Artificial Intelligence: From Speculation to Reality
The term 'Artificial Intelligence' encompasses a broad concept and has been defined in various ways by researchers. According to the U.S. National Institute of Standards and Technology, artificial intelligence is “an engineered or machine-based system capable of producing predictions, recommendations, or decisions that influence real or virtual environments to achieve a set of designated goals.” According to a recent report prepared by former UK Prime Minister Tony Blair and William Hague, a member of the British House of Lords, to highlight the importance of artificial intelligence for the future of the country, artificial intelligence is considered “the most important technology of our generation” and is believed to have “an impact level comparable to that of the internal combustion engine, electricity, and the internet.”
In recent years, alongside other digital transformation efforts, the use of and research in artificial intelligence in the finance sector has accelerated significantly. The application of artificial intelligence in finance has been a subject of discussion for a long time, and these discussions often stretch our imagination. As artificial intelligence increasingly becomes a reality, how it takes shape and will continue to evolve in everyday applications is a matter of curiosity for all of us.
Machine Learning's Dominance in the Financial Sector
Many of the artificial intelligence technologies currently used in financial areas fall under the category of 'Machine Learning' (ML). According to a survey published by the Bank of England and the Financial Conduct Authority in October 2022, 72% of the firms surveyed are using or developing ML applications. These ML applications have evolved over time to their current state and have reached a position where they can be more fully integrated into daily operations. Furthermore, 79% of ML applications are in the final stage of development, meaning they are being used extensively in a major business area and/or are critical for some business sectors.
According to the survey mentioned above, the trend of using and developing ML in financial terms is expected to continue, and it is predicted that the median number of ML applications used by firms will increase by 3.5 times within the next three years.
Types of Machine Learning (ML):
• Supervised learning; a process where labeled input data is processed by an algorithm, and the algorithm produces a set of rules that can be applied to new (unlabeled) input data, thus enabling the prediction of the correct labels.
• Unsupervised learning; a process where when unlabeled input data is given to an algorithm, the algorithm tries to detect underlying patterns such as similar behavior groups or relationships.
• Reinforcement learning; an algorithm that receives unlabeled input data operates in a dynamic environment and tries to develop a strategy to maximize positive outcomes through a system supported by rewards and penalties.
Machine Learning and Automatic Decision-Making Relationship
The use of artificial intelligence does not always require autonomous machines operating without human supervision, as often claimed. Supervised and unsupervised learning methods are not designed to influence any action on their own, but they can be used with automation interfaces that can directly trigger real-world outcomes. Reinforcement learning algorithms, by their nature, produce results in a dynamic environment; these results can either directly create real-world effects or trigger processes that require human intervention depending on the application's needs.
In the financial sector, human intervention is often necessary to ensure appropriate regulatory standards are met. There are various methods for human intervention, for example:
Human-in-the-Loop: Requires human approval for each decision.
Human-on-the-Loop: Allows human intervention during the design process and monitoring of system operation.
Human-in-Command: Involves a person who oversees the overall functioning of the system and decides on its use for specific decisions.
Current Deployment Areas In the financial services sector, a wide range of artificial intelligence activities is reported, varying across institutions, sub-sectors, and jurisdictions. To give some examples:
Risk Management: One of the earliest adopted categories, this includes tools for monitoring, detecting, and managing operational, market, credit, and regulatory risks.
Customer Verification and Interaction: Artificial intelligence is used for verifying customer identification information and in customer interactions, particularly through 'chat-bots'.
Insurance: Artificial intelligence is employed for sales support (e.g., increasing risk sensitivity in pricing) and claims management (e.g., regulating payments based on real-world events).
Asset Management: AI techniques are used to support portfolio management. While the analysis of historical performance data has been used over time, today, increasingly diverse data sources and techniques are utilized.
Algorithmic Trading: Rule-based algorithms have long been used in the trading market, but today AI techniques include 'algo-wheels' that allow choosing between alternative trading strategies.
Advisory: Many robo-advisors use rule-based algorithms. When AI techniques are used, they can generate outputs to support decisions made by humans
Reducing Practical Barriers to Adoption – Commoditization
While widespread discussions on the theoretical limits of artificial intelligence technology continue, the inclusion of AI tools in these processes has gradually reduced the related practical barriers. Initially, developing artificial intelligence required expert computer scientists who created specialized codes on specialized hardware, but today this technology has become much more accessible due to the availability of open-source AI software, cloud-based hosting, and processing capabilities, and the development of new tools and facilities.
Over the past few years, the growth in AI Services (AI as a Service - AIaaS) offered by major cloud providers like AWS Sagemaker and Google Cloud AutoML Engine has been notable. These platforms and tools enable organizations to upload and manage data, as well as train various popular machine learning algorithms on this data. The latest development is AI Services offered with a range of 'plug, play, and use' tools. These are typically provided via an application programming interface (API) and can perform general machine learning tasks such as image recognition, speech recognition, translation, and virtual assistants. These tools can be rapidly integrated to deploy AI solutions without any machine learning experience.
The Impact of Generative AI in Finance
Generative AI refers to the capability of machine learning (ML) tools to produce original content such as text, images, sounds, videos, or code, based on the input data used for training.
