Civil Law and UAE Algorithmic Bias in Legal Decision-Making Systems

1. Introduction

Algorithmic bias in legal decision-making systems refers to unfair, inaccurate, discriminatory, or systematically distorted outcomes produced by an algorithm used in the administration of justice or in legal decision-making.

In the UAE, this issue is becoming increasingly important because courts, tribunals, government authorities, lawyers, experts, and dispute-resolution institutions may use digital systems for:

• case classification and prioritisation;
• legal research;
• document review;
• evidence analysis;
• translation and transcription;
• fraud detection;
• risk assessment;
• dispute-resolution assistance;
• prediction of litigation outcomes;
• automated administrative decisions; and
• AI-assisted judicial case management.

The central civil-law question is:

Who bears legal responsibility when an algorithm produces a biased or procedurally unfair legal outcome?

UAE law does not generally treat an AI system as an independent legal person. Responsibility therefore normally has to be connected to the court, authority, institution, developer, vendor, employer, operator, or other legally responsible human or juridical person.

A particularly important current-law development is that the UAE's new Federal Decree-Law No. 25 of 2025 promulgating the Civil Transactions Law entered into force on 1 June 2026, replacing the former 1985 Civil Transactions Law. Older UAE cases therefore remain useful as jurisprudential guidance, but they were generally decided under the previous Code.

2. Meaning of Algorithmic Bias

Algorithmic bias occurs where an automated or AI-assisted system produces systematically distorted results.

It may arise from:

A. Biased training data

If historical legal decisions contain institutional or demographic patterns, an AI system trained on those decisions may reproduce them.

B. Incomplete data

An algorithm may make a poor recommendation because important facts were excluded.

C. Proxy discrimination

An algorithm may not expressly use a protected characteristic but may use a substitute variable that indirectly produces discriminatory outcomes.

For example:

geographical location → socioeconomic profile → risk score

The system may therefore produce discriminatory results without explicitly asking for a person's protected characteristic.

D. Design bias

The developer may select objectives that favour efficiency, cost reduction, or risk minimisation over individual fairness.

E. Automation bias

Human decision-makers may assume that an AI recommendation is correct merely because it was produced by a sophisticated system.

F. Feedback loops

If earlier decisions affect future training data, the algorithm can continuously reinforce the same pattern.

3. Algorithmic Bias in Legal Decision-Making

Legal decision-making is particularly sensitive because it directly affects:

• property;
• contracts;
• compensation;
• employment;
• family rights;
• commercial disputes;
• administrative rights;
• access to justice;
• procedural rights; and
• personal reputation.

An algorithmic recommendation in an ordinary commercial application may cause inconvenience.

An algorithmic recommendation in a judicial or governmental system may affect a person's legal rights and remedies.

Consequently, the standard of governance should be considerably higher.

4. UAE Civil-Law Framework

Algorithmic bias should be analysed through several overlapping areas of UAE law.

4.1 Civil Transactions Law

Civil law provides the fundamental framework concerning:

• obligations;
• contractual duties;
• good faith;
• causation;
• fault;
• damage;
• compensation;
• abuse of rights;
• interpretation of legal relationships; and
• liability.

The new Civil Transactions Law is particularly relevant because technological systems do not eliminate traditional civil-law questions.

The fundamental inquiry remains:

Was there a legally relevant duty, was that duty breached, did the breach cause legally recognised harm, and what remedy follows?

4.2 Evidence Law

The UAE's Federal Decree-Law No. 35 of 2022 on Evidence in Civil and Commercial Transactions is highly relevant.

Algorithmic decisions may generate:

• electronic records;
• system logs;
• audit trails;
• model outputs;
• electronic communications;
• expert reports;
• data records; and
• other forms of electronic evidence.

A party challenging algorithmic bias may therefore need to establish:

1. what data the system used;
2. how the algorithm operated;
3. what output it produced;
4. whether the output was reproducible;
5. whether the data were complete;
6. whether the system was altered;
7. whether human intervention occurred; and
8. whether the algorithm actually caused the disputed decision.

4.3 Electronic Transactions and Trust Services

Federal Decree-Law No. 46 of 2021 provides an important legal framework for electronic transactions and trust services.

Its significance is that digital processes can have legal consequences, but electronic validity does not automatically establish substantive fairness.

An electronically generated decision may be authentic while still being legally challengeable because the underlying decision-making process was defective.

4.4 Personal Data Protection

Federal Decree-Law No. 45 of 2021 concerning Personal Data Protection is also relevant.

AI decision-making may involve extensive personal information.

Important concerns include:

• lawful processing;
• data accuracy;
• security;
• purpose limitation;
• responsible handling of personal information; and
• protection against inappropriate use of personal data.

Thus, an algorithm can create two separate legal problems:

data-protection problem + civil-liability problem.

