Civil Law and UAE Epistemic Boundaries of Machine Legal Reasoning

1. Introduction

Epistemic boundaries of machine legal reasoning refers to the limits on what an AI system can legitimately know, infer, establish, explain, or decide in a legal dispute.

In the UAE, this issue is becoming increasingly important because legal practice is moving toward:

• artificial intelligence;
• automated legal research;
• machine-generated pleadings;
• predictive analytics;
• digital evidence;
• automated contracts;
• blockchain;
• smart forms;
• algorithmic decision-support; and
• specialised digital-economy courts.

The important distinction is:

A machine may process legal information, but processing information is not necessarily the same thing as establishing a legally authoritative conclusion.

This distinction is particularly visible in the DIFC, where the Courts have expressly recognised both the usefulness and the risks of generative AI. The DIFC Courts' Practical Guidance Note No. 2 of 2023 requires verification of AI-generated material, transparency concerning its use, attention to confidentiality and bias, and avoidance of over-reliance on AI. It expressly states that AI should assist rather than replace the human decision-making involved in preparing evidence and submissions.

The analysis below therefore treats onshore UAE civil law as the primary framework, while using DIFC authorities where they provide the clearest judicial treatment of AI, evidence, reasoning and digital legal systems.

2. Meaning of “Epistemic Boundary”

An epistemic boundary is a boundary concerning knowledge and justification.

In legal reasoning, the court must distinguish between:

1. facts that have actually been proved;
2. facts that are merely alleged;
3. evidence from which reasonable inferences can be drawn;
4. expert opinions;
5. legal rules;
6. disputed interpretations;
7. assumptions;
8. predictions; and
9. conclusions that cannot properly be established from the available material.

An AI system may produce an answer to all of these questions, but the existence of an answer does not establish its legal validity.

For example:

AI system: “The defendant is liable.”

That statement does not establish:

• what evidence proves the defendant's conduct;
• whether the evidence is authentic;
• whether the witness is credible;
• whether the legal rule applies;
• whether causation is established;
• whether an exception applies; or
• whether the conclusion is supported by the applicable UAE law.

Thus:

computational output ≠ legal proof

and

probabilistic prediction ≠ judicial finding.

3. UAE Legal Framework

There is no single UAE statute called an “AI Civil Liability Code.”

Instead, the relevant framework is distributed across:

• the Civil Transactions Law;
• Evidence legislation;
• Electronic Transactions and Trust Services legislation;
• data-protection rules;
• sector-specific AI and technology regulation;
• procedural rules;
• judicial practice; and
• special jurisdictions such as the DIFC.

The UAE's Federal Decree-Law No. 46 of 2021 on Electronic Transactions and Trust Services recognises electronically formed contracts and automated electronic transactions. Article 11 provides that contracts may be made between automated electronic media and can be legally valid and enforceable.

This establishes an important distinction:

The law can recognise an automated process without treating the automated system itself as the ultimate legal decision-maker.

4. AI Can Execute; Courts Determine Legal Meaning

A machine can:

• retrieve cases;
• classify documents;
• detect patterns;
• compare contractual clauses;
• calculate damages;
• identify potentially relevant authorities;
• generate legal text;
• organise evidence;
• predict possible outcomes.

But legal adjudication additionally requires:

• interpretation;
• evaluation of competing evidence;
• procedural fairness;
• attribution of responsibility;
• determination of credibility;
• application of normative standards;
• explanation of the decision.

Therefore:

5. Case Law

Because reported UAE cases specifically concerning the epistemology of machine legal reasoning are still limited, the following cases are used as authorities for the underlying principles of AI reliability, evidence, reasoning, human judgment and legal accountability.

The first group includes directly relevant DIFC AI cases; the later authorities demonstrate the judicial principles that constrain automated reasoning.

Case 1 — Klesta Eshja & Hair Creators Salon LLC v Salah Masri & Others, CFI 066/2024

This is one of the most important recent UAE/DIFC authorities concerning generative AI in litigation.

