AI Bias and Administrative Decision-Making

Artificial Intelligence is increasingly being used by public authorities to process information, assess risks, identify patterns, detect irregularities and assist administrative decision-making. While AI can improve efficiency and consistency, it can also produce biased or discriminatory outcomes.

AI bias becomes an important issue in Administrative Law when an algorithm is used to make or influence decisions concerning government benefits, licences, taxation, public employment, regulatory action, welfare schemes or other legal interests.

The central concern is that an apparently neutral technological system may produce outcomes that are arbitrary, discriminatory, inaccurate or unfair. Since administrative authorities exercise public power, AI-assisted decisions must remain consistent with constitutional and administrative-law principles.

Meaning of AI Bias

AI bias refers to a systematic tendency of an artificial intelligence system to produce outcomes that are unfair, discriminatory, inaccurate or disproportionately disadvantageous to particular individuals or groups.

Bias may arise from:

• Training data;
• Historical administrative practices;
• Incomplete datasets;
• Incorrect data;
• Algorithmic design;
• Selection of variables;
• Human assumptions;
• Proxy variables; or
• The way an AI system is deployed.

AI bias does not necessarily mean that the developers deliberately intended to discriminate.

A system may produce biased outcomes even when no discriminatory intention exists.

AI Bias in Administrative Decision-Making

Administrative decision-making involves the exercise of public power.

When AI is introduced into this process, the traditional structure may change from:

Facts → Human Officer → Legal Rules → Decision

to:

Data → Algorithm → Risk/Recommendation → Human Review → Decision

or, in some cases:

Data → Algorithm → Automated Decision

The second model creates additional legal concerns because errors or biases can become embedded in the technological system itself.

Examples of AI Bias in Administration

AI bias may arise in various administrative contexts.

Welfare Administration

An AI system may incorrectly classify individuals as ineligible for welfare benefits because of inaccurate or incomplete records.

Tax Administration

A risk-assessment system may disproportionately flag certain categories of taxpayers for investigation.

Public Employment

An automated recruitment system may rank candidates differently because its training data reflects historical patterns of selection.

Licensing

An AI system may assign different risk scores to similarly situated applicants.

Regulatory Enforcement

An algorithm may identify certain persons or businesses as high-risk because of variables that operate as indirect proxies for legally irrelevant characteristics.

Public Services

An automated system may make it more difficult for particular groups to access government services because of digital or data-related barriers.

Sources of AI Bias

AI bias can enter an administrative system at several stages.

Biased Training Data

AI systems learn from existing data.

If historical data reflects discrimination or unequal treatment, the AI system may reproduce those patterns.

For example, if historical administrative decisions consistently disadvantaged a particular group, training an AI model on those decisions may cause the system to treat the historical pattern as a basis for future predictions.

Incomplete Data

A dataset may fail to adequately represent certain groups.

If information about particular communities or circumstances is missing, an AI system may produce less accurate results for those individuals.

Incomplete data can therefore produce unequal outcomes even without intentional discrimination.

Incorrect Data

Government databases may contain:

• Incorrect identity information;
• Outdated records;
• Duplicate records;
• Incorrect addresses;
• Incorrect financial information; or
• Mistaken classifications.

If such information is used by an AI system, the resulting decision may also be incorrect.

Algorithmic Design

Bias can also arise from the design of the model itself.

The choice of:

• Variables;
• Weighting;
• Thresholds;
• Categories;
• Training methods; and
• Evaluation criteria

can influence the outcomes produced by an AI system.

A system can therefore be biased even where its underlying dataset appears neutral.

Proxy Variables

A proxy variable is information that appears neutral but may indirectly correspond to a legally sensitive characteristic.

For example, a system might use a geographical or socioeconomic variable that indirectly produces significantly different outcomes for particular communities.

The legal question is not simply whether the sensitive characteristic was directly entered into the algorithm, but whether the system produces an unlawful or unjustified distinction.

Human Bias in AI Systems

AI does not necessarily eliminate human bias.

Humans determine:

• What problem the AI system will solve;
• What data will be collected;
• Which variables will be used;
• What outcomes are desirable;
• What thresholds should be applied; and
• How the system will be evaluated.

Human assumptions can therefore become embedded within technological systems.

Feedback Loops

AI systems can create feedback loops.

