Access to credit is fundamental to modern life. It opens doors to education, homeownership, and entrepreneurship, which together drive social mobility. And as the cost of living rises around the world, affordable and reliable credit has become a lifeline for many.
Yet, depending on where one lives, securing credit can still be a daunting and opaque process. Borrowers face complex requirements and lengthy assessments, while lenders grapple with evaluating risk in economies that are digitizing faster than ever. Compounding these challenges is a growing wave of financial scams that exploit the very technologies meant to improve security.
According to Sumsub, the number of deepfakes detected across industries in the Asia Pacific region rose more than fifteenfold between 2022–2023, reflecting a wider global trend. “Forced verification” tactics, where fraudsters use manipulated identities to bypass know-your-customer (KYC) checks, have become increasingly sophisticated and accessible. A 2024 Oliver Wyman report found that financial scams across Southeast Asia have been expanding by roughly 50% per year since 2020, with total losses estimated at around USD 5 billion.
These figures suggest that defensive measures are not keeping pace with evolving threats. As a result, borrowers may grow more cautious and withdraw from formal financial channels, while lenders tighten screening processes, sometimes excluding creditworthy applicants.
AI-driven credit evaluation and risk modeling
Speaking on November 12 at the Singapore FinTech Festival, FinVolution vice president Lei Chen said artificial intelligence, when applied responsibly, can enhance transparency, accuracy, and security in credit evaluation.
Chen noted out that an increasingly digital economy now generates a wealth of real-world data that can offer more accurate and equitable assessments than traditional credit scores alone.
Consider, for instance, a farmer on the outskirts of a provincial town who needs financing to purchase feed and tools ahead of the planting season but lacks a sufficient credit score due to limited banking history. Data points such as consistent utility payments, mobile transactions, or timely supplier settlements could demonstrate financial reliability more effectively than conventional scoring systems. Similar cases can be found among small business owners and gig workers whose income flows outside traditional banking structures but whose digital activities reflect responsible financial behavior.
But how might such data be accessed and utilized without compromising user privacy? FinVolution proposes the use of federated learning techniques. This approach distributes model training across multiple data sources, with computation performed locally and only aggregated insights shared. By keeping data anonymized and avoiding centralized storage, the method reduces security risks and strengthens privacy safeguards.
According to Chen, FinVolution applies this principle through what he described as a flywheel-like model. By enabling a secure flow of data-driven insights, the company can continuously refine the algorithms deployed across various applications, including credit scoring. Over time, these iterations may enhance KYC processes and improve product offerings, thereby attracting more users and ultimately creating a self-reinforcing cycle of optimization.
Approaches like the one Chen described may have a particularly strong impact in emerging markets such as Indonesia, which, according to DBS Bank, is home to the world’s fourth largest unbanked population. By incorporating nontraditional data, FinVolution’s models help “credit invisibles” gain fairer access to financing.
Building trust and transparency
Meanwhile, AI has unlocked multimodal verification capabilities that enable financial providers to authenticate applicants through a wider range of identifiers. This expands on the concept of two-factor authentication to include voice, facial, and behavioral recognition, among others. Chen expects such capabilities to strengthen verification processes and make fraud significantly harder to execute.
He emphasized, however, that advanced technology is not new in the credit business. Deep learning models already play a key role in detecting suspicious activity. But he cautioned that accuracy must be balanced with interpretability. “The black box nature of deep models poses a significant challenge,” he said. “Chain-of-thought and knowledge graphs can enhance interpretability by offering clear reasoning paths.”
Modern credit scoring systems, Chen explained, now combine several methodologies: unsupervised learning to identify hidden customer segments, data from credit bureaus for baseline evaluation, and large models integrating structured, textual, and visual data from alternative sources. These hybrid systems can improve accuracy and reduce bias, but they are not sufficient on their own to identify deepfakes.
To drive progress in this area, FinVolution has been a strong proponent of collaboration, actively engaging the fintech community through initiatives that promote transparency and responsible AI adoption. The company organizes an annual data science competition, with this year’s edition focusing on visual deepfake detection. It also partners with conference organizers such as the International Joint Conference on Artificial Intelligence and the Conference on Information and Knowledge Management to jointly advance the ethical use of AI in finance.
Ultimately, Chen envisions machines becoming extensions of human capability, and that reality may already be near. He noted that agentic AI applications are already emerging in areas such as customer service, where AI agents can now act as “digital twins” to support human staff. However, he does not foresee a future where humans are replaced entirely. Instead, Chen believes human expertise will remain essential in determining how AI models are applied meaningfully.
One facet of this, he said, involves aligning model type and size with the tasks they are best suited for. According to Chen, smaller AI models can deliver significant impact through stable and efficient performance in domain-specific use cases, while larger models, which excel in general understanding and reasoning, are better suited for exploratory or frontier applications.
To advance work in these areas, FinVolution operates two dedicated platforms: E-LADF, which focuses on large model development, and Zeta, which supports the building and testing of agentic AI applications. Both are designed to drive human-led innovation in financial technology.
Listed on the New York Stock Exchange, FinVolution has reportedly worked with more than 130 financial institutions and served over 36 million borrowers, facilitating more than USD 157 billion in cumulative loans. Founded in 2007, the company is headquartered in China and maintains operations in Indonesia, the Philippines, and Pakistan.
This article was published in partnership with FinVolution Group.