Abstract

In the age of rapid AI advancements, the widespread adoption of LLMs has particularly revolutionized a wide array of applications in our society. As these models have evolved into complex reasoning systems, concerns have been raised regarding their embedded biases. Such bias can reinforce stereotypes and perpetuate existing inequalities, especially when models generate unfair outcomes in decision-making contexts. This creates numerous ethical and legal issues, particularly in highly regulated industries like the banking sector.

This paper investigates the presence of bias in LLMs within the banking industry, using synthetic decision-making scenarios. By utilizing concepts concerning social bias in LLMs and counterfactual fairness, the study explores how decision outcomes differ when demographic variables (ethnicity, age, gender) are varied within the same scenario. The research adopts a mixed methods design, beginning with qualitative interviews, and followed by an experiment where prompts in Swedish were developed and tested to statistically measure bias. The evaluation included both explicit experiments, where demographic attributes were directly varied, and implicit experiments, where names revealed demographic proxies. The findings demonstrate the existence of bias in LLMs, most notably in the explicit tests, yet persisting in the implicit proxies. While certain results counteracted existing social inequities, others reinforced them. Nonetheless, all variations in the experimental results violate the principles of individual, group, and counterfactual fairness. This indicates the importance of adhering to the standards of fairness, accountability, and transparency in order to ethically deploy LLMs within banking.