognition → behavioural classification → productivity score.

The case supports the principle that workplace surveillance requires a proportionality analysis.

25. Österreichische Post AG v Österreichische Datenschutzbehörde

C-300/21

Principle

The CJEU examined compensation arising from GDPR violations.

Algorithmic discrimination significance

Where discriminatory profiling also constitutes unlawful personal-data processing, data-protection remedies may supplement discrimination remedies.

Thus, a claimant may potentially have:

Equality claim

Data-protection claim

Consumer/employment claim

depending upon the circumstances.

26. Indian Constitutional Framework

India does not currently have a single comprehensive statutory doctrine specifically called algorithmic discrimination law.

Algorithmic discrimination may nevertheless be challenged through existing constitutional and statutory principles.

The most important constitutional provisions include:

Article 14

Equality before law and equal protection of laws.

Article 15

Prohibition of discrimination on specified grounds.

Article 16

Equality of opportunity in public employment.

Article 19

Relevant freedoms where algorithmic restrictions affect protected expression or occupation.

Article 21

Life and personal liberty, including privacy and dignity in appropriate circumstances.

27. E.P. Royappa v State of Tamil Nadu

(1974) 4 SCC 3

Principle

The Supreme Court linked equality with protection against arbitrariness.

Algorithmic relevance

An algorithm may be challenged where its operation is:

• arbitrary;
• irrational;
• unexplained;
• inconsistent;
• based upon irrelevant factors.

The government cannot necessarily defend an arbitrary automated decision simply by saying:

"The computer generated it."

28. Maneka Gandhi v Union of India

(1978) 1 SCC 248

Principle

State action affecting fundamental rights must satisfy standards of fairness and reasonableness.

Algorithmic relevance

Where government uses AI to make decisions affecting individuals, technological automation does not remove the requirements of lawful and fair procedure.

This can become especially important where:

• benefits are denied;
• licences are cancelled;
• travel is restricted;
• individuals are classified as risks.

29. A.K. Kraipak v Union of India

(1969) 2 SCC 262

Principle

Natural justice and protection against bias can apply to administrative decision-making.

Algorithmic relevance

An algorithm can potentially introduce:

• hidden bias;
• conflicted datasets;
• institutional bias;
• discriminatory variables.

Human decision-makers cannot necessarily avoid natural-justice requirements by delegating the analytical stage to software.

30. State of West Bengal v Anwar Ali Sarkar

AIR 1952 SC 75

Principle

Classification under equality law must satisfy constitutional requirements.

Algorithmic significance

Algorithmic systems constantly classify people into categories:

• high risk / low risk;
• suitable / unsuitable;
• eligible / ineligible;
• trustworthy / untrustworthy.

The constitutional question can therefore become:

Is the classification based upon an intelligible differentia and rationally related to a legitimate objective?

This is a foundational Indian authority for understanding algorithmic classification.

31. Elements of an Algorithmic Discrimination Claim

A claimant will normally need to establish, depending upon the applicable legal regime:

1. Protected characteristic or protected group

For example:

• sex;
• disability;
• race/ethnicity;
• age;
• religion;
• nationality.

2. Algorithmic decision

The claimant must identify the automated or algorithm-assisted system.

3. Differential treatment or disparate impact

The system treats a person/group less favourably or produces a disproportionate disadvantage.

4. Causal connection

There must be a legally sufficient connection between the algorithm and discriminatory outcome.

5. Absence of lawful justification

For indirect discrimination, the defendant may have an opportunity to establish objective justification depending on the relevant law.

6. Harm or legally actionable disadvantage

Examples include:

• rejection;
• dismissal;
• lower salary;
• denial of credit;
• higher price;
• denial of benefits;
• exclusion from education.

32. Evidence Required

Algorithmic discrimination cases frequently depend upon technical and statistical evidence.

Important evidence includes:

• source code where obtainable;
• model documentation;
• training-data information;
• feature lists;
• decision logs;
• audit reports;
• statistical outcomes;
• selection rates;
• rejection rates;
• false-positive rates;
• false-negative rates;
• demographic impact analysis;
• comparator evidence;
• internal communications;
• risk assessments;
• testing documents.

33. Statistical Evidence

Suppose an employer's AI system evaluates:

10,000 applicants

The selection rate is:

• Group A: 30%
• Group B: 10%

That disparity does not automatically prove unlawful discrimination.

But it is a powerful reason to investigate:

• why the disparity exists;
• which variables cause it;
• whether the model is necessary;
• whether less discriminatory alternatives exist;
• whether the difference is legally justified.

This is where principles associated with ** Griggs and Teamsters** become particularly useful.

34. The Problem of Explainability

A claimant may ask:

"Why was I rejected?"

The business may respond:

"The model gave you a score of 37."

That may not be a meaningful explanation.

An effective explanation may require information about:

• relevant factors;
• decision process;
• data used;
• significant errors;
• human intervention;
• review procedures.

However, legal requirements differ between jurisdictions and contexts.

35. Human Oversight

Human involvement is not automatically enough.

Consider:

AI produces discriminatory score

↓

Employee sees score

↓

Employee automatically approves it

Calling this "human decision-making" may not meaningfully solve the problem.

Effective human oversight should involve:

• ability to question the model;
• ability to override it;
• sufficient information;
• authority to correct errors;
• awareness of potential bias.

This issue is especially significant following modern European automated-decision case law.

36. Defences

A defendant may argue:

A. No protected characteristic was used

But this does not necessarily defeat indirect or proxy discrimination.