Suppose an authority uses an algorithm to identify certain areas as high-risk and consequently conducts more inspections there.

The increased number of inspections produces more recorded violations.

That new data is then fed back into the system, causing the area to be classified as even higher risk.

The system may therefore reinforce an initial assumption rather than independently verifying it.

AI Bias and Article 14

Article 14 of the Constitution guarantees equality before the law and equal protection of the laws.

This makes Article 14 particularly relevant to AI bias in public administration.

An AI system may raise Article 14 concerns where it produces:

• Arbitrary classifications;
• Unjustified differential treatment;
• Discriminatory outcomes;
• Unequal application of administrative rules; or
• Decisions based on irrelevant considerations.

E.P. Royappa v. State of Tamil Nadu (1974)

The Supreme Court established an important connection between Article 14 and protection against arbitrary State action.

The principle is relevant to AI-assisted administration because algorithmic decision-making remains an exercise of State power.

Maneka Gandhi v. Union of India (1978)

The Supreme Court strengthened the relationship between Articles 14, 19 and 21 and emphasised fairness and non-arbitrariness in governmental action.

An AI system used by a public authority must therefore operate consistently with these constitutional requirements.

AI Bias and Article 21

AI-based administration may also affect Article 21 where algorithmic decisions affect:

• Personal liberty;
• Dignity;
• Privacy;
• Livelihood;
• Access to essential services; or
• Other interests protected by law.

The seriousness of the constitutional concern will depend upon the nature and consequences of the administrative decision.

AI Bias and Privacy

AI systems frequently require large amounts of personal information.

The greater the quantity and sensitivity of the information processed, the greater the potential consequences of misuse or inaccurate classification.

Justice K.S. Puttaswamy (Retd.) v. Union of India (2017)

The Supreme Court recognised privacy as a constitutionally protected right.

AI-driven administration may therefore raise privacy concerns where government authorities collect, combine, analyse or infer information about individuals.

The legal framework must consider the purpose, necessity and safeguards associated with such processing.

Data Quality and Administrative Fairness

Data quality is an important component of fair AI administration.

An administrative authority should consider whether data used by an AI system is:

• Accurate;
• Relevant;
• Sufficient;
• Current;
• Properly sourced; and
• Appropriate for the administrative purpose.

A technically sophisticated AI system cannot produce a legally reliable outcome if the underlying information is materially defective.

AI Bias and Natural Justice

Natural justice may become particularly important where an AI system contributes to an adverse administrative decision.

Its requirements depend upon the nature of the decision and the governing legal framework.

Relevant safeguards may include:

• Adequate notice;
• Opportunity to respond;
• Access to relevant adverse information;
• Impartial consideration;
• Reasons for the decision; and
• An effective mechanism for review or correction.

Right to Know the Basis of the Decision

An individual affected by an AI-assisted administrative decision may need sufficient information to understand the basis of the outcome.

For example, if a government authority denies a benefit because an AI system classifies the applicant as ineligible, the person may need to know:

• The applicable eligibility criteria;
• The relevant information used;
• Whether the information was accurate;
• The role played by the AI system; and
• How the decision can be challenged.

Without such information, meaningful administrative or judicial review may become difficult.

AI Bias and the Duty to Give Reasons

Reasons are important for identifying whether an administrative decision is lawful.

S.N. Mukherjee v. Union of India (1990)

The Supreme Court recognised the importance of recording reasons in administrative and quasi-judicial decisions, subject to recognised exceptions.

In an AI-assisted decision, the authority should not simply rely upon an unexplained algorithmic output where the law requires reasons.

A statement such as “the system identified the applicant as high-risk” may not, by itself, adequately explain the legal and factual basis of an adverse decision.

Explainability

Explainability refers to the ability to understand how an AI system reached or contributed to an outcome.

In administrative decision-making, explainability is relevant because individuals and reviewing courts may need to understand:

• What factors influenced the decision;
• Whether legally relevant factors were considered;
• Whether irrelevant information affected the outcome;
• Whether the system operated consistently; and
• Whether the decision was based on reliable information.

The level of explanation required may depend upon the nature and consequences of the decision.

AI Bias and Administrative Discretion

Administrative authorities often possess discretion within statutory limits.