The defendants' amended defences had been prepared substantially with the assistance of AI and contained false references and misleading material. The Court struck out the amended defences and ordered the relevant costs consequences.

Legal significance

The case demonstrates that:

AI-generated legal content does not acquire evidential or procedural validity merely because it appears legally sophisticated.

A lawyer or litigant remains responsible for:

• checking authorities;
• verifying propositions;
• ensuring that citations exist;
• ensuring that the factual material is accurate.

Epistemic boundary

The machine's ability to generate a plausible legal proposition does not establish:

“This proposition is legally true.”

This is the classic problem of syntactic fluency without epistemic reliability.

6. Case 2 — Stelian Gheorghe v BSA Ahmad Bin Hezeem & Associates LLP & Another, CFI 045/2025

This case involved concerns that the claimant's claim form and evidence might have been partly generated by AI.

The Court noted errors in the material and observed that errors of law have no place in witness evidence filed by lawyers. The Court ultimately dealt with the jurisdictional issue and stayed the proceedings in favour of arbitration.

Legal significance

The case demonstrates another important boundary:

AI cannot transform an advocate's legal argument into witness evidence.

A witness statement has a particular evidential function.

A machine-generated statement may contain:

• legal conclusions;
• propositions the witness never personally knew;
• reconstructed facts;
• invented context.

The legal system therefore needs to distinguish:

what the witness knows

from

what the machine believes the law to be.

7. Case 3 — Khaled Salem Musabeh Humad Al Mheiri v John Cameron, [2025] DIFC CA 008

This Court of Appeal decision is extremely important for the epistemic problem of machine reasoning.

The Court emphasised the need for adequate reasons, particularly where serious findings such as fraud are made. It stated that the evidence relied upon, factual findings and reasoning process should be identifiable so that the parties and appellate court can understand how the conclusion was reached.

Relevance to AI

This creates a powerful legal boundary for algorithmic adjudication.

A machine might output:

“Probability of fraud: 87%.”

But that is not equivalent to:

1. identifying the evidence;
2. determining which facts are proved;
3. explaining credibility;
4. applying the legal test for fraud;
5. explaining the inference;
6. giving reasons capable of appellate review.

Principle

Legal reasoning must be sufficiently explainable to support judicial accountability.

An opaque algorithm therefore cannot automatically substitute for a reasoned judicial determination.

8. Case 4 — Obie v Osric, CFI 095/2025

In this case, the DIFC Court of Appeal considered findings of fraudulent misrepresentation and whether the findings were adequately supported.

The Court emphasised that the appeal jurisdiction involved questions concerning law, miscarriage of justice and procedural fairness. It also distinguished the informal evidential structure of the Small Claims Tribunal from ordinary litigation.

Epistemic relevance

The case demonstrates that legal reasoning depends upon the procedural context.

An AI system cannot simply apply one universal reasoning model to:

• ordinary civil litigation;
• small claims;
• arbitration;
• employment disputes;
• expert evidence;
• appeals.

The evidential rules and intensity of judicial scrutiny differ.

Thus:

Legal reasoning is context-dependent.

A machine trained on a general legal corpus may fail if it does not understand the procedural environment in which a proposition operates.

9. Case 5 — Oakley v Oliver, CFI 047/2025

The DIFC Court considered an application to introduce new evidence on appeal.

The Court referred to the rules governing new evidence and recognised that an appellate court may draw factual inferences justified by the evidence. Permission for new evidence depends on specified legal criteria, including its probable importance and whether it could reasonably have been available earlier.

Relevance to machine reasoning

AI systems constantly generate or discover new information.

But:

New information is not automatically admissible evidence.

A machine may locate:

• an email;
• a database entry;
• a document;
• a digital record;
• an internet archive.

The court must still determine:

• authenticity;
• relevance;
• procedural admissibility;
• reliability;
• whether the opposing party has an opportunity to respond.

Therefore:

machine discovery ≠ evidentiary admissibility.

