One of the great promises of artificial intelligence has always been objectivity. Machines, unlike people, do not have bad days. They do not favour candidates because they remind them of themselves. They do not become impatient, tired or distracted.
For financial services, that consistency is enormously valuable. Credit decisions, underwriting, servicing and customer communications all benefit when similar cases receive similar treatment. Customers expect fairness and regulators increasingly expect firms to demonstrate it.
Human beings possess qualities that remain difficult to codify such as judgment, experience and context. Occasionally even mercy when a rule needs to be bent for the greater or individual good.
We know that exceptional circumstances exist because life is rarely lived inside a spreadsheet. The best lending decisions have always combined evidence with professional judgement. Artificial intelligence does not replace that but rather it supports it.
That distinction matters because AI learns from the information we provide. If the data reflects historical bias, the model may faithfully reproduce it. The technology is not choosing to discriminate. It is recognising patterns in the examples it has been shown.
THE ROLE OF GOVERNANCE
There are well known examples of this. Recruitment systems have learned to favour candidates who resemble those hired previously. Image recognition systems have performed less effectively where training data lacked diversity. Left unchecked, AI can simply automate yesterday’s assumptions.
That is precisely why governance matters. The answer is not to avoid AI but to seek to design it properly. And we see this first hand. Many lending decisions are already governed by transparent rules where applicants must meet minimum legal requirements, affordability thresholds must be satisfied and documentation must be complete. These are objective decisions that technology has been handling successfully for years.
Artificial intelligence extends that capability rather than replacing it. It helps identify patterns, prioritise cases and surface insights that might otherwise be overlooked. The final responsibility, however, remains with people.
LESSONS LEARNED
That means carefully considering which data should and should not influence a decision. During experimentation, for example, we found that models given access to names and postcodes quickly began drawing conclusions about geography and ethnicity that had no place in responsible lending.
Those variables were removed, not because the technology had failed but because we had learned something about the importance of the data itself – we applied judgment and wisdom to consistently incorrect behaviour.
The industry now has sophisticated ways of monitoring these issues. Statistical measures can identify whether outcomes differ across groups, while frameworks such as IBM’s AI Fairness 360 allow organisations to test models for unintended bias before they reach production.
Perhaps the biggest shift AI brings is not automation but transparency because human decision making has always contained unconscious bias that is remarkably difficult to measure.
U.S. fair-lending law prohibits lenders from making credit determinations that disparately affect minority borrowers if those determinations are based on characteristics unrelated to creditworthiness.
Using an identification under this rule, we show risk-equivalent Latinx/Black borrowers pay significantly higher interest rates on GSE-securitized and FHA-insured loans, particularly in high-minority-share neighbourhoods. We estimate these rate differences cost minority borrowers over $450 million yearly.
THE OPPORTUNITY TO IMPROVE
AI gives us something we have rarely had before in the shape of an opportunity to inspect, test, challenge and continually improve how and why decisions are made.
That does not mean as a result that every answer becomes perfect but it does enable us to actively identify and manage bias rather than simply hope does not exist.
Ultimately, fairness has never been about removing people from the process. It has always been about ensuring similar circumstances receive similar treatment while allowing room for human judgement when exceptional situations arise.
Artificial intelligence is exceptionally good at the first of those tasks but, for all our faults, we human beings remain essential for the second.
The future is unlikely to belong to organisations that ask AI to replace judgement but it will belong to those that combine machine consistency with human wisdom, governance and accountability. That is the future of better decision making.
