1. Introduction

Artificial Intelligence (AI) model drift is an emerging issue in Singaporean civil law, particularly in financial services, banking, insurance, investment management, automated credit assessment and algorithmic trading.

AI model drift occurs when an AI system's predictive performance or decision-making behaviour changes over time because of changes in input data, economic conditions, consumer behaviour, market conditions or the relationship between variables used during training.

When an AI model becomes unreliable and causes financial losses, affected individuals, businesses and investors may seek civil remedies against the organisations responsible for developing, deploying, maintaining or supervising the system.

Examples include:

An AI trading system generating increasingly inaccurate investment signals.

A bank's credit-scoring model incorrectly rejecting creditworthy customers.

An insurance pricing model producing discriminatory or inaccurate premiums.

An AI fraud-detection system incorrectly blocking legitimate transactions.

An automated investment system generating substantial portfolio losses.

An AI lending system incorrectly assessing borrowers' repayment capacity.

A financial institution failing to detect deterioration in an AI model's performance.

Singapore's legal framework does not presently establish a separate cause of action specifically called AI model drift liability. Instead, claims would generally be assessed under established principles of contract, negligence, misrepresentation, statutory duties, data protection and financial regulation.

Important case-law qualification: There is no established body of reported Singapore Supreme Court decisions specifically determining liability for financial losses caused by AI model drift. The cases discussed below are relevant Singapore authorities on analogous legal principles. They should not be mistaken for judgments directly involving AI model drift.

2. Meaning and Nature of AI Model Drift

AI model drift refers to a deterioration or alteration in an AI system's performance after deployment.

It can occur even where the original model was properly developed and tested.

A. Types of model drift

1. Data drift

Data drift occurs when the statistical distribution of incoming data changes compared with the data used during model training.

Example: A bank trains its credit-risk model using historical economic conditions. A sudden economic downturn changes borrowers' financial circumstances, making previous predictions unreliable.

2. Concept drift

Concept drift occurs when the relationship between input variables and the predicted outcome changes.

Example: An investment model previously predicts market movements based on interest rates, but changes in monetary policy and investor behaviour make those relationships unreliable.

3. Prediction drift

Prediction drift occurs when a model's outputs change significantly over time, potentially affecting financial decisions.

4. Performance drift

Performance drift occurs when measurable accuracy, reliability or predictive effectiveness deteriorates.

5. Feedback-loop drift

This happens when the system's own decisions influence future data, creating a cycle that reinforces earlier errors.

Example: An automated lending system repeatedly rejects a particular category of borrowers. Because fewer members of that category receive loans, the model receives less information about their actual repayment performance.

Data Drift vs. Concept Drift and Why Monitoring for Them is Important | WhyLabs

Monitor ML Models | Deepchecks

AI in Insurance: How Machine Learning Prices Risk — Edu AI Blog

10

3. Legal Framework Governing AI Model Drift in Singapore

Several areas of Singapore law may become relevant.

A. Contract law

Contract law is particularly important where financial institutions purchase AI systems from technology providers or engage vendors to operate automated decision-making services.

Potential contractual obligations include:

Maintaining agreed model-performance standards.

Providing accurate representations about model capabilities.

Conducting regular testing and validation.

Monitoring changes in model performance.

Maintaining system availability.

Providing adequate warnings about limitations.

Complying with agreed risk-management procedures.

A breach may arise where a provider fails to perform contractual obligations and that failure causes foreseeable financial loss.

Relevant legislation includes the Civil Law Act 1909 and the Contract Law principles developed through Singapore case law.

B. Tortious negligence

Negligence may arise when an organisation fails to exercise reasonable care in developing, validating, monitoring or deploying an AI model.

A claimant generally needs to establish:

A duty of care.

Breach of the applicable standard of care.

Factual causation.

Legal causation and remoteness.

Recoverable damage.

For example, a financial technology company may face a negligence claim if it fails to monitor a known deterioration in its credit-risk model and that failure causes financial damage to a client.

However, a poor prediction or unsuccessful investment does not automatically establish negligence.

C. Misrepresentation

A financial institution or AI vendor may incur liability if it makes false representations concerning:

Model accuracy.

Historical performance.

Predictive reliability.

Risk-management capabilities.

Testing and validation.

Ability to detect model drift.

The Misrepresentation Act 1967 may become relevant to contractual claims involving actionable misrepresentations.

D. Financial regulatory obligations

The Monetary Authority of Singapore (MAS) supervises financial institutions operating in Singapore.

