Topics / Alignment & Safety
Alignment & Safety Podcasts
9 shows and 4,650 episodes sit under Alignment & Safety.
Survey of how the top podcasts are covering AI alignment as a research problem and AI safety as a deployed-harm problem — Brain Inspired, Data Skeptic, Daily Paper Cast, Shell Game
Photo: Brecht Corbeel on Unsplash · Unsplash License
Swipe · Alignment & Safety only
Swipe through Alignment & Safety
Save time discovering your next favourite podcast by listening to just the best minute of an episode.
The alignment & safety slate · every cover a link
All 9 alignment & safety shows →
Next to this on the shelf
Subtopics to explore
12 subtopics
Before you press play
A listener's guide to Alignment & Safety podcasts
The interesting moment usually comes after a model has done something impressive. A host asks what would happen outside the demo, a researcher names the test that still fails, or somebody notices that “helpful” depends rather heavily on who gets to define it. On audio, those pauses matter. Good alignment and safety episodes make uncertainty audible instead of polishing it into a verdict.
The shows worth hearing approach the problem from different distances. Some stay close to benchmarks and system behaviour; others examine fairness in deployed algorithms or follow AI agents into ordinary life. Together, they make the subject feel less like one grand philosophical dispute and more like a series of difficult engineering and social choices.
Tests with teeth
Methods, metrics and experimental results arrive at a brisk pace on Daily Paper Cast, the pick if the research layer is what you want first. Its synthetic co-hosts follow a dependable question-answer-restatement rhythm, which helps when the paper itself is dense. Here the clearest way in is PAWBench: How Far Are We from Probabilistically Aligned World Modeling?. FrontierChallenge: Evaluating Scientific Workflow Completion widens the evaluation question, while PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents points toward systems whose behaviour changes over time.
The vocabulary is technical and the delivery slightly didactic. That is useful when you want the structure of a paper without a long preamble, though listeners looking for argument, doubt and human friction may want to pair it with one of the shows below.
Data Skeptic works at a more conversational pace. The host knows the machinery but keeps asking guests to translate it into plain stakes, often by proposing an analogy and letting them correct it. Fairness becomes concrete in Social Choice for Fair Recommendations. Interpretability gets its own careful treatment in Disentanglement and Interpretability in Recommender Systems, and Healthy Friction in Job Recommender Systems is a good reminder that a system’s smoothest outcome may not be its most desirable one.
Before the benchmark
Alignment arguments often hide assumptions about agency, intelligence and what counts as successful behaviour. Brain Inspired is where those assumptions get room to breathe. These are leisurely academic interviews, comfortable with technical language and long conceptual detours.
BI 241 Johannes Jaeger: Agency and the Cyborg Myth is an obvious bridge into the subject. BI 240 Cristopher Moore: Cognition and Computational Complexity and BI 233 Tom Griffiths: The Laws of Thought push further underneath it. The show rarely rushes toward a policy conclusion; its value is in making the premises feel less settled than they did an hour earlier.
That depth also marks the boundary with neighbouring topics. When the main question is how models arrive at answers, head to AI Reasoning; systems built to act across longer tasks sit under AI Agents. Stay here when the concern is how behaviour gets evaluated, constrained and made answerable to human intentions.
When systems meet people
Shell Game supplies the narrative counterweight. Its host-narrated investigations follow AI voice-clone agents operating on his behalf, with generated dialogue placed inside measured, often dry storytelling. The gap between an agent’s eager performance and the host’s quiet reaction does much of the critical work.
After The One-Human Unicorn, the sequence proceeds through titles as plainly ominous as Episode 3: This is Law and Episode 5: Winners Win. This is safety heard from inside an experiment rather than discussed from a lectern. The show trusts the listener to notice the unease without announcing it after every scene.
FT Tech Tonic is quieter still. It opens inside personal situations, lets guests speak at length and then lifts the individual story toward the larger question with restrained narration. Artificial intimacy: The day your chatbot dies is the sharpest starting point for the human consequences of dependence on an AI system. AI Labs: Are Anthropic really the good guys? turns toward the institutions making those systems and the claims they make about themselves.
