The same technology that helps banks identify abnormal behaviour is making deception cheaper, faster and more convincing. The contest is becoming less about who has AI and more about whose data, feedback loops and controls learn faster.
Artificial intelligence is creating a peculiar symmetry in financial crime. The defensive side of banking has used machine learning for years to score transactions, detect anomalous behaviour and prioritise investigations. Now the offensive side is acquiring comparable automation. Generative models can write persuasive phishing messages, fabricate documents, clone voices, produce synthetic faces and help less-skilled criminals assemble attack tools that once required specialised expertise. The result is not a simple story in which AI makes fraud worse. It is an arms race in which both sides become faster at the same time.
The banking side is already deeply invested. In March 2026, the European Central Bank reported that nearly 90% of significant euro-area banks were using AI, with fraud and cybercrime detection the most common use case, deployed by more than half of banks. That is evidence that AI is no longer a laboratory tool for financial crime teams; it is becoming part of core risk infrastructure.
Yet the threat is scaling just as rapidly. The 2026 INTERPOL Global Financial Fraud Threat Assessment said AI-enhanced fraud was 4.5 times more profitable than traditional methods in the activity it assessed and warned that agentic AI can increasingly automate whole fraud workflows. The exact profitability ratio should be read as an INTERPOL assessment rather than a universal market statistic, but the direction is clear: automation is changing the economics of deception as well as the economics of defence.
CORE ARGUMENTAI does not automatically give criminals the advantage. Banks have richer transaction histories, regulated identities, network visibility and the ability to delay or stop money. But criminals can iterate cheaply, attack outside bank perimeters and increasingly manipulate the customer rather than the account. The decisive advantage will belong to the side that combines data, identity, behavioural intelligence and rapid feedback most effectively.
Banks already use AI where traditional rules are weakest
Conventional fraud controls are built around known patterns: a card used in an unusual country, a login from a new device, a transfer far larger than the customer normally makes. Rules remain useful, but criminals learn the rules too. Once a pattern is consistently blocked, attackers change device fingerprints, transaction sizes, timing, beneficiary accounts or social-engineering scripts. The weakness of static controls is that they recognise yesterday’s fraud unusually well.
Machine learning changes that model by combining many weak signals. A payment may not look suspicious in isolation, but its device, counterparties, timing, velocity, account history and links to other accounts can form a pattern. The BIS Innovation Hub’s Project Hertha demonstrated the potential at payment-system level. In a synthetic dataset covering 1.8 million accounts and 308 million transactions, payment-system analytics helped banks and payment service providers identify 12% more illicit accounts than they would otherwise have found. For previously unseen criminal behaviours, the improvement was 26%.
Those results are important but should not be overstated. Hertha used simulated data rather than live customer transactions, and the BIS explicitly described the analytics as a supplementary tool rather than a complete solution. Even so, the project illustrates why AI can favour defenders when data is aggregated across a network. Criminals frequently distribute activity across many accounts and institutions; a single bank sees fragments, while a payment system can see relationships.
This network perspective is also central to the BIS discussion of AI and financial-crime controls. Machine-learning systems can use transaction behaviour, KYC information and network patterns to find combinations that rules-based monitoring may miss. Inference: as payment systems become faster and more interconnected, the value of cross-institution pattern recognition will rise because criminals can move stolen funds through mule networks before any one institution has enough local evidence to understand the full chain.
The criminal advantage is not better intelligence; it is cheaper persuasion
The most disruptive use of generative AI may not be technical intrusion at all. It is persuasion at scale. Fraud has always depended on convincing a victim, employee or institution to accept a false identity or false story. Generative AI lowers the cost of producing that story in the language, tone, accent and visual format most likely to succeed.
INTERPOL’s 2026 assessment describes an increasingly industrialised fraud ecosystem in which AI-generated content, automation and specialised criminal services are combined. A separate Operation First Light 2026 involved 97 countries and territories, led to 5,811 arrests and intercepted $293 million in illicit assets linked to social-engineering scams and associated laundering. That operation is not evidence that every case involved AI. It is evidence of the scale and international organisation of the fraud infrastructure into which AI tools are now being inserted.
The technology also broadens the labour pool available to criminals. FINRA’s 2026 regulatory oversight report warns that GenAI can generate false identification documents, deepfake audio and video, polymorphic malware and malicious tools that allow people without advanced technical skills to conduct more sophisticated attacks. This matters because the economics of cybercrime depend partly on scarce expertise. If a model can produce working code, translate scripts, personalise phishing messages and create synthetic media, capability diffuses to a larger number of attackers.