10. Case 6 — Union Bank of India (DIFC Branch) v Velocity Industries LLC & Others, [2020] DIFC CFI 025

The Court considered evidence given remotely from another jurisdiction.

It distinguished questions of:

• witness capacity;
• procedural admissibility;
• foreign legal restrictions; and
• the weight to be attached to evidence.

The Court explained that whether remote evidence could be received was a matter governed by the forum's procedural law, while the weight of evidence remained a separate question.

AI relevance

This distinction is extremely useful for machine-generated evidence.

Three questions must be separated:

Question 1 — Can the material be received?

Question 2 — Is it authentic/reliable?

Question 3 — What weight should the court give it?

An AI system may be capable of generating or processing evidence, but those three legal determinations remain distinct.

11. Case 7 — Gate Mena DMCC v Tabarak Investment Capital Ltd, [2024] DIFC DEC 002

This case is particularly important for the emerging Digital Economy Court.

In the retrial, the Court required expert evidence concerning cryptocurrency and specifically identified an expert issue concerning whether BTC could be characterised as “money” or “currency.”

Epistemic significance

This demonstrates that even in a highly technological dispute, the Court does not simply allow technology to determine the legal characterisation of the technology itself.

Instead:

technical evidence → expert assistance → legal evaluation → judicial conclusion

The machine may assist with the technical material, but legal classification remains a judicial function.

12. Case 8 — Techteryx Ltd v Aria Commodities DMCC & Others, [2025] DIFC DEC 001

The Digital Economy Court dealt with a dispute concerning stablecoin reserves and cryptocurrency.

The Court expressly observed that whether cryptocurrency constitutes “currency” was not yet a settled proposition of common-law adjudication and proceeded cautiously in describing the asset.

Epistemic boundary

This case demonstrates why machine reasoning cannot simply infer:

“Most databases classify Bitcoin as currency, therefore Bitcoin is legally currency.”

Legal classification may require:

• statutory interpretation;
• consideration of precedent;
• examination of the particular legal context;
• assessment of competing legal concepts.

A machine's statistical classification cannot itself create a legal category.

13. The Six Major Epistemic Boundaries

The UAE legal environment reveals at least six major boundaries.

Boundary 1 — Information vs Knowledge

AI can retrieve information.

But legal knowledge requires:

information + relevance + reliability + legal context.

For example, an AI system might retrieve an old UAE Civil Code case.

But if the case interpreted the 1985 Civil Transactions Law, the system must determine whether the proposition remains applicable after the new Civil Transactions Law effective 1 June 2026.

Merely finding the case is therefore insufficient.

14. Boundary 2 — Prediction vs Proof

Machine-learning systems operate probabilistically.

They can say:

“Cases with these characteristics historically produced outcome X.”

But civil adjudication asks:

“Has this claimant proved the elements required by law in this particular case?”

Those are different questions.

Prediction

What is likely?

Legal proof

What has been established according to the applicable evidentiary standard?

A prediction cannot automatically substitute for proof.

15. Boundary 3 — Pattern Recognition vs Legal Interpretation

AI is extremely effective at pattern recognition.

For example, it can identify:

• similar clauses;
• similar cases;
• recurring fact patterns;
• comparable damages;
• similar judicial language.

But legal interpretation can require choosing between competing principles.

Example:

Contract clause A appears similar to clauses previously upheld.

That does not automatically answer:

• Was the clause incorporated?
• Was consent valid?
• Was there duress?
• Does mandatory law override it?
• Does public policy apply?
• Has the contract subsequently been modified?

Similarity is therefore evidentially useful but legally incomplete.

16. Boundary 4 — Statistical Correlation vs Causation

AI may identify that:

companies using technology X experience fewer disputes.

But this does not prove:

technology X caused the reduction.

The same problem appears in civil liability.

An AI system might identify:

pollution source A is statistically associated with contamination B.

The court must still determine whether there is legally sufficient causation.

This is especially important for:

• environmental liability;
• medical negligence;
• product liability;
• AI liability;
• financial loss;
• professional negligence.