For the faster news cycle, Daily Tech News Show brings two informed, conversational hosts who reason aloud and disagree gently. OpenAI Slows Down and Focuses on Safety - DTNS 5335 places safety inside the daily technology brief, while Anthropic is (almost) Not a Supply Chain Risk - DTNS 5342 shifts attention toward operational exposure. The tone is lighter than the subject, but never careless.
To step back and compare these concerns with the rest of the field, go up to the broader Artificial Intelligence page. Move across to Frontier Models for the lab and model race, or to AI Regulation when laws, enforcement and institutional authority become the main event. Alignment and safety stays closest to the awkward middle: what systems do, how anyone can know, and who carries the cost when the answer is wrong.
Where to start
Begin with Daily Paper Cast and PAWBench: How Far Are We from Probabilistically Aligned World Modeling? if you want an evaluation problem in technical form. Choose Data Skeptic and Disentanglement and Interpretability in Recommender Systems for a patient researcher interview with practical stakes. For narrative tension, start Shell Game at The One-Human Unicorn. If the human relationship with AI matters most, try FT Tech Tonic and Artificial intimacy: The day your chatbot dies.
Featured shows below are part of the PlayNext catalog. Updated weekly.
Featured shows
9 shows, described rather than listed
Each one comes with a short write-up and the role it plays in this guide.
Paul Middlebrooks runs three-hour conversations on how brains and learning machines diverge. The format produces alignment conversations that aren't soundbites — the answer to 'is this dangerous' is usually 'here's what would have to be true for that question to even be tractable.'
Short episodes that surface alignment papers in their original packaging — RLHF variants, mechanistic interpretability, jailbreak taxonomies, scalable oversight. Doesn't ask you to feel anything about it.
Kyle Polich's bench on fairness, evaluation, recommender harms, and explainability is exactly the literature the public alignment debate hand-waves through. Under-recognized in the AI-safety conversation.
Evan Ratliff turning his AI voice clone loose on banks, customer-service systems, and people in his life. The cleanest field-reporting entry in our corpus on what deployed AI is doing right now.
The 'Artificial intimacy' series — 'How to fall in love with AI,' 'A teenager's last conversation,' 'The AI therapist that ended a marriage,' 'The day your chatbot dies.' The most sustained narrative reporting in the corpus on deployed AI's downstream harms.
Lewis Goodall, Emily Maitlis, and Jon Sopel cover safety when it crashes into Westminster — usually as deepfakes, election interference, or political misuse. The framing is political, not technical, which is part of why it lands.
Where deployed-safety incidents land first. 'Anthropic's Move to Secure the World,' 'US DoD Officially Declares Anthropic a Supply Chain Risk,' 'Anthropic and OpenAI Draw a Red Line for the Pentagon.'
Long-form interviews with the people running the labs — #494 Jensen Huang, #490 'State of AI in 2026.' Treat as raw material, not as a guide; the editorial work is yours.
Reid Hoffman and Aria Finger work the optimist-with-receipts corner. Useful counterweight to the doom-or-utopia binary that dominates a lot of the discourse.
Where to start
4 routes in, one per reader
Paul Middlebrooks treats 'is this dangerous' as a research question with prerequisites, not as a podcast bit.
The 'Artificial intimacy' series is the most sustained narrative AI-safety journalism in our corpus.
Evan Ratliff's voice-clone season is observed AI-safety, not speculated.
British political journalism that takes AI seriously when it touches Westminster.
Glossary
Jargon you'll hear on Alignment & Safety podcasts
The research problem of making sure an AI system pursues goals its operators actually want, instead of optimizing a proxy that breaks down at scale. Distinct from safety.
The broader surface of deployed-AI harms — misuse, deepfakes, voice clones, jailbreaks, misinformation, downstream effects on people who didn't choose to interact with the system. Much of the public AI-safety conversation actually concerns this rather than alignment proper.
Reinforcement Learning from Human Feedback. The training step where humans rate model outputs to teach a model what helpful, honest, and harmless answers look like. The dominant alignment technique in production today.
Research that tries to understand what computations a model is actually performing internally — what neurons or circuits represent which concepts. The technical ground for many alignment claims.
An adversarial input crafted to make a model behave outside its intended constraints. Jailbreak taxonomies are an active alignment research area.