“The challenges facing financial institutions this year include the need to plug the growing educational gap and the ongoing shortage of qualified financial crime professionals to conduct effective client due diligence. With more work and fewer resources, firms must look to leverage cutting-edge technology to create a centralized financial crime ecosystem – including machine learning and artificial intelligence – to reduce the growing skills gap and, ultimately, mitigate the risk of further enforcement action over the coming 12 months.” Rory Doyle, Head of Financial Crime Policy, Fenergo, January 2024.
That quote encapsulates the challenges financial institutions are facing in ensuring they have a robust financial crime ecosystem that genuinely prevents financial crime and stands up to regulatory scrutiny.
Customer due diligence (CDD) is often referred to as the bedrock of an organization’s financial crime operating model and its design and execution are key to the success of all associated controls, such as sanctions screening or transaction monitoring.
The key question that is being asked as part of CDD is “what risk does this individual or entity pose to my organization?” That’s answered by collating all relevant information that can be used to assess that risk. Who is the customer? Where do they live? What do they do to generate their income? Who do they do business with? These are questions that an organization can answer relatively easily utilizing information provided by the customer (proof of identity, proof of address, etc.) or by interrogating registers of publicly available information.
Beyond this are questions relating to a customer’s broader activities. Has this individual been involved in financial crime (fraud, sanctions, money laundering)? Has this entity been involved in cases of corruption? Has this customer been involved in activities out of sync with the risk appetite of the financial institution?
How does an organization approach these questions? Typically using some form of adverse media screening.
How has adverse media screening worked historically
Adverse media screening has traditionally been manual, time-consuming, and prone to human error. Some organizations still rely on Google (or its competition) string searches, e.g. For example, by searching: “CompanyABCLtd” AND (“fraud” or “corruption” or “money laundering” or “sanctions” or “criminal activities” or “regulatory fines” or “lawsuits” or “legal issues” or “bankruptcy”).
This approach returns a theoretically unlimited number of articles from the internet. In turn this needs to be reviewed by the investigator to determine whether these search results pertain to the subject of interest and how relevant and material they are to assessing the risk that subject poses.
In more recent times, organizations have moved to more automated methods – some utilizing commercially available databases of articles updated regularly by the vendor and screening them in a similar fashion to sanctions name screening (screening customers against core sanctions lists such as OFAC). Others use more advanced vendor solutions that utilize machine learning and natural language processing to sift through vast amounts of global data sources.
Both options are limited to the pre-defined data sources, determined by the vendor or financial institution and both continue to generate significant volumes of false positives; either the subject of the article is not the subject of interest, or the content of the article is not material enough to be considered relevant.
Enter: Large language models (LLMs)
However, recent advancements in artificial intelligence (AI), particularly large language models (LLMs), offer a promising alternative. Before discussing the advantages and disadvantages of LLMs in this arena, let us get a better understanding of what LLMs are and why they might be considered for this activity.
LLMs, such as those developed by OpenAI, represent a transformative development in the field of AI, having emerged prominently over the past decade. These models, built on vast amounts of data, excel at understanding and generating human-like text, facilitating a broad spectrum of applications from automated content creation to sophisticated dialogue systems.
Key players in the landscape of LLMs include GPT (Generative Pre-trained Transformer) and BERT (Bidirectional Encoder Representations from Transformers), each distinctive in its architecture and capabilities. For instance, GPT models are renowned for their ability to produce coherent and contextually appropriate text, while BERT excels in understanding the nuances of language within a given context. As these technologies evolve, they continue to push the boundaries of what machines can comprehend and accomplish, signaling a new era of AI that closely mimics human cognitive functions.
By leveraging LLMs, financial institutions can streamline adverse media screening processes, enhance accuracy, and fortify their defense against financial crime. Depending on their deployment, an institution could reduce their associated overheads by 50% or more; however, the likelihood is that this benefit gain would be traded off against a greater number of more genuine cases for consideration. As ever in financial crime, it’s about finding the balance between effectiveness and efficiency.
Benefits of LLMs in adverse media screening
First and foremost, the attraction to utilizing LLMs is that they have the capability to process vast amounts of unstructured textual data in a fraction of the time it would take a human analyst. This efficiency allows financial institutions to conduct adverse media screening more rapidly, enabling quicker decision-making and reducing operational costs realistically by 50% or more.
One of the biggest problems currently faced by institutions in this field is the reliance on human investigators to review articles flagged during screening to determine if the subject of the article is in fact the person of interest being reviewed. This is exacerbated if the subject has a common name. LLMs possess contextual understanding and semantic analysis capabilities, meaning that they can discern nuances in language, identify relevant information, and differentiate between false positives and genuine risks with greater precision.
Let’s look at a hypothetical example to bring this to life:
Client Information:
- Name: John Smith
- Occupation: CEO of ‘Ecomax’, a technology firm
- Location: San Francisco, California
- Recent activities: Negotiations for a major merger with a competitor
Article Details:
- Article 1: Mentions a John Smith involved in a fraud case in New York.
- Article 2: Discusses John Smith, CEO of ‘Ecomax’, speaking at a tech conference in Los Angeles.
