Implementation and Evaluation of an Anti-Fraud Prototype Based on Generative Artificial Intelligence for the Ecuadorian Financial Sector

Objectives: This study aims to develop a more economical and simpler alternative to the current anti-fraud systems in the popular and solidarity financial sector in Ecuador. Given the growing threat of banking fraud in Latin America and the regulatory requirements of the Superintendence of Popular and Solidarity Economy of Ecuador, a solution based on generative artificial intelligence (AI) is proposed to identify fraudulent transactions on mobile and web banking platforms. Theoretical Framework: The implementation of anti-fraud systems in the financial sector is a complex challenge due to the evolving nature of fraud and the limitations of current systems, such as Custodian360, Monitor Plus, and Sentinel, which are perceived by more than 50% of surveyed professionals as costly and complex to implement. In this context, artificial intelligence, and specifically generative language models like GPT, present an opportunity to develop more accessible and effective solutions. Method: A prototype was developed using the LLaMA3 GPT engine (8B parameters) and the Ollama framework, in addition to OpenAI models like ChatGPT 3.5 and 4. The methodology included a proof of concept in which a prompt was parameterized in Ollama, specifying fields of transactional frames in JSON format and defining anti-fraud rules in natural language. These rules were shared by Dycotein, enabling the identification of potentially fraudulent transactions. Results and Discussion: The results indicated that models with fewer parameters, such as LLaMA3 8B and ChatGPT 3.5, are not suitable for the effective implementation of anti-fraud systems due to their limited capacity to handle complex rules. However, ChatGPT 4 showed promising results with simple rules, suggesting that while they are not a complete replacement for traditional programmatic systems, generative AI models can complement existing solutions, especially in less complex scenarios. Research Implications: The findings of this study highlight the need for continued research and development of more advanced generative AI models for fraud detection, as current technologies are not yet fully effective for complex scenarios. Additionally, the importance of balancing cost and effectiveness in the implementation of anti-fraud systems is emphasized, especially in sectors with limited resources such as the popular and solidarity financial sector. Originality/Value: This article presents an innovative approach to the use of generative artificial intelligence for the development of anti-fraud systems in a financial sector with budgetary constraints. By exploring the potential of advanced models like ChatGPT 4, the research opens new avenues for the creation of more accessible and adaptable solutions, offering significant value to the industry and academia in a context of increasing financial fraud in Latin America.