Civil Law and AI Financial Market Surveillance Liability in Europe

1. Introduction

AI financial market surveillance liability concerns civil liability arising from the use, failure, misuse, or defective operation of artificial-intelligence systems used to monitor financial markets.

Financial institutions, trading venues, investment firms and regulators increasingly use automated systems to detect:

insider dealing;

market manipulation;

spoofing and layering;

wash trading;

unusual trading patterns;

suspicious orders;

abnormal price movements;

algorithmic trading risks;

sanctions and compliance violations;

fraud;

potentially abusive trading strategies.

An AI surveillance system can create several different types of civil dispute:

False positive: legitimate trading is incorrectly identified as suspicious.

False negative: actual market abuse is not detected.

Incorrect risk classification: a trader or transaction receives an inaccurate risk score.

Algorithmic trading error: surveillance or execution systems interact incorrectly.

Data error: incomplete or incorrect market data produces an incorrect conclusion.

Model bias: the system systematically treats certain transactions or traders differently.

Cybersecurity failure: manipulation of surveillance data causes incorrect alerts.

Failure to report: an institution fails to report suspected market abuse.

Improper disclosure: confidential surveillance information is disclosed.

Regulatory reliance: a public authority relies on defective automated surveillance.

The European legal framework combines MiFID II, MiFIR, MAR, EMIR, DORA, GDPR, the AI Act and national civil/contract/tort law.

2. Basic Liability Structure

The basic model is:

AI Surveillance System → Data/Input → Algorithmic Analysis → Alert/Decision → Human or Regulatory Action → Financial Consequence → Damage

The central legal question is:

Who should bear responsibility when an AI surveillance system produces an incorrect result or fails to detect market abuse?

That question cannot be answered simply by saying that "the AI made a mistake."

The court normally needs to examine:

who operated the system;

who controlled it;

what contractual duties existed;

whether the system complied with applicable regulation;

whether the system was defective;

whether a human decision-maker relied upon the AI;

whether the resulting loss was foreseeable;

whether causation can be established.

3. European Regulatory Framework

A. Market Abuse Regulation — MAR

Regulation (EU) No 596/2014 is central to market surveillance.

MAR addresses:

insider dealing;

unlawful disclosure of inside information;

market manipulation;

detection and prevention of market abuse;

suspicious transaction/order reporting.

Market operators and investment firms therefore have significant surveillance obligations.

For AI systems, the important issue is:

Was the AI surveillance mechanism reasonably capable of detecting the relevant abusive behaviour?

A failure of an automated system does not automatically establish civil liability, but it can become important evidence concerning compliance with regulatory and contractual obligations.

4. MiFID II

Directive 2014/65/EU contains extensive requirements relating to:

investment firms;

algorithmic trading;

organisational controls;

record keeping;

systems and controls;

client protection;

market integrity.

Algorithmic trading firms must maintain appropriate systems and controls.

AI surveillance may therefore become part of an investment firm's broader compliance architecture.

5. MiFIR

Regulation (EU) No 600/2014 provides additional rules concerning financial markets, transparency and trading information.

AI surveillance systems may process:

transaction data;

order-book information;

trading venue data;

client/order information.

Errors in such data can propagate through surveillance models.

6. DORA

The Digital Operational Resilience Act (DORA) is particularly important for AI financial-market systems.

It addresses ICT risks in the financial sector, including:

ICT risk management;

incident management;

operational resilience;

testing;

third-party ICT providers;

ICT-related contractual arrangements.

An AI surveillance platform dependent upon cloud infrastructure or an external technology provider may therefore create a chain of contractual and regulatory responsibilities.

7. AI Act

Financial-market surveillance systems must also be examined under the EU AI Act where the relevant system falls within its scope.

The AI Act establishes requirements relating to:

risk management;

data governance;

technical documentation;

record keeping;

transparency;

human oversight;

accuracy;

robustness;

cybersecurity.

However:

An AI Act infringement does not automatically establish a private damages claim.

Civil liability still requires an applicable private-law basis and, where necessary, proof of damage and causation.

8. GDPR

AI surveillance can involve personal data.

