Predictive technologies: trained on the past, judging the future
from Slop Happens Hosted by Claudio Luís Vera and Bianca Prins
In a past life, Claudio ran a web design boutique that always had the Magic 8-Ball on the conference table, consulted with total mock-seriousness whenever a decision got stuck. Turns out that's not a bad starting point for evaluating predictive AI. So — are today's algorithms that much better than consulting the Magic 8-Ball? Not by as much as you'd think. A system running at 76% accuracy is only 26 points better than a coin flip. And that gap is exactly where things get real: it's the difference between who gets released before trial, who gets flagged as a hiring risk, who gets classified as a legitimate military target — and, in Allegheny County, Pennsylvania, which parents got to keep their kids. This week on Slop Happens, we also dig into an angle that doesn't get enough airtime: hiring algorithms that seek clones of a company's best performers will lock out disabled candidates, and produce different (and differently biased) results depending on whether the training data comes from the US or the EU. When the algorithm gets it wrong, no one ever gets fired for keeping silent. Which is why checks and balances on faulty AI are so important.
Embed this episode
No transcript yet.