Its regulatory framework is relevant to AI deployment, including:

Technology risk management.

Outsourcing arrangements.

Operational resilience.

Model risk management.

Governance and accountability.

Fairness and responsible AI practices.

The MAS FEAT principles address fairness, ethics, accountability and transparency in the use of AI and data analytics in financial services.

These regulatory principles may help establish the expected standard of responsible institutional governance, although breach of a guideline does not automatically establish a private civil cause of action.

E. Personal Data Protection Act 2012

The PDPA may become relevant where AI model drift involves inaccurate, improperly used or inadequately protected personal data.

Potential issues include:

Inaccurate customer information.

Unauthorised use of personal data.

Improper disclosure.

Failure to protect personal information.

Inappropriate automated profiling involving personal data.

The PDPA does not create an automatic right to compensation for every inaccurate AI decision. The applicable statutory requirements and available remedies must be examined separately.

4. Civil Liability for Financial Loss Caused by AI Model Drift

The central question is whether an organisation can be held legally responsible when an AI model deteriorates and causes financial loss.

The answer depends on the relationship between the parties, the contractual obligations, the nature of the model's use and the organisation's knowledge of the risk.

A. Liability of AI developers

An AI developer may face liability where it:

Supplies a model with known technical defects.

Misrepresents the model's predictive accuracy.

Fails to disclose significant model limitations.

Neglects agreed testing and validation requirements.

Fails to correct identified defects.

Provides inadequate warnings about foreseeable financial risks.

For example, a technology company supplies a predictive trading model to a Singapore investment firm. The developer knows that the model's accuracy has deteriorated following a change in market conditions but fails to disclose this information.

If the investment firm continues using the system and suffers financial losses, contractual or negligence claims may arise depending on the applicable duty and contractual allocation of risk.

B. Liability of financial institutions

Banks, insurers and investment management companies may be responsible for their own failures in supervising AI systems.

Potential grounds include:

Failure to establish adequate model governance.

Inadequate validation before deployment.

Failure to monitor model performance.

Ignoring material performance warnings.

Inadequate human supervision.

Failure to maintain appropriate operational controls.

Failure to investigate unusual automated decisions.

A financial institution cannot necessarily avoid responsibility simply by arguing that the loss was generated by an algorithm.

C. Liability of AI vendors and third-party service providers

Financial institutions increasingly outsource AI development, cloud computing, data processing and algorithmic monitoring.

A vendor may face contractual liability where it fails to meet agreed service levels or technical obligations.

However, the financial institution's responsibility towards its own customers must be assessed separately. Outsourcing does not automatically transfer all legal responsibility to the technology provider.

D. Liability for autonomous financial decisions

Particular difficulties arise when AI systems independently:

Execute investment transactions.

Approve or reject loans.

Adjust insurance premiums.

Assess financial risk.

Allocate investment portfolios.

Detect suspicious transactions.

Recommend financial products.

The more significant the financial consequences, the more important appropriate testing, supervision and accountability become.

5. At Least 8 Important Singapore Case Laws

The following cases provide judicial principles that can be applied by analogy to AI model drift financial-loss litigation.

These are actual Singapore decisions, but none should be described as a direct judicial determination of AI model drift liability.

Case 1. Spandeck Engineering (S) Pte Ltd v Defence Science & Technology Agency

Citation: [2007] 4 SLR(R) 100; [2007] SGCA 37.

Legal principle: Duty of care and pure economic loss.

This landmark decision established Singapore's general framework for determining whether a defendant owes a duty of care in negligence.

The Court of Appeal adopted a unified test involving:

Factual foreseeability of damage.

Legal proximity between the parties.

Policy considerations that may negate the imposition of a duty.

The decision is particularly important because it recognised that pure economic loss is not automatically excluded from recovery in negligence.

Application to AI model drift:

Suppose a financial institution engages an AI developer to provide a model for investment-risk assessment. The developer knows that the model will be used to make financial decisions affecting identified clients.

If the developer negligently fails to disclose a known deterioration in the model, the claimant may argue that the developer owed a duty of care.

However, the claimant must establish sufficient proximity and overcome any relevant policy considerations.

eLitigation+1

Case 2. Go Dante Yap v Bank Austria Creditanstalt AG

Citation: [2011] 4 SLR 559; [2011] SGCA 39.

Legal principle: Investment management and assumption of responsibility.

This case involved losses suffered by an investor following the Asian financial crisis.

The Court of Appeal considered the bank's obligations under a discretionary investment management agreement.

The contractual arrangement was important because the bank had been given discretion to manage investments on the customer's behalf.