Examples include:

trader identities;

employee information;

client information;

behavioural profiles;

trading histories;

communications;

transaction patterns.

Where personal data are processed, GDPR principles may become relevant.

Important provisions include:

lawful processing;

transparency;

purpose limitation;

data minimisation;

accuracy;

automated decision-making;

security;

rights of access and objection.

9. Important Case Law

Because AI-specific financial-surveillance judgments remain limited, many of the most useful authorities concern algorithmic decision-making, market manipulation, investor protection, financial information, causation and damages.

They should therefore be identified as direct or analogical authorities, rather than described as cases that actually decided liability for modern AI surveillance.

10. Case 1 — Spector Photo Group NV v Commissie voor het Bank-, Financie- en Assurantiewezen

Case C-45/08

CJEU, 23 December 2009

This is a foundational European market-abuse case.

The Court examined the concept of insider dealing under the EU market-abuse framework and the significance of possession of inside information when trading.

Importance for AI surveillance

An AI surveillance system may detect trading patterns suggesting insider dealing.

But an algorithmic alert is not necessarily equivalent to legal proof of insider dealing.

The legal analysis must distinguish:

AI suspicion → investigation → evidence → legal finding

The system may generate an alert, but the ultimate legal classification requires application of the substantive market-abuse rules.

Principle

Algorithmic identification of suspicious trading must be distinguished from the legal determination that market abuse actually occurred.

This distinction is essential for avoiding false-positive liability.

11. Case 2 — Geltl v Daimler AG

Case C-19/11

CJEU, 28 June 2012

The Court interpreted the concept of inside information, including the requirement of sufficient specificity.

The case is important because market-surveillance systems must distinguish between:

ordinary market information;

rumours;

incomplete information;

sufficiently precise inside information.

AI relevance

An AI system may flag thousands of transactions and communications.

If its model uses an excessively broad definition of suspicious information, it could generate:

false positives;

unnecessary investigations;

improper disclosures;

reputational harm.

Principle

Automated surveillance must apply the legally relevant definition of inside information rather than treating every unusual or predictive market signal as legally significant information.

12. Case 3 — Daimler AG v BaFin

Case C-19/11

The Geltl litigation is particularly useful in understanding how European law defines inside information and when information becomes sufficiently precise.

For AI surveillance, this creates an important methodological problem:

A prediction is not necessarily inside information.

An AI model may predict that a company's share price will rise.

That does not automatically mean that the model has discovered legally protected inside information.

Therefore:

Prediction ≠ inside information

and

statistical anomaly ≠ market manipulation.

13. Case 4 — Genil 48 SL and Comercial Hostelera de Grandes Vinos SL v Bankinter SA

Case C-604/11

CJEU, 30 May 2013

The Court considered investor-protection obligations under MiFID.

The case concerned investment services and the obligations applicable to investment firms.

AI relevance

Suppose an investment firm uses an AI surveillance system to classify client behaviour as suspicious or high risk.

The firm remains responsible for complying with applicable investor-protection requirements.

It cannot necessarily argue:

"The algorithm made the classification."

Principle

Automated financial systems operate within the regulated responsibilities of the financial institution using them.

This is particularly important where AI is incorporated into compliance and risk-management systems.

14. Case 5 — Petruchová v FIBO Group Holdings

Case C-208/18

CJEU, 3 October 2019

The case concerned financial contracts for difference and the application of EU jurisdictional rules to consumer/investment relationships.

Although not an AI-surveillance case, it is relevant to determining the legal context surrounding online financial transactions.

AI relevance

AI surveillance increasingly monitors:

retail trading;

automated trading;

CFDs;

derivatives;

cryptocurrency-related transactions.

The case demonstrates the importance of correctly identifying the legal relationship between:

trader → broker → platform → financial institution.

This affects jurisdiction, contractual duties and available remedies.

15. Case 6 — Kolassa v Barclays Bank

Case C-375/13

CJEU, 28 January 2015

The case concerned investor claims and jurisdiction in connection with financial instruments.

The Court examined the connection between an investor's loss and the place where relevant damage occurred.

AI relevance

AI financial surveillance frequently operates across borders.

For example:

German trader → French broker → Luxembourg platform → Dutch cloud provider → London market.