Identity verification is becoming an adversarial contest
Banks built digital onboarding around the assumption that identity evidence is expensive to fake. A passport image, selfie, voice recording or video call used to carry some evidentiary weight because creating a convincing forgery required effort. Generative AI weakens that assumption by making images, voices and documents easier to manufacture and easier to vary.
The Financial Crimes Enforcement Network said in November 2024 that it had observed an increase in suspicious-activity reporting describing suspected deepfake use, particularly fraudulent identity documents intended to bypass verification and authentication. The alert is especially significant because it was based on reports from financial institutions rather than hypothetical laboratory attacks. By 2026, the EBA was likewise warning that AI could be used to forge documents, generate deepfakes, automate laundering schemes and evade detection.
The defensive response cannot therefore be “better facial recognition” alone. If the input itself can be synthetic, identity must be tested through multiple independent signals: document provenance, device reputation, account history, behavioural patterns, trusted data sources, liveness checks, transaction context and, for higher-risk events, independent human verification. Inference: the strongest identity architecture will increasingly resemble a risk model rather than a one-time check. The question will not be whether a document looks real, but whether the entire digital behaviour around the person is coherent.
Fraud is shifting from stealing credentials to manipulating the customer
Stronger authentication has changed the attacker’s target. The joint EBA-ECB payment-fraud report found that strong customer authentication remains effective against the fraud types it was designed to prevent, particularly card fraud. Yet total reported payment fraud across the European Economic Area rose to €4.2 billion in 2024 from €3.5 billion in 2023, and the authorities highlighted the growing importance of manipulating payers themselves.
The same pattern appears in the United Kingdom. UK Finance’s Annual Fraud Report 2026 reported £1.28 billion of payment-fraud losses in 2025, up 4%. The banking industry prevented £1.68 billion of unauthorised fraud, but authorised push-payment losses rose to £576.4 million, up 19%, across 248,070 cases. The contrast is revealing. Banks can become better at stopping unauthorised transactions while criminals shift toward persuading customers to authorise the transfer themselves.
AI is well suited to that shift. A fraudster does not need to defeat encryption if a realistic voice call, video message or tailored conversation persuades the account holder to perform the transaction. The traditional boundary between cybersecurity and fraud therefore weakens. Identity theft, phishing, account takeover, malware and social engineering become components of one adaptive attack chain.
The FBI’s November 2025 account-takeover alert gives a measure of the exposure even without isolating AI-enabled cases: since January 2025, the FBI’s Internet Crime Complaint Center had received more than 5,100 complaints about account-takeover fraud, with reported losses above $262 million. The schemes often impersonated financial-institution staff and manipulated victims into surrendering credentials or one-time codes. Generative AI can make the impersonation layer more convincing, but the underlying vulnerability remains human trust.
The counterargument: AI may ultimately favour the defenders
There is a credible case that the balance of power will eventually favour banks. Criminals may gain access to the same general-purpose models as everyone else, but banks possess something more difficult to copy: proprietary behavioural data. They know how a customer normally logs in, where the customer sends money, what devices are trusted, what merchant patterns are typical and how suspicious flows connect across time. Large institutions also have the legal ability to block transactions, freeze accounts, contact customers and share information with law enforcement.
The ECB’s adoption data supports this view: fraud and cybercrime detection is already the leading AI use case among major euro-area banks. The BIS Hertha experiment suggests that a wider network view can materially improve detection of novel patterns. And UK Finance data showing £1.68 billion of unauthorised fraud prevented in 2025 demonstrates that defensive controls already stop large volumes of attempted theft.
The U.S. Treasury has nevertheless identified a structural weakness in that advantage. Its report on AI-specific cybersecurity risks in financial services highlighted a “fraud data divide”: large institutions have much more historical data with which to train models, while smaller banks and credit unions may lack comparable resources. That means AI can improve system-wide defence while simultaneously widening the gap between institutions. Criminals will naturally probe the weakest firms, onboarding providers and payment endpoints rather than attack the best-defended bank head-on.
Defensive AI can create new risks of its own
More automation does not eliminate model risk. A fraud model can learn from biased or incomplete labels, overreact to legitimate behaviour, or become stale when criminals change tactics. False positives are not a minor nuisance: they can block customers from their own money, increase call-centre demand, push vulnerable users into alternative channels and create conduct or discrimination concerns. A model that stops more fraud but incorrectly freezes thousands of legitimate payments may not be a better control.
The EBA’s June 2026 Risk Assessment Report shows why the risk is broader than fraud operations. Supervisors are increasingly concerned that frontier AI systems may improve attackers’ ability to discover and exploit software vulnerabilities, while banks still face capacity constraints in responding quickly. The defensive institution therefore has two AI problems at once: it must use AI effectively, and it must secure the AI-enabled technology stack on which that defence depends.