5. Why Algorithmic Bias Creates a Civil-Law Problem

Algorithmic bias can create liability through several pathways.

5.1 Faulty design

A developer may design a system that systematically produces unreliable results.

5.2 Faulty data

Incorrect or incomplete data can lead to incorrect decisions.

5.3 Negligent deployment

A system may be technically functional but inappropriate for the particular legal purpose.

5.4 Lack of human review

A decision-maker may simply accept the AI recommendation without meaningful examination.

5.5 Failure to monitor

Bias can emerge after deployment because data, populations, or circumstances change.

5.6 Failure to explain

A completely opaque system may make it difficult for an affected person to challenge an adverse decision.

5.7 Inadequate cybersecurity

Manipulation of training data or system inputs can produce biased or false outcomes.

6. Algorithmic Bias and Due Process

A fundamental principle is that technology should not eliminate procedural fairness.

Where an AI-assisted system affects a person's rights, an adequate framework should normally provide:

• notice of the decision;
• meaningful opportunity to challenge it;
• human review;
• access to relevant reasons;
• reliable evidence;
• independent evaluation;
• correction mechanisms; and
• judicial review where legally available.

The objective should not be merely:

“The computer produced this result.”

The legally meaningful question is:

“Can the decision be justified under applicable law and supported by reliable evidence?”

7. Human-in-the-Loop Principle

The safest model for UAE legal decision-making is a human-in-the-loop system.

The AI can:

• identify patterns;
• organise documents;
• classify cases;
• identify potentially relevant authorities;
• detect inconsistencies;
• assist with translation;
• generate summaries; and
• provide recommendations.

But the final legal judgment should remain the responsibility of the legally authorised decision-maker.

This is especially important because an algorithm does not possess judicial responsibility merely because it generates an apparently rational recommendation.

8. Important UAE Case Laws

There is currently very limited reported UAE jurisprudence directly deciding algorithmic bias in judicial AI systems. Therefore, the following cases should be understood as foundational or analogous authorities concerning evidence, expert assessment, contractual obligations, causation, judicial reasoning, digital evidence, and compensation.

Case 1: Federal Supreme Court Appeal No. 538 of 2016 (Civil), 18 December 2017

The UAE Federal Supreme Court considered the relationship between interconnected contractual obligations and the requirement of good faith in performance.

Relevance to algorithmic bias

The principle is important because AI systems are normally embedded within a broader contractual and institutional relationship.

For example:

• a government authority contracts with an AI vendor;
• the vendor supplies a decision-support system;
• the authority deploys it;
• an affected person suffers harm.

The system's output cannot be considered independently from the duties governing the parties' relationship.

Principle: contractual and legal obligations must be performed consistently with applicable legal standards and good faith.

9. Federal Supreme Court Appeal No. 941 of 2019 (Commercial), 24 March 2020

This case concerned the legal characterisation of liability and the distinction between contractual and tortious responsibility.

Relevance

Algorithmic harm may involve several possible legal relationships.

For example:

Developer → contractual liability

Employer → contractual/employment liability

Operator → operational negligence

Third party → possible tort liability

Correct legal characterisation is therefore essential.

An injured party should not automatically be forced into one liability category merely because the harmful decision was produced electronically.

10. Federal Supreme Court Appeal No. 826 of 2017 (Civil), 31 December 2018

This authority is particularly useful for algorithmic decision-making because it illustrates the importance of technical expertise in resolving technically complex questions.

Relevance

Courts distinguish between:

• legal questions; and
• technical or specialised questions.

Algorithmic bias frequently requires expert analysis of:

• source data;
• model architecture;
• error rates;
• statistical patterns;
• system logs;
• validation methods;
• data integrity; and
• causal connection.

An expert may explain whether a system produced a particular pattern, but the ultimate legal conclusion remains for the court.

11. Federal Supreme Court Appeals Nos. 1012 and 1023 of 2022 (Commercial), 17 January 2023

These cases reinforce the distinction between legal questions and technical matters requiring expertise.

Relevance to AI bias

A court should not simply accept an AI vendor's technical assertion that:

“The algorithm is accurate.”

Instead, technical claims may require examination through appropriate evidence and expert analysis.

An expert can investigate:

• algorithmic performance;
• data integrity;
• system methodology;
• technical causation; and
• statistical anomalies.

The court ultimately determines their legal significance.

12. Federal Supreme Court Appeal No. 872 of 2023 (Commercial), 1 November 2023

This case is particularly useful for modern digital disputes.

The case involved technical examination of audio material and demonstrated that technical authenticity does not necessarily resolve the legal significance of the material.

Relevance to algorithmic evidence

This distinction is extremely important.

There are two separate questions:

Question 1:
Is the algorithmic record technically authentic?

Question 2:
Does the record establish the legal proposition asserted by the party?

