AI vs AI: who decides when evidence can be synthetic?

One technology, two sides of risk

AI can make impersonation, documents and fraudulent requests more convincing; it can also help defenders connect anomalies, behaviour and transactions. But a machine-generated or machine-detected signal is not, by itself, a verified conclusion. Which independent evidence would change the decision?

When seeing and hearing are not enough

In Horizon Scan AI and Deepfakes, published on 22 December 2025, FATF describes a video conference involving impersonated executives and a USD25 million transfer (Box 0.2). The source does not identify the firm or jurisdiction. The operational lesson is to separate apparent recognition, authority verification and payment approval, rather than expect every employee to recognise every deepfake.

Illustrative control: an urgent request apparently from a senior executive should be checked through a previously verified independent channel, authorised powers and a second review before execution. These are proposed professional controls, not a specific mandatory AMLA technology checklist.

Synthetic identity and remote onboarding

Voice cloning, altered documents and synthetic identities can pass individually plausible checks without establishing a coherent identity. The proposed test connects identity, document, device, account ownership and declared activity; liveness or a convincing video does not resolve the issue by itself. An inconsistency requires independent corroboration and proportionate escalation, not automatic rejection.

Italy’s FIU communication of 8 June 2026 records deepfakes, synthetic identities and benefits from cooperation between anti-fraud and AML controls. Practical application: connect the identity inconsistency with subsequent account behaviour, rather than treating onboarding and transaction review as separate explanations.

An AI risk score is not a decision

A system flags a customer as high risk. An analyst needs to understand the data and relationships behind the signal, test data quality and currency, consider alternative explanations and record unresolved concerns. Automatic closure and automatic escalation without context can both fail.

Detection, investigation and network analysis help connect identities, devices, counterparties and flows. To counter automation bias, the reviewer tests at least one alternative explanation and records missing data, signal confidence and the reason for any override. The system prioritises the review; it does not predetermine the customer decision.

Criminal AI → Defensive AI → Human Decision

The contest is not resolved by asking which AI is more powerful. Defensive advantage emerges when reliable data, connected signals, collaboration and technology reach a reviewer who can challenge the output and explain the decision. AMLA describes AI as a present risk to identity verification and CDD and, on the defensive side, is building analytics, AI/ML and structured information-sharing capabilities; this is not a universal requirement to buy an anti-deepfake detector.

What the global comparison shows

The 2026 picture converges beyond Europe. INTERPOL describes the industrialisation of fraud, with AI, deepfakes and shared criminal infrastructure increasing scale and speed. In the United States, FinCEN links deepfakes with documents and identities used to circumvent verification and authentication. On the defensive side, Singapore is testing AI/ML models on historical data from multiple banks to identify higher-risk transactions and accounts; in the UK, the FCA is using synthetic data to test AML techniques without sharing real customer data. The lesson is not that one model wins: stronger defence combines data quality, collaboration, testing, governance and human decision-making.

What AMLA is building on the defensive side

AMLA’s 2026–2028 work programme explicitly connects technology and innovation with supervision: it envisages data analytics and AI/ML capabilities, automated data collection, anomaly detection and advanced analytics to support supervisory decision-making. The same document calls for safe AI governance to control risks such as model drift and bias. The useful lesson for an obliged entity is not to copy the authority’s tools, but to apply the same principle: automation to find and connect signals, governance and human judgement to turn them into a defensible decision.

Data quality, confidentiality and a reviewable decision

Before moving data between tools or organisations, establish purpose, legal basis, necessary data and permissions. Record source, currency, limitations and checks. Keep the model output, independent verification and accountable decision distinct: a technical log alone does not explain why the signal was accepted or challenged. These are proposed control-design questions, applied within the relevant legal framework.

• Separate supervisory, group-wide and inter-operator information sharing: each has its own scope and access rules.
• The AMLA database consultation concerns confidential supervisory exchange; it does not authorise unrestricted customer-file sharing between private firms.
• Use authorised environments and proportionate access. Preserve evidence, reviewer, reason for override and review trigger without circulating unnecessary customer data.

Distinguish applicable rules and expectations

Under the AI Act, establish the role and classification of the system: Article 14 oversight concerns in-scope high-risk systems, not every AML application. AMLR and AMLA standards address risk, governance and controls; they do not prescribe a generic AI solution or one detector. Connecting the case to the applicable rule avoids turning a proposed good practice into a non-existent duty.