Human Oversight and Explainability in AI-Assisted Credit Decisions: A Survey Experiment among Banking Professionals

The intervention consists of a 2×2 between-subject randomized survey experiment embedded in the questionnaire. All respondents are presented with the same hypothetical scenario in which a customer applies for a €20,000 five-year personal loan and an artificial intelligence system, after assessing standard financial information, recommends rejecting the application. Respondents are then randomly assigned to one of four experimental conditions that vary independently along two dimensions: human oversight and explainability. In the human-oversight condition, a qualified professional reviews the AI recommendation and the underlying information before the final decision and may confirm or modify it; in the no-oversight condition, no such review takes place. In the explanation condition, the AI system provides information on the factors that contributed positively or negatively to its recommendation, whereas in the no-explanation condition no information is provided about the factors underlying the recommendation. The four treatment arms therefore allow us to estimate the separate and joint effects of human oversight and explainability on respondents’ reactions to AI-assisted credit decisions.

2026-10-01

2026-10-31

- (1) Trust: “Thinking about the procedure you have just read, how much trust would a final decision on this application, reached through this procedure, generate in you?” (0 = no trust at all, 10 = complete trust).
- (2) Delegated authority: “What maximum weight would you be willing to allow the AI system’s recommendation to have in the final decision?” (0% = no weight, 100% = fully determines the decision).

- (3) Procedural acceptance: “To what extent would you consider it appropriate for your institution to use a procedure like the one described to make credit decisions?” (0–10).
- (4) Perceived accountability/control (0–10).
- (5) Perceived transparency/understanding (0–10).

We conduct a between-subject 2 x 2 factorial survey experiment embedded in an online questionnaire administered to employees of the Spanish banking sector. Two binary factors are randomized independently at the individual level. The first factor is Human Oversight: before the final decision is adopted, a qualified professional either reviews the AI recommendation and the information used by the system and may confirm or modify the recommendation, or no such professional review occurs. The second factor is Explanation: the AI either provides a concise explanation identifying the factors that most influenced its recommendation and the direction of those influences, or provides no information about which factors influenced the recommendation. All respondents read the same underlying loan-application vignette and receive the same AI recommendation to decline the application. The four experimental conditions differ only in the two randomized procedural features. Outcomes are elicited immediately after treatment exposure, followed by manipulation checks.

- Common scenario
A person applies to the respondent's institution for a EUR 20,000 personal loan with a five-year maturity. The bank uses an AI system that analyzes the economic and financial information normally used in this type of decision. The applicant has stable employment, net monthly income of EUR 2,400, no payment defaults during the previous three years, existing financial obligations of EUR 850 per month, and approximately 70% utilization of available credit-card limits. After analyzing this information, the AI system recommends declining the credit application.

Treatment manipulations
- Human oversight - no: the AI recommendation is not reviewed by a qualified professional before the final decision is adopted.
- Human oversight - yes: before the final decision is adopted, a qualified professional reviews the AI recommendation and the information used by the system and may confirm or modify the recommendation.
- Explanation - no: the system communicates its recommendation but provides no information on which factors influenced it or in which direction.
- Explanation - yes: together with its recommendation, the system states that employment stability and the absence of recent defaults influenced the recommendation positively, while current indebtedness and high utilization of available credit influenced it negatively, with the negative factors receiving greater overall weight.

Not available

Computer-based random assignment implemented in the online survey platform by the survey provider. Human oversight and explanation are independently randomized with equal probability at first exposure to the experimental module.

Individual respondent

No

-

1,000 employees of the banking sector

250 employees of the banking sector per treatment arm

The planned sample is approximately N = 1,000, with equal allocation across the four treatment cells (approximately 250 respondents per cell). For a continuous 0-10 outcome with standard deviation 2.5, a two-sided test with alpha = 0.05 and 80% power implies the following approximate minimum detectable effects (MDEs).