Best AI Research and AI Safety Podcasts in 2027

Tape recorder and research papers in a dim library under cool blue light

The 12 best AI research and AI safety podcasts for 2027, from Dwarkesh Podcast and MLST to 80,000 Hours, AXRP and the CSIS AI Policy Podcast.

AI / AI Education / AI podcasts

The best AI research and safety podcasts for 2027 are Dwarkesh Podcast and Machine Learning Street Talk for research interviews, the 80,000 Hours Podcast and AXRP for AI safety, and the CSIS AI Policy Podcast and ChinaTalk for policy and geopolitics. We picked 12 shows that take capability claims and risks seriously, grouped them into research, safety and policy, and checked each show's ratings, cadence and contacts on 19 November 2026. AXRP is the one irregular show on the list and is flagged as such. For each podcast we explain who it suits, how much time it asks for and who produces it, because several are made by labs, foundations or think tanks.

Key takeaways

  1. Start with one research show and one safety or policy show; the two angles correct each other.
  2. Dwarkesh Podcast reports about 1.8 million views and listens per episode, by far the largest audience in this niche.
  3. Lab-produced shows such as Google DeepMind: The Podcast explain research well but are not independent reviews of it.
  4. The safety shows disagree with each other on how large AI risk is. Hearing both the 80,000 Hours Podcast and Doom Debates' skeptical guests gives a fairer picture than either alone.

Which AI research and safety podcasts go deepest?

Ranked by how much a practitioner or technology leader gets out of each one. "Checked" names the source behind each number and when we looked.

Research interviews

  1. 1.48M YouTube subscribers · about 1.8M views and listens per episode (publisher) · 4.4 on Apple (640 US ratings) · about weekly

    For: Researchers, strategists and serious generalists who want the hardest questions put to AI leaders.

    Dwarkesh Patel describes his interviews as deeply researched, and the recent run spans Paul Christiano on near-term AI risk and an episode arguing that the shift that hit mathematics is coming for AI development. The publisher reports about 1.8 million views and listens per episode across platforms. Episodes are 45 to 100 minutes and work fully in audio.

    Checked: Apple Podcasts US page (4.4, 640 ratings), YouTube channel page and dwarkesh.com sponsor page (per-episode average across platforms), 19 Nov 2026

  2. 221K YouTube subscribers · 4.5 on Apple (103 US ratings) · about weekly, around 90 min

    For: Technical listeners who want long, skeptical discussions of reasoning, cognition and AI capability claims.

    Machine Learning Street Talk mixes AI research with cognitive science and philosophy of mind, with recent guests including Joscha Bach and Leo de Moura on proofs no human can read. Tim Scarfe and Keith Duggar push back on hype as a matter of format. Expect about 90 minutes an episode, roughly weekly.

    Checked: Apple Podcasts US page (4.5, 103 ratings), RSS feed, YouTube channel page and mlst.ai, 19 Nov 2026; no sponsor email published outside YouTube sign-in

  3. 45.6K YouTube subscribers · 4.5 on Apple (102 US ratings) · 2 to 3 episodes a week, often 90 min or more

    For: Builders and researchers who want long interviews on interpretability, biosecurity and frontier agents.

    Nathan Labenz's interviews often go deep on safety-adjacent research, such as Goodfire's interpretability work and DNA screening for biosecurity. The show also runs AI:AM news episodes. At two or three long episodes a week, it rewards selective listening by guest.

    Checked: Apple Podcasts US page (4.5, 102 ratings), RSS feed, YouTube channel page and the show's sponsor page, 19 Nov 2026

  4. 4.8 on Apple (197 US ratings) · roughly monthly, about 40 min · Google DeepMind YouTube: 943K subscribers

    For: Listeners who want accessible, well-produced explanations of frontier research from inside one lab.

    Hannah Fry's series explains Google DeepMind projects such as AlphaGenome 2 and AI watermarking in about 40 minutes. It publishes roughly monthly. Because the lab produces it, it is best used for understanding the research, not for judging it.

    Checked: Apple Podcasts US page (4.8, 197 ratings), RSS feed, YouTube channel page and deepmind.google podcast page, 19 Nov 2026; no business contact (lab-produced show)

  5. 3.9 on Apple (11 US ratings) · about weekly, mostly 8 to 12 min audio essays plus longer interviews

    For: Researchers and engineers following open models, post-training and benchmark debates.

    Nathan Lambert's feed is mostly 8 to 12 minute audio versions of his essays, with occasional long interviews. Recent pieces covered the benchmarker's dilemma around Artificial Analysis v5 and the state of open-source AI. Its short format makes it the easiest research show to keep up with.

    Checked: Apple Podcasts US page (3.9, 11 ratings), RSS feed and interconnects.ai, 19 Nov 2026

AI safety and risk

  1. 98.4K YouTube subscribers · 4.4 on Apple (375 US ratings) · about 2 episodes a week, 15 min briefs to 4-hour interviews

    For: People who want rigorous coverage of AI risk, governance and careers in AI safety.