- Article 3: Features a John Smith arrested for a disturbance at a sports event in Miami.
LLM Analysis Process:
- Extraction of Relevant Information:
- The LLM extracts all mentions of “John Smith” along with associated activities and locations from each article.
- Contextual Comparison:
- Article 1: The LLM notes the location (New York) and context (fraud case) do not match the client’s profile as a tech CEO in San Francisco.
- Article 2: Recognizes the professional title and company name, aligning perfectly with the client’s details. It further validates relevance by the tech conference context, which is likely given the client’s industry.
- Article 3: Identifies the location (Miami) and context (sports event disturbance) as inconsistent with the known professional and personal activities of the client.
- Risk Assessment:
- Article 1 and 3: Classified as false positives due to discrepancies in location, context, and nature of the activities compared to the client’s known background.
- Article 2: Tagged as relevant, requiring further review or immediate attention due to the direct match of professional identity and activities.
- Learning and Updating:
- The model updates its understanding based on feedback from analysts about the accuracy of its matches and misclassifications, improving future screenings.
As organizations grow in scale and financial crime matures, often expanding their operations globally, the ability to scale adverse media screening processes efficiently becomes indispensable. LLMs provide the flexibility to accommodate increasing data volumes without compromising performance. Coupled with the logical extension of the use case to consider not just financial crime risk exposure, but potentially reputational risk, green risk and beyond, a model that can scale proportionally is vital.
Finally, the ability of LLMs to continuously learn and adapt gives them the edge over traditional solutions. Through machine learning algorithms, these models refine their understanding of language patterns and evolve over time. This adaptability ensures that adverse media screening remains effective in detecting emerging risks and evolving tactics employed by malicious actors.
Hallucinations and other considerations for LLMs
Any significantly transformative step in financial crime risk management must be considered carefully, and although the benefits outlined above would seem to create a compelling case for using LLMs to deliver a more effective and efficient model for identifying adverse media risk, there are key factors that have to be taken into account.
The most important and often discussed is that of reliability – driven primarily by bias and hallucinations.
LLMs rely heavily on the quality and diversity of the data used for training. Biases present in the training data, such as skewed representations or incomplete datasets, can propagate into the model’s outputs, leading to inaccurate or unfair outcomes. Ensuring the integrity and representativeness of training data poses a significant challenge for deploying LLMs in adverse media screening (see choice of model below). Further, the model can hallucinate and make mistakes analyzing texts (e.g. study in “Journal of the American Medical Informatics Association” in 2021 where GPT misdiagnosed symptoms of a heart attack as possible anxiety or indigestion). Various anti-hallucination strategies exist to prevent this, such as providing a longer context window, confirmation questions, better and more precise prompting, and using a more advanced model; this all adds complexity to the design.
Then comes the ability to interpret and explain the model. Despite their remarkable performance, LLMs often operate as black-box systems, meaning the decision-making process is not transparent or easily interpretable. This lack of explainability raises concerns regarding regulatory compliance and accountability. Financial institutions must grapple with the challenge of reconciling the need for transparency with the inherent complexity of LLMs. One can validate their behavior by using the anti-hallucinations techniques explained earlier. This means that although an organization may not understand how the model generated the output, they can interpret and explain it by validating it extensively.
Finally, consideration must be given to privacy and ethics. The advantage of LLMs accessing vast quantities of data to assess what risk is posed by an individual or an entity is balanced with the fact that LLMs have the potential to uncover sensitive information about individuals that may not be relevant to the screening process.
An example might be if the person of interest is CEO of a multinational corporation. Deploying the LLM model, it identifies articles relating to accounting irregularities at their company that are wholly relevant to the assessment; however, it also flags articles pertaining to the CEO’s divorce proceedings, which include allegations of personal misconduct. Clearly the latter is sensitive and personal, and not relevant to a financial crime investigation (unless there is a connection between the two cases clearly).
Striking a balance between effective risk management and respecting privacy rights presents a formidable challenge for financial institutions utilizing LLMs.
In conclusion: Huge potential, careful navigation
The adoption of LLMs for adverse media screening holds immense potential to revolutionize the way financial institutions mitigate risks associated with financial crime. Moving away from harvesting volumes of adverse media and screening it against the person of interest (typically an institution’s customer) to asking the core question “what risk does this customer pose to my organization?” is the step change financial crime professionals have been looking for.
By leveraging the efficiency, accuracy, scalability and adaptive learning capabilities of LLMs, organizations can strengthen their compliance efforts and bolster their defenses against emerging threats.
However, realizing the full benefits of LLMs requires addressing significant challenges related to data bias, hallucinations, interpretability, privacy and regulatory compliance. Financial institutions must navigate these challenges thoughtfully and responsibly to harness the transformative power of LLMs while upholding ethical principles and regulatory requirements.
In reality this carries a “Don’t try this at home” warning and organizations are best to engage the services of experts in the industry to work hand in hand to develop a best fit solution. In doing so, they can pave the way for a more effective and resilient approach to adverse media screening in the ever-evolving landscape of financial crime prevention.
Ben Rayner, is Regional Head, UKIMEA, at Silent Eight.
