Guide to Artificial Intelligence in Jersey's Finance Industry

The financial and professional services industry is undergoing continued change because of emerging technologies. Current research focusses mainly on how these technologies will change the industry at a higher level, with little detail given on how the day-to-day work of employees will be impacted and what practical skills they will need to perform this work and remain relevant in the job market.

Companies that embrace Artificial Intelligence (AI) and support upskilling to help professionals adapt to, and benefit from, changes brought about by AI will have the most success attracting — and retaining — top talent.

Financial services stand out as the only industry in which the share of professionals with AI skills and the speed at which they are adding AI skills to their profiles is above average. This is an example of how industries, beyond tech, have the potential to be not only early adopters but drivers of AI innovation.

Jersey Finance’s fintech ambition is to strive to be the easiest international finance centre to do business with remotely, in a digital world. Our fintech strategy centres around two overarching strategic ‘drivers’: ‘enhancing client experience’ and ‘driving efficiency.’

These drivers have been put in place to guide our approach to fintech, ensuring that Jersey’s finance industry can continue to deliver world-class services to international clients in an increasingly virtual world. They also empower firms to support greater productivity and satisfy growing regulatory and reporting requirements, as well as skills shortages.

Working in partnership with Grant Thornton UK, we are proud to launch a new series of insights into AI, diving into the key topics surrounding the use of AI by the financial services community, both globally and specifically in Jersey.

Released over the coming months, we will examine AI from a variety of viewpoints, looking at several topics, from the genesis of AI, key technologies used now and what skills will be needed in the future.

The Evolution of AI

The rapid rise of AI has not come out of the blue. The history of AI as a field of systematic research traces its origins back as far as the 1950s, when the term ‘artificial intelligence’ was originally coined. The evolution of AI since then can be defined across several distinct phases, each marked by significant advancements in theory, methodology and application. The continued development has been characterised by several distinct periods of intense growth, decline and resurgence.

The ‘First Golden Age’ of AI began in the mid–20th century, fuelled by optimism and significant advances in computational power and algorithms, leading to pioneering programs in problem-solving and logic.

This section will look at the specific technologies that make up Artificial Intelligence (AI) as well as capabilities that are expected to be developed in the future.

AI is typically classified according to how the machine learning algorithm enhances its capabilities. The field of AI consists of three primary approaches. These approaches are referred to as ‘Supervised Learning,’ ‘Unsupervised Learning,’ and ‘Reinforcement Learning,’ and give AI systems the flexibility to tackle diverse problems within varied applications:

Supervised Learning

Supervised Learning is a type of machine learning approach that involves predicting the value of a pre-defined target/outcome based on known inputs. This approach requires labelled data for the model training, which allows the algorithm to capture the relationship between the input labels (features) and the target outcome. Supervised learning can be seen in action in your email folder: separating spam emails into a spam folder in your inbox is a type of machine learning that uses supervised learning.

Unsupervised Learning

In contrast to supervised learning, unsupervised learning identifies patterns and relationships in data without predefined outcomes, making it exploratory and useful when there isn’t a lot of labelled data, or the available labelled data is expensive. It can include cluster analysis for grouping similar observations and association analysis for identifying relationships between variables. Such methods can provide insights such as “customers interested in X often show interest in Y and Z as well”.

Reinforcement Learning

This technique involves an ‘agent’ exploring an environment to identify optimal actions through trial and error, relying on a reward function for feedback. This approach is beneficial in settings with unknown optimal actions and dynamic problem types like robotics or game playing and is used in areas such as autonomous vehicles and dynamic pricing within financial services, where the environment frequently changes.

In addition to the three primary approaches, there are a number of other classifications of AI which make up the full suite of technologies.

Machine Learning

Traditional machine learning algorithms are designed to perform tasks like prediction or classification by analysing input data to identify underlying patterns. Various algorithms have been developed in traditional machine learning, with differing functions depending on the nature of the data inputs and the specific task that the algorithm is applied to. To function effectively, traditional machine learning relies on extracting specific characteristics from data through a process called feature engineering, which is where raw data is transformed into data that can be used by the machine learning algorithms.