AI may assist officials in exercising this discretion, but it should not automatically replace the legally required exercise of judgment.

For example, if an authority is required to consider individual circumstances, an official who mechanically follows an AI-generated recommendation in every case may fail to exercise the discretion entrusted by law.

This can raise concerns regarding fettering of administrative discretion.

AI Bias and Relevant Considerations

Administrative authorities must generally consider factors that are relevant to the statutory purpose.

An AI system may process hundreds of variables, but quantity does not equal legal relevance.

A system may therefore be technically sophisticated while still being legally defective if it relies upon irrelevant information.

Judicial review can examine whether the factors influencing the decision were legally relevant.

AI Bias and Improper Purpose

Administrative power must be exercised for the purpose for which it was granted.

AI systems may create opportunities to repurpose government data for different objectives.

For example, information collected for welfare administration should not automatically be used for an unrelated enforcement purpose without appropriate legal authority.

Use of AI for an improper purpose may therefore provide a basis for administrative-law challenge.

AI Bias and Proportionality

Where AI-assisted administration interferes with Fundamental Rights, proportionality may become relevant.

For example, an AI-based surveillance system may collect extensive personal information to identify a relatively limited category of regulatory risks.

The court may examine:

• The objective pursued;
• The connection between the AI system and that objective;
• Whether less restrictive alternatives exist where applicable;
• The extent of interference with rights; and
• The safeguards available.

Modern Dental College & Research Centre v. State of Madhya Pradesh (2016)

The Supreme Court recognised proportionality as an important constitutional method for examining restrictions on rights.

AI Bias and Judicial Review

AI-assisted administrative decisions remain subject to judicial review.

Courts may examine:

• Legality;
• Jurisdiction;
• Procedural fairness;
• Relevant and irrelevant considerations;
• Arbitrariness;
• Unreasonableness;
• Proportionality where applicable;
• Violation of Fundamental Rights; and
• Improper purpose.

The use of AI does not create a separate category of administrative power outside judicial review.

AI Bias and Wednesbury Reasonableness

The principle of Wednesbury reasonableness concerns administrative decisions that fall outside the range of reasonable decisions available to the authority.

Associated Provincial Picture Houses Ltd. v. Wednesbury Corporation (1948)

The principle provides a traditional framework for reviewing unreasonable administrative decisions.

If an authority blindly relies upon an obviously defective AI output without examining its reliability or consequences, the manner in which the decision was reached may become relevant to judicial review.

AI Bias and Mala Fides

Mala fide administrative action involves improper or bad-faith exercise of public power.

AI bias should not automatically be described as mala fide.

An AI system may produce discriminatory outcomes because of defective data or design without any dishonest intention by the officials involved.

Therefore, bias and mala fides are distinct concepts.

A finding of mala fides requires appropriate evidence and should not be inferred merely from the existence of an adverse algorithmic outcome.

AI Bias and Automation Bias

Automation bias occurs when human decision-makers place excessive reliance on computer-generated recommendations.

An official may assume that:

“The computer has calculated it, so the result must be correct.”

This can be problematic where the official is legally required to exercise independent judgment.

Human oversight should therefore involve genuine evaluation rather than merely approving an automated recommendation.

AI Bias and Human Oversight

Human oversight can reduce the risk of unfair automated outcomes.

A responsible official should, where appropriate, be capable of:

• Reviewing the AI output;
• Examining the underlying data;
• Identifying errors;
• Considering individual circumstances;
• Seeking additional information; and
• Rejecting an unreliable recommendation.

The greater the consequences of a decision, the more important meaningful review may become.

AI Bias in Welfare Administration

Welfare administration is particularly sensitive because incorrect decisions can directly affect people’s access to essential benefits.

AI may be used to:

• Identify beneficiaries;
• Detect duplicate claims;
• Assess eligibility;
• Detect suspected fraud; or
• Prioritise applications.

However, incorrect data or biased classifications may result in wrongful exclusion.

A fair system should therefore provide appropriate mechanisms for:

• Verification;
• Correction;
• Human review;
• Grievance redressal; and
• Appeal where applicable.

AI Bias in Public Employment

AI may be used in recruitment or personnel administration.

Potential applications include:

• Screening applications;
• Ranking candidates;
• Identifying qualifications;
• Predicting suitability; and
• Detecting inconsistencies.