An AI system may therefore produce a genuine, technically accurate output that nevertheless does not legally prove the conclusion for which it is being relied upon.

13. Federal Supreme Court Appeal No. 79 of 2020 (Civil), 17 February 2020

This authority concerned admissions and the evidentiary consequences of statements made by parties.

Relevance

AI-generated summaries, automated communications, or machine-generated records should not automatically be treated as equivalent to a human admission.

The court must examine:

• who generated the information;
• whether the person authorised the communication;
• whether the record was altered;
• whether the system accurately represented the underlying communication; and
• whether legal attribution can be established.

Thus:

machine generation ≠ automatic legal attribution.

14. Federal Supreme Court Appeal No. 261 of 2000 (Civil), 17 September 2000

This case involved older forms of electronic communication, including telegram/fax-type communications, and questions concerning reliability and attribution.

Importance

Although the technology predates modern AI, the underlying legal principle is highly relevant.

Courts have historically had to determine:

• whether electronic communications can be relied upon;
• whether they can be attributed to a person;
• whether they are authentic; and
• what evidentiary weight they should receive.

The same principles now apply in a technologically more sophisticated environment.

15. Federal Supreme Court Cassation No. 880 of 2021 (Civil), 15 November 2021

This is particularly important concerning damages and loss of opportunity.

The court recognised that compensation may extend to future damage and loss of opportunity where legally established.

Relevance to algorithmic bias

Suppose an AI system incorrectly:

• rejects a legal claim;
• denies access to a service;
• produces a discriminatory risk assessment; or
• causes an individual to lose a commercial opportunity.

The claimant may potentially seek compensation where the relevant elements of liability and causation are established.

This demonstrates why algorithmic bias cannot be treated merely as a technical defect.

It can become a civil damages issue.

16. Federal Supreme Court Appeal No. 950 of 2019 (Criminal), 4 February 2020

This case concerned WhatsApp-related digital evidence.

Relevance

Although it is a criminal rather than civil case, it illustrates an important broader principle:

digital evidence requires judicial evaluation rather than automatic acceptance.

This is relevant to AI because algorithmic outputs should similarly be evaluated for:

• authenticity;
• attribution;
• reliability;
• context; and
• evidentiary weight.

It should not be assumed that an electronically generated result is necessarily correct merely because it comes from a technological system.

17. Federal Supreme Court Appeal No. 1001 of 2022 (Criminal), 9 May 2023

This case involved WhatsApp communications and issues concerning attribution and factual evaluation.

Relevance to AI systems

The case supports the broader proposition that digital information must be connected to the relevant person and evaluated within its factual context.

For algorithmic decision-making, this means asking:

• Who operated the system?
• Who supplied the data?
• Who authorised the decision?
• Who relied on the output?
• Was there human intervention?

These questions become central to attribution.

18. Federal Supreme Court Appeal No. 1094 of 2022 (Criminal), 24 January 2023

This case concerned digital evidence, including WhatsApp/video material.

Relevance

It demonstrates the importance of judicial evaluation of electronically stored or transmitted information.

For AI litigation, digital evidence can include:

• model outputs;
• prompts;
• logs;
• metadata;
• database records;
• model versions;
• system instructions; and
• audit trails.

The existence of such material does not eliminate the court's responsibility to assess its reliability and relevance.

19. Case-Law Summary

20. Who May Be Liable for Algorithmic Bias?

Algorithmic bias can involve multiple actors.

20.1 AI Developer

Potential responsibility may arise where the system was:

• negligently designed;
• inadequately tested;
• trained on defective data;
• insufficiently monitored; or
• marketed for an inappropriate legal purpose.

20.2 AI Vendor

The vendor may have contractual obligations concerning:

• accuracy;
• security;
• maintenance;
• documentation;
• performance;
• updates; and
• compliance.

20.3 Government Authority or Institution

Where an authority adopts an AI system, it may have responsibilities concerning:

• lawful deployment;
• supervision;
• human review;
• procedural fairness;
• data governance; and
• correction mechanisms.

20.4 Human Decision-Maker

A judge, official, arbitrator, or administrator should not blindly adopt an AI recommendation.

Human decision-makers retain responsibility for decisions within their lawful authority.

20.5 Employer

Where an employee uses an AI system negligently in the course of employment, questions of employer responsibility may arise under applicable civil-law principles.

21. Causation Is Essential

The mere existence of algorithmic bias does not automatically establish civil liability.

A claimant generally needs to establish a legally sufficient connection between:

biased system → decision → legally recognised harm.

For example:

biased training data → incorrect risk score → denial of opportunity → financial loss.

The claimant may need evidence establishing each significant link.

This makes expert evidence particularly important.

22. Statistical Bias vs Legal Bias

These concepts should not be confused.

