Interview with Gao Jiasheng: Exploring AI‑Driven Anti‑Money Laundering Strategies Amid Evolving Attack‑Defense Dynamics

1. Current Challenges and Regulatory Landscape

Rapid technological advancement has spurred innovative financial products while also giving rise to more concealed and diversified money‑laundering schemes. Regulatory pressure is intensifying: FATF launched its fifth round of mutual evaluations in 2024, and China’s revised Anti‑Money‑Laundering Law took effect on 1 January 2025, imposing stricter risk‑based obligations, dynamic customer due‑diligence, and higher penalties.

2. Six Dimensions of AI‑Enabled AML

Gao outlines six AI‑driven dimensions that transform AML from passive compliance to proactive, intelligent, and continuous defense. These include (1) shifting to proactive risk prediction and real‑time alerts, (2) leveraging large‑model AI for workflow automation and cost reduction, (3) achieving full‑process dynamic perception of data, (4) enabling fine‑grained, personalized risk decisions through multimodal feature fusion, (5) bridging the “last mile” between technology and business by allowing domain experts to co‑build models, and (6) creating a collaborative closed‑loop among business, technology, and compliance.

3. Model Optimization with AI

A case study of a bank shows that replacing a semi‑annual expert‑model update with a dynamic‑monitoring‑indicator mechanism reduced the optimization cycle from roughly two weeks to two days, cutting manual effort while preserving model effectiveness.

4. Decision‑Oriented AI Innovations

Three breakthrough innovations are highlighted: (i) deep optimization of traditional expert models using AI, (ii) decision‑oriented AI that combines multimodal feature fusion, dynamic sample‑weight adjustment, and “one‑size‑fits‑all” adaptive monitoring, and (iii) large‑model and generative AI agents that automate case‑report generation, regulatory interpretation, and KYC assistance.

5. Multimodal Feature Fusion Architecture

The proposed architecture integrates numeric, categorical, time‑series, graph‑structured, and unstructured text features. Numeric and categorical data are embedded, time‑series are captured by a Transformer encoder, relational networks are modeled with Graph Neural Networks, and textual remarks are processed by a pre‑trained BERT model. Cross‑attention fuses these modalities, markedly improving detection accuracy and generalization.

6. Dynamic Weight Adjustment and Similarity Matching

Positive and negative samples identified by analysts are vectorized and stored in a knowledge base. When a new alert arises, similarity search retrieves the most comparable historical cases; if similarity exceeds a threshold, the system automatically down‑weights the risk score, reducing false positives.

7. Dual‑Baseline Adaptive Parameter Optimization

By constructing personal‑behavior and group‑behavior baselines, the system converts static model parameters into quantifiable business features, enabling per‑customer, “thousand‑person‑one‑face” risk monitoring. Parallel multi‑feature computation architecture further boosts batch processing efficiency.

8. Large‑Model and Generative AI Agents

Leveraging models such as DeepSeek, the company builds AI agents capable of self‑learning, automatic analysis, and intelligent reasoning across the AML workflow, including suspicious‑case identification, report generation, client identity verification, and regulatory document interpretation. A pilot with a leading city‑commercial bank demonstrated faster, more accurate risk identification and higher‑quality reports.

Conclusion

Gao asserts that AI is turning AML from a compliance baseline into a core competitive advantage for banks, supporting real‑time risk monitoring, model explainability, and alignment with digital‑intelligence transformation goals.