The case illustrates how the nature of an investment-management relationship can establish legal proximity and support a duty of care.

Application to AI model drift:

Where a bank uses an AI model to make discretionary investment decisions for a customer, the bank's responsibility may be more substantial than where it merely supplies general information.

A financial institution cannot automatically escape its obligations by delegating investment decisions to an AI model.

eLitigation+2

Case 3. Deutsche Bank AG v Chang Tse Wen

Citation: [2013] 1 SLR 1310; [2013] SGCA 49.

Legal principle: Negligence in financial advisory relationships.

The case involved substantial investment losses arising from financial products purchased through Deutsche Bank.

The trial judge had awarded the investor damages for negligence.

However, the Court of Appeal rejected the finding that the bank owed the investor the particular advisory duty identified by the trial judge.

The Court stressed the importance of examining the actual relationship between the bank and investor, including the contractual documents, assumption of responsibility and circumstances of the advice.

Application to AI model drift:

A customer cannot automatically establish negligence simply because a financial institution used an inaccurate AI model.

The court would need to examine:

Whether the institution undertook to provide advisory or discretionary management services.

Whether the claimant reasonably relied on those services.

Whether the contractual documents limited the institution's responsibilities.

Whether the alleged loss fell within the scope of any duty owed.

This decision is particularly useful for determining the boundary between AI-generated information and AI-assisted investment management.

eLitigation+1

Case 4. IPP Financial Advisers Pte Ltd v Saimee bin Jumaat

Citation: [2020] SGCA 47.

Legal principle: Negligent misrepresentation and financial investment losses.

This decision concerned an investor who relied on representations made by financial advisers about the risks of an investment in an algorithmic foreign-exchange trading platform.

The investor claimed that the advisers had misrepresented the investment risks.

The trial judge found the advisers liable in negligence and imposed vicarious liability on their employer.

The Court of Appeal examined when actionable damage from negligent misrepresentation occurs and how limitation periods apply.

The judgment also considered the distinction between entering into a flawed investment transaction and suffering actual damage when the relevant risks materialise.

Application to AI model drift:

This is especially relevant because the underlying investment involved an algorithmic trading platform.

Consider a financial adviser who represents that an AI trading model has been thoroughly tested and is capable of generating reliable returns.

If those statements are negligently made and the investor relies on them, a claim for negligent misrepresentation may arise.

The claimant must establish the relevant elements, including actionable representation, duty, reliance and damage.

The case also highlights the importance of identifying when the financial loss actually occurred for limitation purposes.

eLitigation+1

Case 5. NTUC Foodfare Co-operative Ltd v SIA Engineering Company Ltd

Citation: [2018] SGCA 41.

Legal principle: Causation and the scope of negligence liability.

The Court of Appeal considered the requirements of negligence, including factual causation and the relationship between the defendant's conduct and the claimant's damage.

The case provides guidance on the need to establish a legally sufficient connection between the alleged breach and the loss suffered.

Application to AI model drift:

A claimant must establish more than the fact that a financial loss occurred after an AI system was deployed.

For example, a trading firm may lose money during a market crash while its AI model is also experiencing performance deterioration.

The claimant must demonstrate that the model's deterioration and the defendant's alleged failure caused the relevant loss, rather than merely coinciding with it.

eLitigation+1

Case 6. Anwar Patrick Adrian and another v Ng Chong & Hue LLC and another

Citation: [2014] 3 SLR 761; [2014] SGCA 34.

Legal principle: Knowledge, control, vulnerability and proximity.

The Court of Appeal examined the circumstances in which a professional may owe a duty of care to persons who are not directly their clients.

The judgment recognised that the defendant's knowledge of the claimant's circumstances and the defendant's control over the relevant risk may be important proximity factors.

Application to AI model drift:

An AI vendor may have exclusive access to technical information concerning:

Model deterioration.

Changes in training data.

Predictive failures.

Known system defects.

Performance warnings.

If the vendor knows that an identified financial institution or group of clients is exposed to a particular risk, this knowledge may be relevant to the question of proximity.

The decision does not establish automatic liability for vendors, but provides a useful framework for examining responsibility in complex AI supply arrangements.

eLitigation+1

Case 7. Skandinaviska Enskilda Banken AB (Publ), Singapore Branch v Asia Pacific Breweries (Singapore) Pte Ltd

Citation: [2011] 3 SLR 540; [2011] SGCA 22.

Legal principle: Banking due diligence and internal controls.