If AI surveillance causes an incorrect account suspension or trading restriction, several jurisdictions may potentially be involved.

Principle

Cross-border financial AI disputes require careful analysis of jurisdiction, applicable law and the location/nature of the relevant damage.

16. Case 7 — Genil 48 / Bankinter and MiFID Investor Protection

The broader significance of the MiFID jurisprudence is that financial firms remain subject to regulatory duties even when sophisticated technology is used.

Therefore:

Technology does not automatically replace the financial institution's legal responsibilities.

This is particularly important for:

AI surveillance;

robo-advisory;

algorithmic trading;

automated suitability assessment;

automated compliance systems.

17. Case 8 — Verein für Konsumenteninformation v Volkswagen

Case C-343/19

CJEU, 9 July 2020

The case concerned the location of damage resulting from unlawful software manipulation.

Although it was not a financial-market surveillance case, it is highly relevant to software-related civil liability and cross-border damage.

AI relevance

The case demonstrates that software can be central to a civil wrong and that digital misconduct can generate geographically distributed economic consequences.

For AI financial surveillance:

defective software → incorrect detection → account restriction/trading loss → cross-border damage.

The Volkswagen litigation therefore provides useful analogical reasoning concerning software-related harm.

18. Case 9 — Schufa Holding AG

Case C-634/21

CJEU, 7 December 2023

This is one of the strongest modern analogies for AI financial-market surveillance.

The case concerned automated credit scoring.

The Court considered the creation of a probability score through automated processing and Article 22 GDPR.

The decision demonstrates that an algorithmic score can have significant legal consequences where another entity substantially relies upon it.

AI surveillance relevance

Imagine a surveillance model producing:

"Market-abuse probability: 97%."

If a financial institution automatically uses that score to:

suspend an account;

terminate a relationship;

refuse services;

restrict trading;

GDPR automated-decision rules may become relevant depending on the circumstances.

Principle

An algorithmic score can have legal significance where it substantially determines consequential treatment of an individual.

19. Case 10 — Dun & Bradstreet Austria

Case C-203/22

CJEU, 27 February 2025

This case further developed the law surrounding automated decision-making and meaningful information about the logic involved.

It is important for AI surveillance because an affected person may need sufficient information to understand and challenge an automated decision.

Application

Suppose an AI system flags a trader:

"High market-manipulation risk."

The trader may want to know:

what data produced the score;

which factors were relevant;

whether erroneous data were used;

whether the model was functioning correctly;

whether the decision was entirely automated;

whether human review occurred.

Principle

Meaningful explanation and the ability to challenge automated processing are important where GDPR automated-decision protections apply.

20. False-Positive Surveillance

A false-positive claim could look like:

Legitimate trading

↓

AI incorrectly detects manipulation

↓

Broker freezes account

↓

Investor cannot trade

↓

Market opportunity lost

↓

Financial loss

Potential civil claims could involve:

breach of contract;

negligence;

wrongful suspension;

GDPR;

financial-regulatory obligations;

damages.

But the claimant would still need to prove the appropriate legal elements.

21. False-Negative Surveillance

A false-negative situation is different.

Example:

Trader manipulates market → AI fails to detect conduct → another investor suffers loss.

The important question becomes:

Does the surveillance provider owe the injured investor a private-law duty?

This is much more difficult than showing that the surveillance system failed.

A regulatory duty to report or monitor does not automatically mean every injured investor has a private damages action.

The claimant may need to establish:

a contractual duty;

a national tort/delict duty;

statutory civil liability;

sufficiently direct causation.

22. AI Surveillance Provider Liability

A technology company may provide:

surveillance software;

AI models;

cloud infrastructure;

data feeds;

analytics;

alert systems.

A defective system could produce:

inaccurate alerts;

missed alerts;

corrupted data;

system outages;

delayed detection.

Liability depends on the contractual relationship.

For example:

Contract A

Bank ↔ AI surveillance provider

The contract may contain:

service levels;

accuracy requirements;

warranties;

liability caps;

indemnities;

cybersecurity obligations.

Contract B

Bank ↔ customer

The bank separately owes obligations to the customer.