17. Boundary 5 — Technical Expertise vs Legal Authority

AI can be extremely useful for explaining technical matters.

For example:

“This blockchain transaction was executed using this cryptographic mechanism.”

But the legal question might be:

“What proprietary right, if any, does the claimant have in the asset?”

The first is technical.

The second is legal.

Gate Mena illustrates this distinction: technical/expert evidence regarding cryptocurrency was considered, while the Court retained responsibility for determining the legal significance of the evidence.

18. Boundary 6 — Explanation vs Justification

An AI system may produce an explanation such as:

“The claim fails because similar cases generally reject such claims.”

This may be an explanation of output, but not necessarily a legal justification.

A legally adequate judgment must identify:

1. applicable law;
2. relevant facts;
3. evidence;
4. findings;
5. reasoning;
6. conclusion.

The DIFC Court of Appeal's emphasis on adequate reasons in Al Mheiri v Cameron demonstrates why this matters.

19. Hallucination as a Legal-Epistemic Problem

One of the most serious limitations of generative AI is hallucination.

A system may produce:

• nonexistent cases;
• incorrect citations;
• invented quotations;
• incorrect statutory provisions;
• outdated legislation;
• wrong jurisdictions.

The Klesta Eshja litigation provides a concrete UAE/DIFC example of AI-assisted pleadings containing false references and misleading material.

The legal lesson is:

Plausibility cannot substitute for verification.

This is why the DIFC Courts' AI Guidance requires independent verification against statutes, case law and credible legal sources.

20. Temporal Epistemic Boundary

This is particularly important in UAE legal research.

AI systems may combine:

• old legislation;
• current legislation;
• repealed provisions;
• foreign cases;
• DIFC authorities;
• onshore UAE cases.

After 1 June 2026, the UAE Civil Transactions Law changed the legislative framework by replacing the former 1985 Civil Transactions Law.

Therefore an AI system must determine:

Which law was in force at the legally relevant time?

For example:

1985–31 May 2026

→ former Civil Transactions Law.

From 1 June 2026

→ Federal Decree-Law No. 25 of 2025 / new Civil Transactions Law.

Failure to distinguish these periods creates a temporal epistemic error.

21. Jurisdictional Epistemic Boundary

An AI system can easily combine:

• UAE onshore law;
• DIFC law;
• ADGM law;
• English common law;
• foreign civil law.

But these are not interchangeable.

For example:

DIFC Courts operate within an independent common-law jurisdiction in Dubai.

The DIFC itself describes its courts as a distinct English-language common-law jurisdiction operating within the UAE.

Consequently:

DIFC precedent ≠ automatic onshore UAE precedent.

An AI legal-reasoning system must identify:

1. court;
2. jurisdiction;
3. applicable law;
4. date;
5. precedential status.

22. AI and Evidence

The UAE's electronic-transactions framework recognises electronic records and automated transactions.

This creates an important distinction between:

Automated generation

A machine produces a document or transaction.

Evidentiary reliability

The court determines whether the document reliably establishes the relevant fact.

Legal effect

The court determines what legal consequence follows.

These are three separate questions.

23. Automated Contracts Are Not Automated Judgments

Article 11 of Federal Decree-Law No. 46 of 2021 recognises automated electronic transactions.

Therefore:

automation can participate in contracting.

But this does not mean:

automation can independently decide disputes.

An automated trading system may enter a contract.

A smart contract may execute performance.

A machine may calculate a contractual amount.

But where parties dispute:

• validity;
• fraud;
• duress;
• interpretation;
• causation;
• damages;

human judicial determination may still be required.

24. DIFC Digital Economy Court

The DIFC's Part 58 expressly permits claims concerning:

• artificial intelligence;
• blockchain;
• digital assets;
• fintech;
• databases;
• cloud systems;
• digital payments;
• robotics;
• automated systems.

It also permits technologically enabled court processes, including smart forms and AI-driven decision-tree systems for collecting information.