Deep Learning

While machine learning falls under the umbrella of AI, deep learning can be considered a subset of machine learning. In contrast to traditional machine learning, deep learning automates the process of extracting information from input data, removing the need for human intervention in feature engineering. By doing so, deep learning can automatically determine the most optimal features for the task at hand, as opposed to the data being assigned features externally by a human.

Generative AI

This type of AI refers to a class of AI systems designed to generate new content or data that resembles and, in some cases, is indistinguishable from human-created content. It is a subset of deep learning. At its core, generative AI works by learning patterns and structures from existing data and then using that knowledge to create new content. This content can span various domains, such as images, text, audio, and video, or a combination of those domains.

Predicting the trajectory of AI development is challenging due to its rapid pace, but several trends are increasingly probable. The anticipated developments in AI are set to deeply impact the finance sector in a myriad of ways, from enhancing customer service to redefining how risk is assessed. In this section we discuss how AI adoption is likely to scale over the next two, four, and six years, and its implications for the future landscape of financial services.

Short-Term (Over the next 2 years):

Medium-Term (In the next 4 years):

Long-Term (Over 6 years):

From transparency issues to concerns surrounding bias, privacy, accountability, and systemic risks, the ethical landscape of AI in finance is varied and complex.

Addressing these ethical considerations is not only a task for individual firms but requires action from multiple stakeholders, including regulators, industry professionals, developers, and governments, to ensure responsible deployment and maximise societal benefits.

With these challenges come opportunities, and Jersey, with our close connections between the industry, the JFSC and the Government is ideally suited to build on our strong reputation of being a trusted provider of professional services to becoming a leader in the ethical use of AI in financial services.

Bias in AI refers to systematic errors or inaccuracies in algorithmic predictions that typically result from skewed or unrepresentative training data, imbalanced data, algorithmic design choices, or implicit assumptions. These biases can manifest in various forms, such as underrepresentation or misrepresentation of certain groups, unequal treatment based on protected attributes, or reinforcement of existing societal stereotypes.

Fairness in AI refers to the absence of discrimination or bias in algorithmic decision-making processes. A fair AI system ensures that individuals from different groups are treated similarly under similar circumstances. This may involve defining appropriate fairness criteria, such as ensuring that the false rejection rates are balanced across different demographic segments, and developing AI algorithms that satisfy these criteria.

One of the key concerns in the use of AI algorithms is the potential for biases. These biases can stem from historical data, societal prejudices, or even the design of the algorithms themselves. A common example of this is how historical lending data disproportionately favours certain demographic groups and that AI algorithms trained on this data may perpetuate those biases. As a result, individuals from historically disadvantaged groups may face obstacles in accessing credit, despite their creditworthiness. To address concerns like these, financial institutions need to implement measures to mitigate biases in AI algorithms used as part of their operations.

One approach involves using bias detection and mitigation techniques to assess the fairness of AI algorithms and rectify any potential disparities. As part of the development of systems, AI developers can incorporate fairness-aware machine learning methods to actively address biases during the training and deployment of AI algorithms. In addition, regulatory bodies can work with industry stakeholders to establish guidelines and standards for fair and transparent use of AI within financial services. This may involve requiring financial institutions to regularly assess and report on the fairness of their AI algorithms, ensuring that decisions are free from discriminatory biases.

Diverse and Representative Data

The first step in addressing bias in AI models is ensuring that training data used to develop algorithms is diverse and representative of the full range of population. This involves collecting data from a wide range of sources and curating datasets accordingly. For example, in the context of facial recognition technology, ensuring diversity includes collecting a wide range of facial data from individuals with varying skin tones, ethnicities, ages, and gender identities. By curating these datasets, AI algorithms can be trained to recognise and classify facial features with greater accuracy and fairness, minimising the risk of biased outcomes.

Human Oversight and Intervention

Integrating human oversight and intervention mechanisms into AI systems can help detect and correct biases that may not be apparent from data alone. Instead of relying solely on accuracy, using alternative evaluation metrics can provide a more comprehensive assessment of model performance, especially on imbalanced datasets.

When including human oversight as a bias prevention mechanism, it is important to recognise that humans can also be blind to biases and therefore shouldn’t be the sole measure used to defend against biases in AI algorithms.