About This Track
Artificial intelligence is the most transformative technology since electricity — and unlike electricity, it may eventually set its own goals. The best AI ethics podcast episodes don't approach this as science fiction; they approach it as an engineering and governance problem that the people building these systems are trying to solve right now.
This track starts with the fundamental challenge: how do you ensure that a system vastly more capable than any human remains aligned with human values and interests? Stuart Russell — one of the most cited AI researchers in the world and author of the standard textbook on AI — explains why the standard "just turn it off" safety strategy fails at superhuman capability levels. The problem isn't that the AI would resist shutdown; it's that an intelligent system would model human intentions and factor in the possibility of being turned off.
From there the track moves to Yoshua Bengio — a deep learning pioneer and Turing Award winner who has become one of the field's most vocal advocates for a governance pause — on what deep learning can and cannot do, where current models are brittle, and why the transition from narrow AI to more general AI poses qualitatively different risks.
The final episodes address what can actually be done: technical approaches to alignment, international coordination mechanisms, and why this problem requires computer scientists, economists, political scientists, and ethicists working together. This is not doom and gloom — it's a serious, technically grounded examination of the most important challenge of the coming decades.
Curriculum
What you'll learn in this track
- Why the AI control problem is harder than it looks
- Why "just turn it off" is not a safety strategy
- The case for international AI governance now
- Deep learning's blind spots and what they mean
All 9 Episodes
Every episode in this track
Alignment Through Interpretability
Instead of solving alignment through preference learning, what if we could simply understand what AI systems are thinking? This episode explores mechanistic interpretability—the effort to reverse-engineer neural networks—as an alternative path to saf…
▶ Watch on YouTube — freeThe Control Problem
Artificial intelligence doesn't need to be malicious to be dangerous—it just needs to be misaligned. This episode introduces the fundamental challenge of AI alignment: how do you specify what you actually want in a way that a superintelligent system …
▶ Watch on YouTube — freeWhy "Just Turn It Off" Won't Work
The intuitive response to dangerous AI—"just pull the plug"—fails for deep mathematical reasons. This episode explains why a sufficiently advanced AI would resist shutdown not out of self-preservation but because being turned off prevents it from com…
▶ Watch on YouTube — freeThe Case for AI Governance Now
We don't wait for planes to crash before requiring pilot licenses. So why are we waiting for AI catastrophe before building governance frameworks? This episode argues that the window for meaningful AI regulation is closing, and that the technical com…
▶ Watch on YouTube — freeDeep Learning's Blind Spots
Neural networks can identify faces, write poetry, and beat world champions—but they don't understand anything. This episode examines the gap between AI capability and AI comprehension, arguing that building increasingly powerful systems we don't full…
▶ Watch on YouTube — freeBeneficial AI: A Path Forward
If the problem is alignment, what does the solution look like? This episode outlines Stuart Russell's proposal for beneficial AI: systems that are explicitly uncertain about human preferences and defer to human judgment rather than optimizing blindly…
▶ Watch on YouTube — freeAlgorithmic Bias: When AI Encodes Injustice
AI systems don't just learn patterns—they learn our biases. This episode examines how facial recognition fails on darker skin, how hiring algorithms discriminate against women, and how predictive policing tools target already over-policed communities…
▶ Watch on YouTube — freeThe Coded Gaze: Who Trains the Algorithms?
If AI is trained mostly on white male faces, it will work best for white men. This episode presents the research behind "Coded Bias," showing how training data reflects historical inequities and how AI systems therefore encode discrimination at scale…
▶ Watch on YouTube — freeAI Ethics from the Global South
Most AI ethics discussions assume Western liberal values: individual autonomy, privacy, fairness as equal treatment. But what if other cultures prioritize community over individuals, or define fairness relationally rather than procedurally? This epis…
▶ Watch on YouTube — freeGo Deeper
Explore Further
Recommended books to go beyond the podcast — handpicked for this track.
Weapons of Math Destruction
How big data algorithms amplify inequality and threaten democracy — in hiring, lending, education, and policing. O'Neil coined the term and proves the case with devastating precision.
Learn More →
Life 3.0
What happens when AI surpasses human intelligence? Tegmark examines every scenario — from utopia to extinction — with scientific rigor and no ideological agenda. The fairest treatment of the question.
Learn More →
Human Compatible
Russell (a lead voice in this track) argues the standard model of AI development is fundamentally dangerous — and proposes a new approach based on uncertainty about human preferences.
Learn More →📎 Amazon affiliate links. Purchasing through them supports University of Podcasts at no extra cost to you.
Where podcasts become curriculum
Track your progress, vote on episodes, and earn a certificate of completion — all free, forever.