Getting rejected for a bank loan can derail a person’s financial life, limiting their ability to buy a home, fund an education, or start a business. To make lending fairer and more objective, financial institutions increasingly rely on artificial intelligence (AI) to screen applicants, while regulators push for strict transparency rules that force banks to explain how these automated systems make decisions.
New research published in The Journal of Social Psychology reveals an unexpected catch in how the public reacts when these lending algorithms discriminate against minority applicants. Not only do people feel less moral outrage when a machine discriminates compared to a human loan officer, but transparently explaining the historical roots of the AI’s bias actually makes the public even less angry.
Looking for a motive inside the machine
Matthias Forstmann and Isabel Sonder, psychologists at the University of Zurich, set out to investigate how people judge ethnic discrimination in the financial sector. Previous research in corporate hiring had already identified a phenomenon known as the algorithmic outrage deficit, which describes the tendency for people to feel less anger when a computer program discriminates than when a human does.
Psychological theory suggests this gap happens because of how human beings evaluate right and wrong. When we witness an unfair act, we automatically try to read the actor’s mind to figure out their underlying motives.
People naturally assume that a human decision-maker has feelings, personal beliefs, and the capacity for malice or prejudice. While observers recognize that an algorithm can perform tasks and make decisions, they do not view the software as having inner feelings or conscious attitudes, making it difficult to accuse a machine of acting out of hatred.
Because moral outrage is the primary emotion that drives citizens to boycott companies, demand policy changes, and hold institutions accountable, a lack of outrage can allow unfair systems to operate unchecked. Forstmann and Sonder wanted to see if this outrage deficit held true in high-stakes credit lending, and how public reactions might shift when a bank provides a transparent explanation for why the bias occurred.
Testing human and algorithmic loan officers
To test their predictions, the researchers recruited 310 participants, primarily university students in Switzerland, for an online experiment. Each participant read two fictional scenarios about a bank that used a two-stage process to review loan applications.
In the first stage of the process, applications were screened and given a rating from one to five stars, with only the highest-rated applications advancing to a team of credit specialists for final approval. In one scenario, a human bank employee handled this initial star rating, while in the other scenario, an AI system performed the exact same task.
Both scenarios revealed that an external audit had uncovered severe ethnic discrimination at the bank. The human employee and the AI system both systematically assigned fewer stars to applicants with foreign-sounding names than to those with domestic-sounding names, even when their financial qualifications were comparable.
To test the impact of transparency, the researchers split the participants into two groups. Half of the participants received no explanation for why the discrimination happened.
The other half were given a structural explanation tracing the bias back to historical inequalities. In the AI scenario, participants read that the software had been trained on decades of past lending data from an era when people with migration backgrounds earned lower incomes and were deemed less creditworthy.
In the human scenario, participants received a parallel explanation stating that the employee’s judgments were shaped by decades of working at the bank during those same historical periods of inequality. After reading each scenario, participants rated their feelings of anger, outrage, and disgust on a seven-point scale, and answered questions about whether the decision-maker held racist or prejudiced motives.
The transparency paradox
The experiment confirmed that the algorithmic outrage deficit extends to financial lending. Across the board, participants reported significantly less moral outrage when the AI system penalized applicants with foreign-sounding names (averaging 4.56 on the seven-point scale) than when the human employee did the exact same thing (averaging 5.81).
Tracing the psychological steps behind these ratings explained why the machine got a lighter moral pass. When an AI system rejected the applicants, participants were much less likely to believe the decision was driven by prejudiced motives, and this absence of perceived malice directly led to lower moral outrage.
When the researchers looked at the impact of transparency, they uncovered what they call a transparency paradox. Providing a structural explanation for the AI’s behavior caused moral outrage to drop from 4.96 (among those given no explanation) down to 4.16.
Telling participants that the algorithm had passively learned its bad habits from historical data made them even less inclined to view the system as having biased motives, which further dampened their anger. By contrast, offering a parallel historical excuse for the human bank employee only reduced outrage from 5.99 to 5.63, a minor shift that was not statistically meaningful.
The researchers suggest this difference occurs because people assume humans are capable of harboring real prejudice regardless of how they learned it. Even when an employee’s bias is framed as an unconscious habit picked up over decades, observers still expect a human being to know better and override those impulses.
The study also uncovered a few secondary twists in how people evaluate automated finance. Participants who reported having a higher baseline knowledge of how AI works showed an even wider outrage gap, feeling notably less anger toward the discriminating algorithm than participants who knew less about technology.
At the corporate level, the researchers measured whether participants felt angry at the bank’s management. Outrage toward the executives remained roughly the same regardless of whether a human or an AI committed the bias. The researchers speculate this may be partly because participants were already annoyed at management simply for handing moral decisions over to a machine before they even learned about the discrimination, which may have obscured the effect of learning about the bias itself.
What this means for AI regulation
These findings point to a major hurdle for policymakers who rely on transparency laws to police automated finance. The authors argue that forcing companies to disclose how their algorithms work might backfire if organizations explain away discriminatory outcomes as the passive byproduct of historical training data.
Instead of sparking public scrutiny and demands for reform, these data-focused explanations act as an excuse that lets the technology off the hook. To keep public accountability intact, the researchers suggest that communications around algorithmic failures should emphasize active human choices, such as the engineers who selected the flawed training data or the managers who failed to audit the software for fairness.
Clear legal frameworks that hold banks strictly liable for discriminatory outcomes may also be necessary to counter the public’s tendency to forgive biased software. The authors note several limitations to their study that open avenues for future research.
Because the participant pool consisted mostly of young, female, and politically left-leaning Swiss university students, broader studies are needed to see how older, more conservative, or less educated populations react, as well as how minority applicants who actually experience credit rejection firsthand view the issue. The researchers also only tested one specific type of transparency: blaming historical data.
Other ways of explaining algorithmic failures might yield very different emotional reactions from the public. Future experiments will need to test whether explanations that highlight human coding errors or corporate negligence actually succeed in raising moral outrage and holding financial institutions to account.
