If ‘All Models Are Wrong,’ Why Do We Give Them So Much Power?
About this episode
Klein asks Brian Christian to explain the 'alignment problem' — the difficulty of getting machine-learning systems to pursue the goals their builders actually intend, a concept Christian traces back to economics. They walk through present-day cases: Amazon scrapped a resume tool that downgraded women's applications, crime-risk models predict arrest rather than crime, and a self-driving Uber killed a pedestrian partly because its training data contained no jaywalkers. They discuss why inscrutable deep-learning models are hard to audit, how AI assistants built on advertising business models may serve their makers over users, and whether recommendation feeds and targeted ads can coexist with human autonomy. Christian also covers what AI research borrowed from human psychology — curiosity, dopamine-driven learning — the ethics of potentially suffering reinforcement-learning agents, and what automation means for jobs, dignity, and status.