The Explosion of Generative AI
Generative AI has rapidly entered our world following the emergence of large-scale, open-source models. These models, built on publicly accessible APIs that produce outputs that feel extremely human-like, have significantly accelerated the adoption of AI technology due to their easy accessibility: While it took nearly two years for Twitter (now known as X) to reach one million users in 2006, Instagram reached the same number in just two and a half months in 2010. In November 2022, OpenAI's ChatGPT application reached this number in only five days, marking the fastest adoption of a technology ever, reaching 100 million users within two months.
What Sets Generative AI Apart?
Generative AI, a sub-branch of machine learning (ML), possesses capabilities beyond the reach of traditional artificial intelligence. It can analyze various types of data, including unstructured data, process large datasets faster, and have a broader range of functions. Basic models can produce original content such as text, images, sound, video, or code. Generative AI, crucial for financial services, can identify instant and continuous trends, facilitating real-time monitoring and forecasting. Its advanced architecture and the ability to process and learn from sequential data also enable it to perform nuanced sentiment analysis.
Despite the excitement about the transformative power of generative AI, the financial services sector has adopted a more balanced approach. Firms experienced in traditional AI and machine learning are beginning to experiment with this new technology and review potential use cases in finance.
Examples of Generative AI Use in the Finance Sector: In the finance sector, companies can choose ready-to-use generative AI tools to apply to their datasets or develop custom tools. Customer service and more sophisticated chatbots, data analysis, coding, and assistance in hiring processes are among the many applications also found in other sectors that are being implemented in finance.
Specifically for the finance sector, the following use cases for models processing financial data can be cited:
Continuous monitoring to better detect fraud and crimes in financial systems.
Personalized financial recommendations and payment alerts.
Comprehensive financial analysis and forecasting.
Preparation and summarization of financial reports.
Informing for compliance with financial regulations.
Providing analyses to improve portfolio and investment risk management.
Below are examples of some major financial institutions working on using generative AI:
BloombergGPT -> A large language model specifically developed for the finance sector, containing 50 billion parameters. It is said to successfully perform sentiment analysis, news categorization, and other financial transactions and has passed tests.
Morgan Stanley -> Utilizes chatbots powered by OpenAI to enable asset management advisors to give more effective advice.
JPMorgan -> Working on a ChatGPT-style AI investment advisor that can make investment choices for its clients and is trying to trademark a program called IndexGPT for financial securities analysis and selection.
Citadel -> This Chicago-based fund is considering acquiring a ChatGPT license for corporate-wide software development and data analysis.
CHALLENGES AND RISKS OF USING ARTIFICIAL INTELLIGENCE IN THE LEGAL FIELD
Challenges and Risks in the Legal Field Regarding Artificial Intelligence
Specific Problems and Challenges of Artificial Intelligence
The use of machine learning (ML) techniques presents specific legal problems and challenges, particularly in determining responsibilities. Addressing these legal issues and challenges is not optional but necessary due to the serious legal and regulatory consequences that can arise from failures. While problems and challenges may vary according to specific applications, many of them arise from the following characteristics of machine learning (ML) techniques:
Trust in Training Data: Unlike rule-based algorithms, machine learning (ML) algorithms are dynamic, and their outcomes depend on the quality of the data they are trained on. This data can be collected from various sources and used over a certain period. This situation necessitates the creation of new processes and controls to maintain data quality at acceptable levels.
Predictability: While the outcomes of rule-based algorithms are predetermined, machine learning algorithms are designed to achieve a certain level of accuracy and can produce different results with the same inputs over time as the model is retrained with new data. These characteristics will conflict with regulations that require strict compliance standards or demand consistent results.
Explainability: In some models using more advanced techniques, outputs may not be explainable as a function of the inputs. Some experts have identified a balance between efficiency and explainability. Although reverse engineering methods are sometimes used to understand the properties of algorithms known as black boxes, these methods do not provide full transparency.
Specific Risks of Artificial Intelligence
There are specific risks associated with the use of artificial intelligence, and these risks can lead to financial losses, reputational damage, and legal sanctions as the use of generative AI increases.
Employee Trials: Now that generative AI is widely available, trials by employees with these tools can further increase existing risks.
Unreliable Outputs: As acknowledged by OpenAI, ChatGPT can produce sentences that are syntactically correct but semantically incorrect, or fabricate non-existent information. This situation creates risks of statements that seem correct but are false and the potential for malicious use of incorrect information, such as harassment, defamation, or the spread of misinformation and fake news, which can have serious effects on market confidence.
Information Limitations: Upon release, ChatGPT's knowledge base was loaded up to September 2021. Therefore, it is not aware of more recent developments and carries the risk of producing outputs with outdated information.
Bias and Discrimination: Considering the size and diversity of the data sets that generative AI can be trained on and the fact that it independently produces content, the potential for bias and discrimination in the outputs of generative AI may be more difficult to manage than in traditional AI.
Copyright Infringement: The use of copyrighted training data, illegal copying of copyrighted works for training, and reproduction of significant portions of copyrighted works in outputs can lead to potential copyright infringement.
Misuse of Personal Data: There are challenging questions about whether the use of publicly available data in training these models is legal and whether the 'fictional' outputs created by Large Language Models (LLMs) comply with the accuracy principle stated in the GDPR. Similarly, fulfilling requests from individuals to block or delete data can create technically challenging situations if that personal data has become embedded in the model."
