Ethical Concerns in AI-Enabled Personal Finance Applications

Abstract

Artificial personal finance (PF) tools in individual financial decision making include budgeting automation, credit risk analysis, robo-advising, financial time-series modeling, purchase analysis and recommendations, and stand-alone financial education. More advanced PF automation may improve on one‘s speed and ease of decision making; however, newer automation also has the potential to spawn a slew of new ethical issues, such as inaccessibility to algorithms, data privacy, algorithmic bias, capricious user manipulation, loss of transparency and individual agency, and user degradation of trust and reliance.

To attempt solving some of these issues, this paper developed an integrative ethical-risk model of artificial PF applications using a transparent simulation data set creation process. In an artificial PF common-future use simulated scenario, a set of 480 conceptually modeled PF application user data points was generated of six multi-item concepts (algorithmic opacity, privacy concern, blinding oversight, apparent user manipulation, PF application trust, and PF application continuance intention). Cronbach‘s alpha reliability and multivariable regression were used to analyze the theoretically proposed interactions.

The data set demonstrated that algorithmic opacity, privacy concern, and perceived user manipulation were negatively associated with simulated PF application user PF application trust, and that high human-override capacity was positively associated with simulated PF application user PF application trust in this simulated multi-concept PF application model. In the simulated PF application trust model, R2 was 38.0%, and in the simulated PF application continuance intention model, R2 was 45.7%.in the simulated PF application trust model, PF  

application trust was a significant positive predictor of continued use, while privacy concern and perceived user manipulation continued to have negative effects on continued use.

An empirically driven, ethically principled PF application model seems to be emerging, in which transparency, automated PF application with minimal data collection, on-demand contestability, PF application and user-centered human intervention, PF application fairness-assessment tools, and non-manipulating user interfaces could form the bones of an ethical-by-design Artificial Intelligence-enabled PF system. As simulated data were used instead of the actual behaviors and attributes of real PF application users, the resultant numerical findings provide analyzable relationships of the variables instead of population estimates; the multi-case models studied here constitute a reproducible foundation for future empirical research.