This demonstrates an important UAE model:

AI can be integrated into judicial administration without eliminating judicial responsibility.

The technology may assist the court.

It does not necessarily become the source of legal authority.

25. AI as an Evidential Tool

AI can legitimately assist with:

Document review

Finding relevant documents in large datasets.

Legal research

Identifying potentially relevant authorities.

Chronology

Constructing timelines.

Contract analysis

Finding inconsistent provisions.

Financial analysis

Calculating transactions and losses.

Digital evidence

Identifying patterns in blockchain or electronic records.

Translation

Assisting multilingual legal work.

But each output requires appropriate verification.

26. AI Cannot Reliably Determine Credibility by Itself

Credibility is particularly difficult.

A machine may analyse:

• speech patterns;
• facial movements;
• textual inconsistencies;
• response times.

But a legal determination of credibility is context-dependent.

A witness may:

• make a linguistic error;
• misunderstand a question;
• have translation difficulties;
• forget a minor detail;
• possess incomplete information.

Therefore:

behavioural prediction is not equivalent to judicial credibility assessment.

This is one of the strongest boundaries against fully automated adjudication.

27. AI and Expert Evidence

AI can function as a technical assistance tool.

But where the dispute requires expert knowledge, the legal system still needs to consider:

• expert qualifications;
• methodology;
• data;
• assumptions;
• competing methodologies;
• reliability;
• cross-examination.

Gate Mena's treatment of cryptocurrency evidence demonstrates this model.

The machine may assist the expert, but the expert's conclusions must remain open to legal scrutiny.

28. Bias and Training Data

The DIFC AI Guidance specifically warns of:

• inaccurate information;
• bias;
• inadequate training data;
• algorithmic limitations.

It recommends understanding the training data and algorithms underlying the AI tool before relying on it.

This creates another epistemic boundary:

A system cannot necessarily know what its training data systematically omitted.

If historical decisions contain structural biases, a predictive model trained on those decisions may reproduce them.

Therefore:

historical pattern ≠ normative correctness.

29. Legal Change and Model Obsolescence

Law changes.

AI models may therefore produce an answer based on:

• repealed legislation;
• superseded cases;
• outdated regulatory guidance.

This is especially important in the UAE because the Civil Transactions Law changed in 2026.

A legally responsible AI system must therefore incorporate:

temporal version control.

For every legal proposition, the system should ideally identify:

law → jurisdiction → date → authority → status.

30. Machine Legal Reasoning and Judicial Discretion

Many civil-law questions involve standards rather than mechanical rules.

Examples include:

• good faith;
• reasonableness;
• abuse of rights;
• proportionality;
• fairness;
• causation;
• mitigation;
• public policy;
• unconscionability.

These concepts cannot always be reduced to fixed mathematical formulas.

A machine can identify previous applications of the standard.

But the final determination may require contextual normative judgment.

31. The Human-in-the-Loop Principle

The most legally defensible UAE model is therefore:

Machine

↓
information retrieval / analysis / prediction

↓

Human lawyer / expert

↓
verification / contextualisation

↓

Court

↓
legal evaluation / procedural fairness / judgment

This is consistent with the DIFC's express warning against over-reliance on generative AI and its statement that AI should assist rather than replace integral human decision-making.

32. AI Cannot Create Legal Authority Merely by Repetition

Suppose an AI system generates the same legal proposition 1,000 times.

That does not make the proposition law.

Legal authority derives from:

• legislation;
• binding precedent where applicable;
• recognised legal principles;
• authoritative judicial interpretation;
• applicable regulations.

Therefore:

frequency of machine output is not a source of legal validity.

This is a fundamental epistemic boundary.

33. AI and the Burden of Proof

A machine prediction should not silently shift the burden of proof.

For example:

“The algorithm indicates that the defendant probably committed fraud.”

That does not necessarily establish the claimant's legal case.

The court must determine whether the legally required standard has been satisfied based upon admissible and reliable evidence.