The case concerned substantial banking losses arising from a prolonged fraud involving a finance manager.

The Court of Appeal examined the banks' failure to exercise proper due diligence and comply effectively with their internal verification procedures.

The decision emphasised that paper compliance with internal procedures is not necessarily genuine compliance.

Application to AI model drift:

A financial institution that has adopted an AI-based credit or investment system should not rely merely on the existence of written model-governance policies.

It should be able to demonstrate that relevant controls are actually implemented.

This includes:

Regular model validation.

Review of unusual outputs.

Investigation of performance deterioration.

Verification of important financial decisions.

Appropriate escalation procedures.

eLitigation+1

Case 8. Poh Chiak Ow v United Overseas Bank Ltd

Citation: [2020] SGHC 275.

Legal principle: Bank responsibility, investment loss and proof of negligence.

The plaintiff invested US$500,000 in PixelTrade and alleged that his relationship manager had made fraudulent representations.

He also claimed that the bank was negligent in failing to protect him against the investment loss.

The High Court dismissed the claim.

Application to AI model drift:

The decision illustrates that a financial institution is not automatically liable for every investment loss suffered by a customer.

Similarly, an institution's use of an AI system does not automatically create a duty to protect customers against all losses arising from third-party investments.

The claimant must establish the relevant legal duty, breach and causation.

laws.sg

6. Comparative Analysis of the Case Laws

7. Monetary Authority of Singapore and AI Risk Governance

Monetary Authority of Singapore (MAS) plays an important role in regulating financial institutions that use AI technologies.

2026 regulatory development

On 7 October 2026, MAS issued its Guidelines on Artificial Intelligence Risk Management for Financial Institutions.

These guidelines apply to financial institutions and all forms of AI technologies. They are scheduled to take effect on 7 October 2027.

Monetary Authority of Singapore+1

A. AI risk governance

Financial institutions are expected to establish appropriate governance structures and management oversight.

Important requirements include:

Board and senior-management oversight.

Clear institutional responsibility for AI risks.

AI inventories.

Risk materiality assessments.

Documented governance procedures.

B. AI lifecycle management

AI risks should be addressed throughout the system lifecycle.

This includes:

Development.

Training.

Testing.

Validation.

Deployment.

Monitoring.

Updating.

Retirement.

C. Continuous monitoring

For model drift litigation, monitoring is particularly important.

Financial institutions should have appropriate mechanisms to identify deterioration in model performance and respond to emerging risks.

D. Third-party risk management

Institutions should assess risks associated with external AI vendors, technology suppliers and outsourced service providers.

E. Human oversight

Automated financial decisions should be subject to appropriate human oversight according to their risk and materiality.

Legal significance: These guidelines may provide important evidence of regulatory expectations and reasonable institutional practices. However, they do not automatically establish civil liability or create a standalone compensation right for customers.

8. Evidence Required in AI Model Drift Litigation

Technical evidence is essential when establishing that model drift caused financial losses.

A. Technical evidence

A claimant may require:

Original model architecture.

Training-data documentation.

Validation reports.

Model-performance records.

Drift-detection reports.

Algorithmic decision logs.

Version histories.

System update records.

Monitoring alerts.

Incident reports.

B. Financial evidence

Financial evidence may include:

Investment transaction records.

Bank statements.

Portfolio valuation reports.

Trading histories.

Credit-loss calculations.

Insurance payment records.

Business financial statements.

Independent financial expert reports.

C. Expert evidence

Expert witnesses may include:

Machine-learning engineers.

AI model-risk specialists.

Financial analysts.

Economists.

Forensic accountants.

Cybersecurity specialists.

Their evidence may help establish whether the model deteriorated, when the deterioration occurred and how it contributed to the claimed financial loss.

9. Legal Remedies Available to Victims

Depending on the cause of action and contractual relationship, claimants may seek several remedies.

A. Compensatory damages

A claimant may seek compensation for losses directly attributable to the defendant's actionable conduct.

For example, if negligent model maintenance causes a financial institution to execute incorrect transactions, damages may be calculated by examining the transactions and the financial position that would reasonably have existed without the breach.

B. Rescission

Where a financial transaction was induced by actionable misrepresentation, rescission may be available subject to applicable legal requirements and bars to rescission.

C. Injunctive relief

A court may grant appropriate injunctive relief where the legal requirements are satisfied.

For example, an affected business may seek to restrain continued use of an unlawfully operated system in appropriate circumstances.

10. Defences Available to AI Developers and Financial Institutions

Defendants may rely on several legal arguments.