B. The algorithm is statistically accurate

Accuracy does not necessarily equal legality.

C. The disparity is accidental

The legal significance of accidental discrimination depends upon the applicable legal regime.

D. Legitimate objective

The defendant may argue that the system pursues a legitimate business or public objective.

E. Proportionality

The defendant may argue that the algorithm is necessary and proportionate.

F. Human made the final decision

The claimant may challenge whether the human review was genuine.

G. Third-party vendor supplied the system

Contracting out algorithmic design does not necessarily eliminate legal responsibility.

37. Remedies

Depending upon the jurisdiction and claim, remedies may include:

• compensation;
• reinstatement;
• correction of an automated decision;
• reconsideration;
• human review;
• injunction;
• declaration;
• deletion/correction of unlawful data;
• modification or withdrawal of the algorithm;
• regulatory penalties;
• non-discrimination orders;
• reasonable accommodation.

In public-law cases, courts may quash an unlawful decision and require the authority to reconsider the matter.

38. Corporate Algorithmic Compliance

Organisations using high-impact AI should establish:

1. Equality impact assessment

Identify potential discriminatory effects before deployment.

2. Dataset audit

Check historical bias.

3. Proxy testing

Identify variables that may reproduce protected characteristics.

4. Statistical monitoring

Compare outcomes across relevant groups.

5. Human review

Create meaningful appeal and override procedures.

6. Documentation

Record:

• model purpose;
• inputs;
• testing;
• limitations;
• validation;
• monitoring.

7. Incident response

Investigate discriminatory outcomes promptly.

8. Vendor due diligence

Contracts with AI vendors should address:

• bias testing;
• audit access;
• data quality;
• compliance;
• incident reporting;
• remediation.

39. Practical Hypothetical — Recruitment

Facts

Company A uses an AI recruitment system.

The system evaluates:

• university;
• employment gaps;
• previous salary;
• writing style;
• location;
• employment history.

After one year, statistical analysis reveals that women are selected at half the rate of men.

Possible claim

Applicants may investigate whether:

• employment gaps operate as a gender proxy;
• previous salary perpetuates historical inequality;
• the model learned from biased historical hiring;
• the system is objectively justified;
• less discriminatory alternatives exist.

Company's defence

The company argues:

"Gender was never an input."

Response

That fact alone is insufficient to eliminate the possibility of indirect discrimination.

The system may still produce discriminatory effects through correlated variables.

40. Practical Hypothetical — Credit

A bank uses AI to calculate creditworthiness.

The system assigns lower scores to customers living in particular postal areas.

Those areas have a disproportionately high concentration of a protected ethnic group.

The bank argues:

"We only use geographic risk."

A claimant could investigate whether the geographical variable functions as a proxy and whether the resulting disparity is legally justified.

This is closely connected to the reasoning concerning apparently neutral practices in ** CHEZ**.

41. Practical Hypothetical — Disability

An employer introduces an AI productivity system.

The system measures:

• typing speed;
• mouse movement;
• response time;
• continuous screen activity.

An employee with a disability receives a low productivity score.

The algorithm does not contain a "disability" variable.

Nevertheless, the system may fail to account for reasonable accommodation.

The relevant legal questions could include:

• disability discrimination;
• indirect discrimination;
• reasonable accommodation;
• proportionality;
• data protection;
• employment law.

HK Danmark, Glor and Guberina are particularly useful analogical authorities.

42. Direct vs Indirect Algorithmic Discrimination

43. Most Important Case-Law Principles

Griggs

Neutral criteria can produce unlawful disparate impact.

Teamsters

Statistics can reveal systemic discrimination.

McDonnell Douglas

Discrimination claims can use structured burden-shifting.

CHEZ

Neutral-looking practices can produce indirect discrimination.

HK Danmark

Disability equality requires consideration of reasonable accommodation.

Glor

Disability discrimination must be analysed substantively.

Guberina

Individual circumstances matter in equality analysis.

SCHUFA

Automated scoring can have significant legal consequences.

Ligue des droits humains

High-impact automated processing requires necessity, proportionality and safeguards.

E.P. Royappa

Indian equality jurisprudence is strongly concerned with arbitrariness.

Maneka Gandhi

Fairness and reasonableness constrain State action.

A.K. Kraipak

Administrative decision-making remains subject to safeguards against bias.

44. Consolidated Case-Law Table

45. Conclusion

Algorithmic discrimination claims represent the application of established equality principles to automated decision-making.

The most important insight is that discrimination does not require an algorithm to contain an explicit discriminatory instruction. Historical data, proxy variables, apparently neutral criteria and optimisation objectives can all produce discriminatory outcomes.

The leading authorities demonstrate several complementary principles:

• ** Griggs and Teamsters** show the importance of disparate-impact and statistical analysis.
• ** CHEZ** demonstrates that apparently neutral practices can create indirect discrimination.
• ** HK Danmark, Glor and Guberina** emphasise substantive equality and disability accommodation.
• ** SCHUFA** demonstrates the legal importance of automated scoring.
• ** Ligue des droits humains** reinforces proportionality and safeguards for high-impact automated processing.
• ** E.P. Royappa, Maneka Gandhi and A.K. Kraipak** provide important Indian constitutional principles concerning arbitrariness, fairness and bias.

Ultimately, an algorithm should not be treated as a legally neutral "black box." Courts and regulators can examine the data used, variables selected, statistical outcomes, discriminatory effects, justification, proportionality, human oversight and the actual consequences for affected persons. The fact that a decision was produced by software does not, by itself, immunise the employer, business, financial institution or public authority from discrimination law.