Statistical bias

The algorithm produces systematically different outcomes between groups.

Legal bias

The result violates a legally protected right, duty, principle, or procedural guarantee.

An algorithm may have statistical differences without every difference necessarily constituting unlawful discrimination.

Conversely, a system may create a legally problematic outcome even when the statistical disparity is difficult to demonstrate.

Therefore courts must examine both statistical evidence and legal standards.

23. Explainability

A major problem with AI systems is the “black box” problem.

If an individual receives an adverse legal outcome, the individual should be able, to the extent legally required, to understand the basis of the decision sufficiently to challenge it.

Explainability does not necessarily mean revealing the entire source code.

It may instead involve providing:

• principal factors considered;
• relevant evidence;
• decision methodology;
• applicable rules;
• human review;
• reasons for the decision; and
• a meaningful avenue for challenge.

24. Right to Human Review

For high-impact legal decisions, a strong governance model should provide meaningful human review.

The human reviewer should have authority to:

1. question the AI output;
2. examine underlying evidence;
3. reject the recommendation;
4. request additional information;
5. correct errors; and
6. provide independent reasons.

A system in which the human merely clicks “approve” is not meaningful human oversight.

25. Algorithmic Bias and Judicial Independence

Judicial independence creates an additional concern.

AI vendors should not effectively determine judicial outcomes through:

• proprietary scoring systems;
• undisclosed predictive models;
• unreviewable rankings;
• hidden variables; or
• commercial optimisation objectives.

The ultimate authority must remain with the legally constituted decision-maker.

Technology should therefore be subordinate to law, not the reverse.

26. Algorithmic Bias and Expert Evidence

Expert evidence is likely to become increasingly important.

An algorithmic-bias dispute may require experts in:

• artificial intelligence;
• statistics;
• cybersecurity;
• data science;
• software engineering;
• digital forensics;
• economics; and
• legal technology.

However, the expert should not replace the court.

The proper division is:

Expert: explains the technical evidence.

Court: determines the legal consequences.

This principle is strongly consistent with the UAE authorities concerning technical evidence and expert assessment.

27. Remedies for Algorithmic Bias

Depending upon the circumstances and applicable procedural framework, possible remedies may include:

1. Reconsideration

The decision can be reviewed using a non-biased process.

2. Correction

Incorrect data can be corrected.

3. Reassessment

The case can be reassessed without relying upon the defective algorithm.

4. Injunctive relief

Where appropriate, continued use of a harmful system may be challenged.

5. Compensation

Where legally established damage has occurred, monetary compensation may be available.

6. Loss-of-opportunity damages

Where the evidence establishes a lost legally recognisable opportunity, the UAE Supreme Court's approach to future damage and loss of opportunity is particularly relevant.

7. Contractual remedies

A defective AI system may give rise to contractual claims against the vendor or service provider.

28. Recommended UAE Governance Model

A responsible UAE legal AI system should contain at least the following safeguards:

Before deployment

• bias testing;
• data-quality testing;
• cybersecurity testing;
• legal compliance assessment;
• independent validation;
• documentation of intended use.

During operation

• human supervision;
• audit logs;
• performance monitoring;
• periodic bias testing;
• incident reporting;
• data protection controls.

After an adverse decision

• explanation;
• human review;
• correction procedure;
• preservation of relevant records;
• independent technical examination where necessary.

29. Important Legal Principle

The most important principle can be stated as follows:

Automation does not transfer legal responsibility from humans and legal entities to the algorithm.

An AI system is a technological mechanism.

The law must still identify:

who designed it → who supplied it → who deployed it → who controlled it → who relied upon it → who suffered harm.

That chain is crucial for civil liability.

30. Conclusion

Algorithmic bias in UAE legal decision-making systems represents a new technological problem but is governed largely through established civil-law principles.

The new UAE Civil Transactions Law provides the contemporary civil-law framework, while the Evidence Law, Electronic Transactions legislation, Personal Data Protection legislation, and sector-specific rules provide additional layers of regulation.

The principal legal issues are:

1. fairness of the algorithm;
2. accuracy of the underlying data;
3. human oversight;
4. explainability;
5. electronic evidence;
6. expert evidence;
7. causation;
8. attribution of responsibility;
9. procedural fairness;
10. compensation for legally recognised harm.

The UAE case law cited above does not mean that the Federal Supreme Court has already decided these cases specifically as “algorithmic bias” cases. Rather, these decisions provide established principles concerning good faith, technical evidence, electronic records, attribution, expert assessment, legal characterisation, causation, and compensation that can be applied to emerging AI disputes.

Ultimately, the appropriate UAE approach is likely to be a human-centred model of algorithmic decision-making: AI can assist legal decision-makers, but lawful authority, judicial reasoning, procedural fairness, and civil responsibility must remain attributable to legally responsible human or juridical actors.