Bias may arise if the training data reflects historical patterns of unequal selection.

Because public employment is subject to constitutional equality requirements, AI-assisted recruitment must comply with applicable legal standards.

AI Bias in Tax Administration

Tax authorities may use AI to identify transactions or taxpayers requiring further examination.

Such systems can improve enforcement efficiency.

However, a risk score should not automatically be treated as proof of wrongdoing.

An individual should not ordinarily be subjected to adverse consequences solely because an algorithm identifies them as high-risk without appropriate legal procedure and evidentiary assessment.

AI Bias in Regulatory Enforcement

Regulators may use AI to identify entities requiring inspection.

An algorithm can assist enforcement, but the final regulatory action must remain grounded in lawful authority and relevant facts.

AI-generated risk indicators should therefore be distinguished from established violations.

AI Bias and Public Procurement

Government procurement of AI systems creates additional concerns.

Authorities must consider:

• Transparency;
• Equal treatment;
• Technical requirements;
• Data security;
• Vendor accountability;
• Auditability; and
• Contractual safeguards.

Tata Cellular v. Union of India (1994)

The Supreme Court established important principles concerning judicial review of government procurement and contracts.

The decision-making process must remain lawful and fair even when the subject matter involves technically complex systems.

AI Bias and Accountability

AI can create uncertainty about who is responsible for a biased decision.

Several actors may be involved:

• Government department;
• Administrative officer;
• Software developer;
• Technology contractor;
• Data provider; and
• System administrator.

From the perspective of Administrative Law, the public authority exercising the statutory power cannot simply transfer public-law responsibility to a private technology provider.

AI Bias and Government Contractors

Government authorities may rely on private companies to develop or operate AI systems.

This can create challenges involving:

• Commercial confidentiality;
• Intellectual property;
• Source-code access;
• Data ownership;
• Security;
• Audit rights; and
• Contractual responsibility.

Where the AI system substantially affects public rights, the government may need sufficient contractual and institutional safeguards to ensure that the system remains subject to lawful oversight.

AI Bias and Right to Information

The Right to Information Act, 2005 may provide an important transparency mechanism where AI systems are used by public authorities.

Depending upon the statutory framework and applicable exemptions, relevant information may include:

• Government policies concerning AI;
• Guidelines;
• Administrative records;
• Contracts;
• System-related information; and
• Decision records.

However, the RTI framework contains statutory exemptions, and not every piece of technical or personal information is automatically disclosable.

AI Bias and Digital Exclusion

Bias can also arise indirectly through unequal access to digital systems.

Individuals may be disadvantaged because of:

• Poor internet connectivity;
• Lack of digital literacy;
• Lack of suitable devices;
• Language barriers;
• Disability; or
• Difficulty navigating complex digital procedures.

An AI-driven administrative system should therefore be assessed not only for algorithmic bias but also for practical accessibility.

AI Bias and Persons with Disabilities

Digital systems may create barriers for persons with disabilities if they are not designed to be accessible.

Administrative authorities should therefore consider whether AI-enabled public services can be meaningfully accessed by all persons who are legally entitled to use them.

AI Bias and Judicial Remedies

Where an AI-assisted administrative decision is unlawful, appropriate legal remedies may be available depending upon the circumstances.

These may include:

• Administrative reconsideration;
• Statutory appeal;
• Certiorari to quash an unlawful decision;
• Mandamus to require performance of a public duty;
• Directions to reconsider a matter;
• Correction of administrative records; and
• Constitutional remedies in appropriate cases.

The particular remedy depends upon the governing law and facts.

Principles for Preventing AI Bias in Administration

A public authority using AI should consider:

Data Quality

Data should be accurate, relevant and sufficiently representative for the intended administrative purpose.

Bias Testing

Systems should be tested for discriminatory or systematically unequal outcomes.

Transparency

The role of AI in significant administrative decisions should be appropriately disclosed.

Human Oversight

Officials should retain meaningful control over significant decisions.

Explainability

Affected individuals should receive adequate information to understand legally significant outcomes.

Record Keeping

Authorities should maintain records sufficient to permit administrative and judicial review.

Correction Mechanisms

Individuals should have appropriate avenues to correct inaccurate information.

Regular Auditing

AI systems should be periodically reviewed for performance, bias and compliance.

