Dr. Scott Zoldi's advice to others on their use of artificial intelligence: Don't get comfortable.
"I've been in this industry and building AI models for a long time — 25 years or more — and about seven years ago, I got really concerned that people are getting too comfortable with AI," said Zoldi, a former theoretical physicist and the chief analytics officer at FICO.
"Twenty-five years ago, you had to go through quite the set of discussions to get anyone to adopt a machine-learning model, whereas today, it seems many people don't even ask the basic sort of questions," he said. "Mistakes are being made."
A major issue is that the new adopters of AI, drawn in by the current hype cycle, are using models without being able to explain how they work. This creates a clear risk of introducing or perpetuating biases that can affect customers, Zoldi warns.
Zoldi, who holds a doctorate in theoretical and computational physics from Duke University, submitted two patents within the past year to address these concerns. (Over his career, Zoldi has written or submitted 138 patents, of which 96 were granted and the rest remain pending).
The first of these new patents, which has been granted, codifies AI model development in a blockchain to create an immutable record of who performed, tested and verified the work. "It's not just a checklist," he said. "It shows when mistakes are made and when things were rejected and when things were remediated … it provides transparency, which is of critical importance."
The second patent, which is pending, is also focused on transparency — in this case, ensuring that the person who built the model documents how it was done, so that they (or someone else) can look at it years later and understand what to monitor about it.
To encourage future generations of experts, Zoldi was the executive sponsor of the first FICO Educational Analytics Challenge, which works with several historically Black colleges and universities to train students to work with complicated data. The first year's focus was on identifying and mitigating bias in data collected by the Consumer Financial Protection Bureau. The upcoming year's focus will likely be on payment fraud data, he said.