1. No duty of care

The defendant may argue that there was insufficient legal proximity to establish a duty under the Spandeck framework.

2. No breach

The defendant may demonstrate that reasonable testing, validation and monitoring procedures were followed.

3. Market volatility

Financial losses may have resulted from independent market movements rather than AI model drift.

4. Contractual exclusions

A defendant may rely on contractual limitations or disclaimers, subject to applicable statutory and common-law restrictions.

5. Contributory negligence

A claimant's own failure to exercise reasonable care may reduce damages where the law permits.

6. Lack of causation

The defendant may argue that the claimant cannot establish that model drift caused the claimed financial losses.

7. Limitation periods

A claim may be defeated if commenced outside the applicable limitation period, subject to statutory exceptions.

11. Limitation Periods for AI Model Drift Claims

Limitation is particularly important because model drift may remain undetected for months or years.

Under the Limitation Act 1959:

Contract claims are generally subject to a six-year limitation period.

Tort claims involving negligence are generally subject to a six-year limitation period from accrual.

Section 24A provides additional rules for negligence actions involving latent damage and the date of knowledge, subject to its statutory requirements and longstop provisions.

In AI litigation, important dates may include:

The date the model first deteriorated.

The date the institution became aware of the deterioration.

The date an actionable breach occurred.

The date actual financial damage was suffered.

The date the claimant discovered relevant facts.

The distinction between the date of a negligent representation and the date actual damage occurs is particularly important under IPP Financial Advisers Pte Ltd v Saimee bin Jumaat.

12. Hypothetical Singapore AI Model Drift Litigation

Consider the following example.

Hypothetical case: Singapore AI Investment Management Dispute

Illustrative facts only

A Singapore investment management company purchases an AI-based trading system from a technology provider.

Initial deployment

The AI model performs satisfactorily during initial testing.

Model deterioration

Changes in market conditions cause the model's predictions to become increasingly inaccurate.

Failure to monitor

The provider fails to communicate significant deterioration despite receiving internal performance warnings.

Financial losses

The investment company continues relying on the model and suffers substantial losses.

Civil proceedings

The investment company brings claims against the technology provider for breach of contract and negligence.

Judicial examination

The court considers contractual responsibilities, duty of care, technical evidence, causation and recoverable damages.

Potential legal issues

Whether the developer breached its contractual obligations.

Whether the developer owed a duty of care.

Whether the model's deterioration was reasonably foreseeable.

Whether the provider knew about the deterioration.

Whether the failure to monitor caused the losses.

Whether market movements independently contributed to the losses.

13. Future Development of Singapore Civil Law

AI model drift is likely to create new legal questions in Singapore's financial services industry.

Important future developments may include:

Potential legislative and judicial reforms

Singapore could further develop its legal framework through:

More specific AI model governance requirements.

Clearer contractual allocation of AI-related risks.

Standardised model validation and monitoring documentation.

Improved access to technical evidence in litigation.

More developed standards for AI-related professional negligence.

Clearer allocation of responsibility between AI developers, financial institutions and outsourced providers.

Appropriate insurance arrangements for AI-related financial losses.

These are possible areas of reform rather than existing standalone statutory rights.

14. Overall Conclusion

Civil liability for AI model drift financial losses in Singapore must currently be examined through established principles of contract law, negligence, misrepresentation, causation and financial regulation.

The most important judicial authority is Spandeck Engineering v Defence Science & Technology Agency, which provides the framework for establishing a duty of care, including claims for pure economic loss.

Other decisions, particularly Go Dante Yap, Deutsche Bank v Chang Tse Wen and IPP Financial Advisers v Saimee bin Jumaat, provide useful principles for investment management, financial advisory responsibilities and negligent misrepresentation.

The October 2026 MAS AI Risk Management Guidelines also represent an important regulatory development for financial institutions, although their commencement is scheduled for October 2027.

Ultimately, the central legal challenge is determining whether an identifiable party failed to exercise reasonable care or fulfil its contractual obligations, whether that failure caused the AI model's harmful financial consequences, and whether the resulting loss is legally recoverable.

For advanced legal research, the most important authorities are:

Spandeck Engineering (2007).

Go Dante Yap (2011).

Skandinaviska Enskilda Banken (2011).

Anwar Patrick Adrian (2014).

Deutsche Bank v Chang Tse Wen (2013).

NTUC Foodfare (2018).

IPP Financial Advisers (2020).

Poh Chiak Ow (2020).

These decisions provide a foundation for analysing future AI model drift financial-loss litigation