My new Unite.AI article looks at what happens when AI learns from human decisions. Hiring records, medical records and everyday language all carry our blind spots, and a model trained on them applies those blind spots to every case it sees. The research in the article covers hiring, healthcare and workplace email.
Thank you to Unite.AI for publishing it. This is a short summary. The full article is here: “AI Doesn’t Eliminate Human Bias. It Scales It.”
Integration over deployment
The usual rollout picks a model, builds it into the workflow and monitors performance. The article proposes integration over deployment: the accountable person works inside the decision, with real authority to change it. Three steps start it. Give every decision the AI screens, scores or approves a named human owner. Audit the history that feeds the model before it becomes policy. Make overriding the model safe, and track the override rate.
What matters is whether that person has the authority, the information and the psychological safety to say, “The machine is wrong.”
What the research shows
The article draws on a Stanford analysis of 4 million job applications, University of Washington testing of more than 550 resumes, Johns Hopkins research on workplace emails, and a Science study of a health care algorithm that used spending as a stand-in for medical need. Each shows a system carrying forward the assumptions built into its data and its metrics.
Read the full article
The article also covers three tests any team can run and the limits of technical fixes. Fairness belongs in the same meeting as accuracy, with people in the room who have been misjudged by systems like these. That idea runs through my book, The Bias Advantage.
Read the full article: AI Doesn’t Eliminate Human Bias. It Scales It. by Liat Ben-Zur, first published on Unite.AI.
