How costly are cultural biases? Evidence from FinTech

GAICIL Insights

The standard script says algorithms import human prejudice.¹ This paper flips it: an algorithm exposes prejudice and makes people richer for abandoning it. How costly are cultural biases? Evidence from FinTech, by Francesco D'Acunto, Pulak Ghosh and Alberto G. Rossi, tackles a notoriously hard measurement problem, the cost of cultural bias in high-stakes economic choices. The authors compare peer-to-peer loans that the same lenders make alone and after observing robo-advised suggestions.

The numbers sting. Unassisted, lenders are more likely to choose co-ethnic borrowers, facing 8% higher defaults and 7.3pp lower returns. Robo-advising leaves diversification unchanged and reduces lending to high-risk co-ethnic borrowers. Lenders in locations with high inter-ethnic animus drive the results, even when borrowers reside elsewhere. Biased beliefs explain the patterns better than a conscious taste for discrimination: lenders rarely override robo-advised matches to ethnicities they discriminated against when unassisted.

The conclusion has a sharp edge. In-group versus out-group discrimination and stereotypical discrimination are prevalent and economically sizable. They make discriminating lenders worse off in consumption utility, because by discriminating they finance loans to borrowers who perform worse than the discriminated borrowers available on the platform. Prejudice turns out to be a bad investment strategy, and the market charges for it.

The same-person design is what makes the evidence bite. Debates about artificial intelligence often focus on algorithms embedding bias. Here a robo-advisor reduces the effects of human bias, improving outcomes for the very people whose choices were distorted. Comparing assisted and unassisted choices by the same individuals removes the usual worry that different people simply chose differently. The algorithm acts like a mirror held up to each lender.

The override result is the psychological twist. If discrimination stemmed from a conscious taste, lenders would reject the algorithm's suggestions. They largely accept them, which points to biased beliefs. Algorithms can correct beliefs by supplying better information, and people appear willing to act on it when it is offered. Many biased lenders are misinformed, and misinformation has a cure.

The authors propose a new research role for robo-advising tools: benchmarks for the types and sizes of biases in decision-making. Coding tools that embed different biases or rules of thumb and comparing unassisted choices with choices after accessing them would let researchers disentangle and quantify alternative biases. AI becomes an instrument for understanding human behavior as well as improving it, a kind of controlled experiment built into a product.

The learning question is next. Future research should study whether exposure to robo-advising lets decision-makers learn optimal choices and develop rules of thumb that help when no robo-advisor is available. Interactive tools might teach goal-setting strategies or provide just-in-time financial literacy. If algorithms educate as well as advise, their benefits extend beyond the moments of use.

For policy, well-designed algorithmic tools can reduce the costs of discrimination in financial markets. Regulators concerned with fairness should evaluate algorithms against the human decisions they replace or assist, measuring outcomes for decision-makers and those affected. In this setting the algorithm reduced biased lending and improved returns, a result that should shape how policy treats such tools. Judging an algorithm against perfection, and ignoring the human baseline, can entrench the more biased option.

Digital platforms double as laboratories. Peer-to-peer lending generates detailed records of choices and outcomes, and robo-advice creates a natural comparison within the same individuals. Similar opportunities exist wherever algorithmic tools enter human decision processes.

Read as a whole, the lesson concerns human and machine decision-making together. The authors call for work across fields on how human and machine-based decisions interact, complement or substitute for each other, as the combination becomes ubiquitous in daily economic decisions. This study offers a template, and evidence that algorithms can help people overcome their own biases, to their benefit and that of others.

Something for fair-lending regulators to sit with: the question to ask of an algorithm is whether it beats the humans it assists. Sometimes the machine is the most effective remedy for discrimination on the table.