Thus:

AI provider's contractual liability and bank's customer-facing liability are not necessarily identical.

23. Financial Institution Liability

A bank or investment firm may remain responsible for:

selecting appropriate surveillance technology;

configuring the system;

monitoring performance;

validating outputs;

responding to alerts;

maintaining human oversight;

keeping records;

complying with MAR/MiFID obligations.

A financial institution cannot necessarily transfer all responsibility to its software vendor.

24. Human Oversight

Human oversight is particularly important.

Consider:

AI = 99% manipulation probability.

A compliance officer automatically closes the account.

Later it is discovered that:

the AI confused two traders with similar identifiers.

Questions arise:

Was human review required?

Was the model's confidence score reliable?

Was the system appropriately validated?

Was the decision proportionate?

Was there a mechanism to correct the error?

The greater the consequence, the greater the importance of appropriate governance and review.

25. AI Model Defects

AI surveillance can fail because of:

1. Data defects

Incorrect market data.

2. Model defects

Incorrect statistical assumptions.

3. Training defects

Insufficient examples of manipulation.

4. Concept drift

Market behaviour changes after model training.

5. Overfitting

Model incorrectly identifies historical patterns as universal.

6. False positives

Legitimate activity is classified as suspicious.

7. False negatives

Manipulation remains undetected.

8. Adversarial behaviour

Traders deliberately modify behaviour to evade detection.

9. Cyberattack

Attackers manipulate the data or model.

26. Market Manipulation and AI

AI surveillance itself must understand the legal concept of manipulation.

Potential conduct includes:

spoofing;

layering;

wash trading;

pump-and-dump schemes;

false orders;

benchmark manipulation;

dissemination of misleading information.

However:

An unusual algorithmic trading pattern is not automatically market manipulation.

Intent, effect, market context and applicable MAR provisions must be considered.

27. Causation

Causation can be particularly complicated.

Example:

AI false positive

→ trading account suspended

→ investor misses a trade

→ price moves

→ investor suffers loss.

The defendant might argue:

the trade might not have been profitable;

the market could have moved differently;

other causes contributed to the loss;

the claimant failed to mitigate the loss.

Therefore:

The existence of an AI error does not automatically prove the amount of financial loss caused by it.

28. Loss Calculation

Possible losses may include:

Direct financial loss

Actual monetary loss caused by an erroneous restriction.

Lost opportunity

Potential profits that could have been achieved.

This is usually more difficult because it requires a counterfactual assessment.

Transaction costs

Additional costs caused by the error.

Hedging costs

Expenses incurred to protect against the consequences of an incorrect suspension.

Reputational loss

May be recoverable in certain circumstances under national law, but requires careful proof.

29. GDPR and Surveillance

Where traders are identifiable individuals, surveillance data can constitute personal data.

Potential issues include:

excessive data collection;

inaccurate data;

unlawful profiling;

excessive retention;

inadequate transparency;

automated decisions.

For example:

AI incorrectly labels trader X as a "high-risk manipulator."

If the label is stored and used to make consequential decisions, GDPR accuracy and automated-processing issues may arise.

30. Data Accuracy

Article 5(1)(d) GDPR requires personal data to be accurate and, where necessary, kept up to date.

Therefore:

Incorrect identity matching can become a serious legal problem.

Example:

Trader A performs suspicious transactions.

AI mistakenly associates those transactions with Trader B.

Trader B's account is then restricted.

Potential issues include:

inaccurate personal data;

unlawful automated decision;

contractual breach;

negligence;

financial loss.

31. DORA and ICT Failures

AI surveillance systems are increasingly dependent on:

cloud computing;

APIs;

market-data providers;

third-party AI vendors;

cybersecurity systems.

If a system outage causes failure to monitor market abuse, the legal analysis may involve:

DORA + contract + operational-resilience obligations + national civil law.

The key issue is again:

Which actor had the relevant legal duty?

32. Evidence in AI Surveillance Litigation

Evidence may include:

Trading evidence

order-book records;

transaction timestamps;

execution logs;

trading patterns.

AI evidence

model version;

model outputs;

confidence scores;

alert history;

false-positive/negative statistics.

Data evidence

market data;

identity records;

customer data;

communications.