Legal Compliance

AI systems must operate within constitutional, statutory and regulatory limits.

Difference Between AI Bias and Traditional Administrative Bias

Traditional Administrative Bias	AI Bias
May arise from personal prejudice	May arise from data, design or deployment
Usually associated with human decision-maker	May arise from algorithmic system
Often easier to identify the decision-maker	Responsibility may be distributed
Bias may be intentional or unconscious	Bias may exist without intentional discrimination
Evidence may involve conduct or statements	Evidence may involve data, model and outcomes
Natural justice addresses some concerns	Natural justice must also account for automated processes
Judicial review examines administrative conduct	Judicial review may examine both process and technological influence

Important Cases

E.P. Royappa v. State of Tamil Nadu (1974)

Connected Article 14 with protection against arbitrary State action. The principle is important for assessing arbitrary or unequal AI-assisted administrative decisions.

Maneka Gandhi v. Union of India (1978)

Strengthened the relationship between Articles 14, 19 and 21 and emphasised fairness and non-arbitrariness in governmental action.

S.N. Mukherjee v. Union of India (1990)

Recognised the importance of recording reasons in administrative and quasi-judicial decisions, subject to recognised exceptions.

Tata Cellular v. Union of India (1994)

Established important principles governing judicial review of government contracts and procurement decisions.

L. Chandra Kumar v. Union of India (1997)

Reaffirmed the constitutional importance of judicial review under Articles 226/227 and Article 32.

State of Tamil Nadu v. P. Krishnamurthy (2006)

Set out important principles concerning judicial review of subordinate legislation.

Modern Dental College & Research Centre v. State of Madhya Pradesh (2016)

Recognised proportionality as an important constitutional method for examining restrictions on Fundamental Rights.

Justice K.S. Puttaswamy (Retd.) v. Union of India (2017)

Recognised privacy as a constitutionally protected right and established an important constitutional framework for government processing of personal information.

Justice K.S. Puttaswamy (Retd.) v. Union of India (2018)

Considered the Aadhaar framework and addressed issues involving digital identity, welfare delivery, privacy and proportionality.

Key Points for Examination

• AI bias means systematic unfairness, discrimination or distortion in AI-generated outcomes.
• AI bias may arise from data, algorithmic design, human assumptions or deployment.
• AI bias can exist even without deliberate discriminatory intention.
• Article 14 is particularly important because arbitrary or discriminatory AI-assisted State action may be constitutionally challenged.
• Article 21 becomes relevant where AI decisions affect liberty, dignity, privacy or other protected interests.
• Puttaswamy (2017) is important for privacy and government data processing.
• Natural justice remains relevant where AI contributes to an adverse administrative decision.
• Reasons are important because unexplained algorithmic outputs can make meaningful review difficult.
• AI systems should not unlawfully replace statutory administrative discretion.
• Automation bias may cause officials to rely excessively on machine-generated recommendations.
• Human oversight should be meaningful, particularly in high-impact decisions.
• Incorrect government data can produce incorrect AI outcomes.
• Algorithmic risk scores should not automatically be treated as proof of wrongdoing.
• AI bias and mala fides are distinct concepts.
• Judicial review can examine AI-assisted decisions through illegality, arbitrariness, procedural impropriety, relevant and irrelevant considerations, unreasonableness and proportionality where applicable.
• Transparency, accountability, data accuracy and effective correction mechanisms are important safeguards against AI bias.
• Technological complexity does not place AI-assisted administrative action beyond judicial review.

Conclusion

AI can improve administrative efficiency, but the use of AI does not eliminate the legal responsibility of public authorities. An AI system may process information quickly and identify patterns that would be difficult for individual officials to detect, but it can also reproduce historical discrimination, inaccurate data, inappropriate classifications and hidden assumptions.

In Administrative Law, the important question is therefore not whether an AI system is technically advanced, but whether its use results in lawful, fair, rational and constitutionally compliant administration.

AI-assisted decision-making must remain subject to Article 14, Article 21, natural justice, proportionality where applicable, the duty to give reasons, judicial review and other relevant legal safeguards.

The objective should be to ensure that AI becomes a tool for improving public administration without allowing technological systems to become a source of arbitrary, discriminatory or unaccountable exercise of public power.