The requirement for cogent evidence in serious allegations such as fraud, highlighted in Al Mheiri v Cameron, illustrates why a probabilistic output cannot automatically replace evidentiary reasoning.

34. AI and Procedural Fairness

Machine reasoning creates procedural questions:

• Did both parties know that AI was being used?
• Can the opposing party challenge the output?
• Can the methodology be examined?
• Can the underlying data be disclosed?
• Is the system biased?
• Can the decision be explained?
• Can the decision be appealed?

The DIFC AI Guidance specifically recommends early disclosure of AI use in proceedings so that concerns can be raised and addressed during case management.

Thus:

AI transparency becomes part of procedural fairness.

35. Confidentiality Boundary

Legal AI may process:

• privileged communications;
• trade secrets;
• personal information;
• litigation strategy;
• financial records.

The DIFC AI Guidance specifically warns practitioners about confidentiality and data-protection risks when information is supplied to generative AI systems.

Therefore the epistemic issue is not simply:

“Can AI understand the information?”

It is also:

“Was the AI legally entitled to receive the information?”

36. AI and Judicial Accountability

If a court were to rely substantially upon an AI system, a difficult question would arise:

Who is responsible for the reasoning?

Possible answers could include:

• the judge;
• the judicial institution;
• the technology provider;
• the programmer;
• the expert.

But legal accountability cannot simply disappear into the algorithm.

The reasoning requirement highlighted by Al Mheiri v Cameron suggests that the ultimate judicial decision must remain attributable to the court and capable of explanation and review.

37. Emerging UAE Model: “Assisted Intelligence,” Not “Artificial Adjudication”

The developing UAE approach can therefore be conceptualised as:

Permitted/valuable

• AI-assisted research;
• AI-assisted document review;
• automated transactions;
• digital evidence analysis;
• smart court forms;
• blockchain analytics;
• AI-assisted drafting.

Legally sensitive

• AI-generated evidence;
• AI-generated witness statements;
• predictive judicial tools;
• automated credibility assessment;
• automated liability determination.

Fundamental boundary

AI can assist the legal process, but legal authority remains grounded in law, evidence, procedure and accountable human adjudication.

38. Practical Test for UAE Machine Legal Reasoning

Before relying on an AI-generated legal conclusion, five questions should be asked:

1. Source

What statute, regulation or judgment supports the proposition?

2. Currency

Was that law in force at the relevant date?

3. Jurisdiction

Is the authority:

• UAE Federal;
• Dubai onshore;
• Abu Dhabi;
• DIFC;
• ADGM;
• foreign?

4. Evidence

What evidence supports the factual assumption?

5. Reasoning

Can the conclusion be explained through a transparent chain from:

fact → evidence → legal rule → interpretation → conclusion?

If the answer to any of these questions is no, the AI output should not automatically be treated as reliable legal reasoning.

39. Summary of the Case Authorities

40. Conclusion

The epistemic boundaries of machine legal reasoning in the UAE can be summarised through eight propositions:

1. Information is not legal knowledge.
2. Prediction is not proof.
3. Pattern recognition is not legal interpretation.
4. Correlation is not causation.
5. Technical expertise is not legal authority.
6. AI-generated text is not automatically evidence.
7. Machine output does not itself constitute legal precedent.
8. Automated assistance does not eliminate accountable human adjudication.

The UAE's developing legal-technology framework is not necessarily hostile to machine reasoning. The DIFC's Digital Economy Court, AI guidance, smart forms and technology-enabled procedures demonstrate substantial institutional acceptance of AI-assisted legal processes.

But the cases involving AI-generated false citations, the requirement for reasoned judgments, treatment of new evidence, expert evidence concerning digital assets, and the distinction between technical and legal questions establish a crucial principle:

The legal system may use machines to expand the capacity of legal reasoning, but the machine's computational output does not by itself establish truth, proof, legal authority or judicial responsibility.

That is the central epistemic boundary of machine legal reasoning under the developing UAE civil-law and DIFC legal-technology framework.