Governance evidence

AI risk assessments;

validation reports;

testing records;

human-oversight procedures;

model-change records.

Vendor evidence

service-level agreements;

warranties;

audit reports;

incident reports.

33. Black-Box Problem

A trader may say:

"I do not know why your AI classified me as suspicious."

The financial institution may respond:

"The model is proprietary."

This creates tension between:

trade secrets

and

effective legal challenge.

European data-protection law can provide certain rights to meaningful information concerning automated decision-making, while financial institutions may also have legitimate confidentiality interests.

The exact balance depends on the applicable law and circumstances.

34. Regulatory Liability vs Civil Liability

This distinction should always be made.

Regulatory question

Did the institution comply with MAR/MiFID/DORA/AI Act requirements?

Civil question

Did the claimant suffer legally recoverable damage because of conduct for which the defendant is legally responsible?

These are not necessarily identical.

For example:

Regulatory failure → administrative sanction.

does not automatically mean:

Regulatory failure → automatic compensation to every affected trader.

35. Liability Matrix

36. Key Cases — Revision Table

37. Practical Legal Test

A European AI financial-surveillance claim can be tested through the following sequence:

Step 1

Was AI actually involved?

Step 2

What function did it perform?

Step 3

Was it:

detection;

scoring;

classification;

reporting;

account restriction;

trading control?

Step 4

What legal duty applied?

MAR;

MiFID II;

MiFIR;

DORA;

GDPR;

AI Act;

contract;

national tort law.

Step 5

Was there an AI error?

Step 6

Was there human review?

Step 7

Was the error foreseeable or preventable?

Step 8

Did the claimant suffer damage?

Step 9

Did the AI error cause that damage?

Step 10

Which actor is legally responsible?

38. Important Distinction: AI Alert vs AI Decision

This is one of the most important issues.

Situation A

AI produces an alert:

"Possible manipulation."

Human investigator investigates and closes the case.

Civil liability is different from:

Situation B

AI automatically:

blocks account → reports trader → terminates relationship.

The second scenario creates much greater questions concerning:

automated decision-making;

proportionality;

human oversight;

procedural fairness;

contractual duties;

GDPR.

39. Future Development

The most significant future development is the combination of:

MAR + MiFID II + DORA + GDPR + AI Act + national civil law

This will increasingly require financial institutions to demonstrate:

AI governance;

model validation;

data quality;

human oversight;

auditability;

cybersecurity;

incident response;

explainability where legally required.

The law is therefore moving from a simple question of:

"Did the trader commit market abuse?"

toward additional questions concerning:

"Was the surveillance architecture itself appropriately designed, governed, validated and operated?"

40. Conclusion

AI financial-market surveillance liability in Europe is fundamentally a question of allocating responsibility between the AI developer, financial institution, trading venue, data provider, human compliance officer and potentially other service providers.

The most important legal principles are:

An AI alert is not automatically proof of market abuse.

A false positive can potentially cause contractual, tortious, GDPR or other civil consequences.

A false negative does not automatically create a damages claim for every investor.

Financial institutions remain responsible for their regulatory and contractual obligations even when AI is used.

AI-generated scores can trigger GDPR automated-decision rules where their legal conditions are satisfied.

Human oversight is particularly important when AI outputs produce significant consequences.

Software and AI-related failures create difficult causation and evidence questions.

Cross-border financial disputes require careful jurisdiction and applicable-law analysis.

DORA adds an important operational-resilience dimension.

The AI Act adds AI governance requirements but does not replace national civil liability.

Core Formula

AI Surveillance + Regulatory/Contractual Duty + Algorithmic Error or Failure + Unreasonable Reliance/Operation + Financial Damage + Causation = Potential Civil Liability

Ultra-Short Revision

Spector Photo Group → insider dealing
Geltl → inside information
Genil 48 → MiFID protection
Petruchová → financial contracts
Kolassa → investor damage/jurisdiction
Volkswagen → software-related harm
SCHUFA → automated scoring
Dun & Bradstreet → algorithmic explanation

Main principle: AI can automate financial surveillance, but automation does not by itself eliminate the legal duties of the institution deploying or relying upon the system.