A collaborative article by Chartis Research and Silent Eight

Jump to: Generative AI | Managing the risks

The ability of generative pre-trained transformer (GPT) technology to create new content and answer challenging questions has captured the public’s imagination. Media hype notwithstanding, artificial intelligence (AI) has long been used by financial services firms to process large volumes of sensitive data quickly and accurately. Advanced techniques such as natural language processing (NLP) now play a key role in the fight against financial crime. In this paper, we examine some of the key challenges financial institutions face when managing financial crime risks – and how NLP can help.

Language models and NLP – the foundations

Historically, NLP has focused on the use of computers to understand, interpret and generate human language. It has progressed through distinct phases: from early rule-based systems reliant on handcrafted linguistic rules, to statistical models that use probabilistic methods, and later to neural networks and word embeddings that capture semantic relationships. The development of transformer-based architectures began with Attention Is All You Need, an important paper in the evolution of machine learning, published in 2017, which kickstarted the shift to large language models (LLMs). Transformers revolutionized NLP by allowing models to capture both short- and long-range dependencies in text, and models like GPT and bidirectional encoder representations from transformers (BERT) have built on this architecture.

Back to top

Generative AI – creating something new

AI and LLMs hold considerable promise for a variety of anti-money laundering (AML) workflow compliance tasks. Specifically, they can:

• Automate document analysis. Organizations can leverage LLMs to automate the analysis of Know Your Customer (KYC) documents. In corporate/institutional KYC, for example, AI can be used to parse articles of incorporation, shareholder agreements or business licenses, delivering the relevant information automatically. At the individual level, AI is currently being used extensively in identity analysis. By employing NLP techniques, institutions can sift through vast numbers of documents, extracting such pertinent information as personal details, identification numbers and financial records.
• Generate summaries. To comply with KYC regulations, firms often have to review extensive and complex regulatory documents. To help in this, they can task LLMs with generating concise summaries of these documents, distilling key information and compliance requirements. GenAI can be applied to suspicious activity reports (SARs), for example, automating their preparation and analysis, and assembling them into the standardized formats required by regulators.

Back to top

Managing the risks

For financial institutions, adopting GenAI and LLMs can come with several risks that they must consider carefully, including:

• Bias and fairness. GenAI can perpetuate biases present in training data and in the engineering prompts used to train the model itself. This can lead to outcomes that may disproportionately target certain populations or risk groups, and may skew results to such an extent that risks may not be identified.
• Model degradation. Without proper maintenance, the performance of LLMs may degrade, leading to inaccurate results. For example, retraining a model with updated data is necessary to keep it current with changing language patterns, regulatory requirements or evolving financial crime typologies.
• Hallucinations. A common issue with LLMs and GenAI is the presence of ‘hallucinations’, whereby a model predicts an outcome rather than retrieving or summarizing information, leading to incorrect answers to queries. False positives or negatives caused in this way can be extremely hazardous.

To mitigate these risks, firms must have robust oversight and governance frameworks. Regular testing and validation ensure that models perform as intended and remain compliant with regulatory standards. Establishing continuous feedback loops allows for ongoing refinement based on the real-world outcomes of models, while monitoring tools can detect and correct biases or errors early.

In managing the risks associated with adopting GenAI in AML processes, financial institutions must prioritize these risk management strategies. This involves implementing robust governance frameworks for the data, training and modeling of an LLM in a manner that matches the intended results and use cases. In this way firms can ensure fairness, transparency and accountability in AI decision-making. Financial institutions should also invest in stringent data privacy and cybersecurity measures to safeguard sensitive customer information and protect against potential breaches. Regular monitoring, auditing and retraining of GenAI models is essential to maintain performance and compliance with evolving regulatory standards.

Back to top