The governance problem becomes sharper when banks rely on external vendors for fraud analytics, identity tools or generative-AI infrastructure. The ECB notes that banks are more likely to rely on third-party providers for fraud detection and cybersecurity than for some proprietary applications. That can accelerate capability, but it also creates concentration and model-dependency risk. If many institutions use similar data providers, liveness systems or model families, a weakness discovered by criminals could propagate across the sector.
The decisive resource is the feedback loop
Fraud models improve when investigators rapidly label what happened, confirmed scam patterns are shared, mule accounts are linked across institutions and customer reports feed back into detection logic. Criminal groups operate with their own feedback loops: which scripts produce replies, which voice clones pass, which banks release payments, which onboarding controls fail and which cash-out routes stay open. The arms race is therefore not simply model versus model. It is feedback loop versus feedback loop.
This is why data sharing matters as much as model quality. One bank may see a new beneficiary account receiving three suspicious transfers. Ten banks together may see the same account receiving hundreds. A telecom provider may see the spoofed call pattern that preceded the transfers. A social platform may see the advertising account that originated the investment scam. A crypto exchange may see the final conversion. Each organisation owns a fragment of the attack; the criminal controls the sequence.
The policy challenge is to connect those fragments without building an indiscriminate surveillance system. Privacy, data minimisation, explainability and due process remain necessary. Project Hertha is useful precisely because it explored whether a limited set of payment-system data could add value without assuming unrestricted access to everything. The next generation of fraud infrastructure will have to demonstrate not only that shared analytics work, but that they can be governed.
What the arms race means for banks, fintechs, regulators and investors
For banks, the strategic priority is to move from isolated fraud tools toward an integrated identity-and-behaviour architecture. Transaction monitoring, onboarding, authentication, device intelligence, customer communications and case management should feed one another. The goal is not to maximise the number of alerts. It is to shorten the time between a new criminal tactic appearing and the institution recognising it across channels.
For fintechs, the risk is that low-friction onboarding becomes a competitive feature until attackers discover it. Firms that outsource KYC, liveness, device scoring or payment processing cannot outsource accountability for how those components work together. The strongest fintechs may be those that can preserve fast customer journeys while escalating dynamically when behaviour becomes inconsistent, rather than forcing every user through maximum friction.
For regulators, the central problem is coordination. The U.S. Treasury’s broader 2024 AI report recommended continued information sharing and adaptation of existing risk-management frameworks rather than treating AI as an isolated technology category. The EBA and ECB payment-fraud data point in the same direction: controls designed for stolen credentials are less effective when the customer is manipulated into authorising the payment. Regulation has to follow the attack path, not merely the technical rail.
For investors, fraud resilience is becoming a measure of operating quality. The important indicators are not only gross fraud losses, but the rate of false positives, reimbursement costs, customer attrition after fraud events, dependence on third-party identity vendors, speed of model refresh and the ability to share intelligence. A bank that spends heavily on AI but still treats fraud, cyber, AML and customer authentication as separate silos may have bought more technology without building a stronger defence.
Conclusion: the future of fraud is a competition in learning speed
AI is arming both banks and criminals because both sides face the same basic problem: detect patterns, predict behaviour and act before the other side can respond. Banks have structural advantages in data, regulated identity, transaction visibility and the power to stop money. Criminals have different advantages: they can experiment cheaply, ignore legal constraints, attack across borders and exploit the fact that people—not just systems—can be deceived.
That makes the end state less dramatic than the idea of an autonomous fraud war, but more consequential. The winning bank will not be the one with the largest model or the most alerts. It will be the institution that can connect identity, behaviour, network intelligence and human investigation into a feedback loop that learns faster than the criminal network attacking it.
Fraud detection is therefore becoming fraud adaptation. AI may improve the defender’s odds, but it also shortens the attacker’s development cycle. In the next phase of financial crime, trust will depend on whether banks can make their learning systems as dynamic as the threats they are built to stop.
References
1. European Central Bank — “AI and the euro area economy” (23 March 2026)
3. INTERPOL — “Global Financial Fraud Threat Assessment 2026” summary (16 March 2026)
4. INTERPOL — “Over 5,800 arrests, USD 293 million intercepted in global fraud bust” (9 July 2026)
5. UK Finance — “Annual Fraud Report 2026” (15 June 2026)
8. European Banking Authority — “Risk Assessment Report - June 2026”
9. European Banking Authority — “Risk Assessment Report - December 2025”
11. FINRA — “2026 Annual Regulatory Oversight Report: Cybersecurity and Cyber-Enabled Fraud”
12. FBI — “Account Takeover Fraud via Impersonation of Financial Institution Support” (25 November 2025)
15. BIS Annual Economic Report 2025, Chapter III — “The next-generation monetary and financial system”