Episode notes
You can train an AI on data containing no bias at all and still end up with a biased system. Understanding why is the difference between a working deployment and a reputational problem.
Alastair McDermott talks to AI ethics specialist Tom Murphy about the failures that do not announce themselves. Recorded December 2024.
They get into:
- What "alignment" actually means, stripped of the science fiction
- Proxies: how postcodes, titles and car colours smuggle protected traits back in
- How to test your own system for subgroup bias, concretely
- The trolley problem, four hundred data scientists, and what they could not agree on
- The cancer drug distribution dilemma, which nobody in the episode resolves
Listen now, wherever you get your podcasts.