    The 80,000 Hours Podcast now describes itself around AI and publishes about twice a week, from short briefs on new reports, such as a redacted METR report, to a four-and-a-half-hour interview with Holden Karnofsky on 39 projects to improve outcomes with AGI. A curated crash-course episode is the best entry point for new listeners.

    Checked: Apple Podcasts US page (4.4, 375 ratings), RSS feed, YouTube channel page and 80000hours.org/podcast, 19 Nov 2026

  2. 108K YouTube subscribers · 4.7 on Apple (111 US ratings) · about weekly, around 70 min

    For: Listeners interested in catastrophic risk, AI governance and the social effects of advanced AI.

    The Future of Life Institute is a nonprofit working to reduce catastrophic risk from powerful technologies, and its weekly podcast reflects that mission. Recent guests include Max Tegmark on losing control of AI agents and Sherry Turkle on artificial intimacy. Episodes run about 70 minutes.

    Checked: Apple Podcasts US page (4.7, 111 ratings), RSS feed, YouTube channel page and futureoflife.org, 19 Nov 2026; email is the owner address in the RSS feed

  3. 4.4 on Apple (9 US ratings) · irregular, 50 episodes, latest August 2026 · 2 to 2.5 hours

    For: Researchers and technical readers who want paper-level discussion of AI safety work.

    Daniel Filan talks with researchers about specific papers and how they might reduce existential risk. Recent episodes covered METR's time-horizon work with David Rein and the AI 2027 scenario with Eli Lifland. It publishes irregularly, with the latest episode in August 2026, so treat it as a library more than a feed.

    Checked: Apple Podcasts US page (4.4, 9 ratings) and RSS feed (latest episode 2 Aug 2026), 19 Nov 2026

  4. 149K YouTube subscribers · 4.1 on Apple (25 US ratings) · 1 to 2 episodes a week, about 90 min

    For: Listeners who want AI risk arguments tested in open debate between opposing views.

    Liron Shapira hosts debates and interviews about AI catastrophe risk, often with guests who disagree with him. Recent episodes included Robert Wright and a debate with a guest who argued there is zero chance of AI doom by 2050. Its 149K YouTube subscribers make it larger than most safety shows.

    Checked: Apple Podcasts US page (4.1, 25 ratings), RSS feed and YouTube channel page, 19 Nov 2026; email is the owner address in the RSS feed

Policy and geopolitics

  1. 4.8 on Apple (59 US ratings) · weekly on Thursdays, about 50 min

    For: Policy, legal and government-affairs professionals tracking AI regulation and national security.

    Aalok Mehta, director of the Wadhwani AI Center at CSIS, hosts a weekly Thursday show on AI regulation, economic effects and national security. Recent episodes covered METR's findings at Anthropic, securing model weights with Sella Nevo and the backlash against data centers. CSIS describes itself as a bipartisan think tank.

    Checked: Apple Podcasts US page (4.8, 59 ratings) and RSS feed, 19 Nov 2026; email is the owner address in the RSS feed

  2. 17.5K YouTube subscribers · 4.3 on Apple (23 US ratings) · 2 to 3 episodes a week, about 70 min

    For: Analysts and leaders who need to understand US-China competition in chips and AI.

    Jordan Schneider's ChinaTalk covers China, technology and US-China relations, and recent emergency episodes dealt with export-control changes and the Nvidia leadership crisis, including one with Dylan Patel. It publishes two or three episodes a week. No other show on this list covers the chip and export-control side as closely.

    Checked: Apple Podcasts US page (4.3, 23 ratings), RSS feed, YouTube channel page and chinatalk.media/about (advertising section), 19 Nov 2026

  3. 4.8 on Apple (1.8K US ratings) · every other Thursday, 45 to 60 min

    For: General listeners who want the societal and humane-technology angle on AI.

    Tristan Harris and Aza Raskin of the Center for Humane Technology publish every other Thursday, as part of the TED Audio Collective. Recent episodes featured Ben Buchanan on whether AI is a race or a trap and new measures of AI safety for children. With 1.8K US Apple ratings, it is the most-rated show in this niche.

    Checked: Apple Podcasts US page (4.8, 1,777 ratings) and RSS feed (email in show notes), 19 Nov 2026

Why listen to AI research and safety podcasts?

Research and safety podcasts explain where AI capability is heading and what could go wrong, months before those questions reach product roadmaps or regulation. Episodes on METR's time-horizon measurements, interpretability and securing model weights all appeared on these shows in 2026, well ahead of mainstream coverage.

They also give listeners the arguments behind headline claims. A news show reports that a lab says its model can do something; Machine Learning Street Talk or AXRP asks how that was measured and what it does not show. That habit is useful for anyone who has to judge a vendor's capability claim or a policy proposal.

For a written framework that connects risk to organizational controls, our AI governance framework and AI security guide pair well with these shows.

How should you balance safety, research and policy shows?

The three groups answer different questions. Research shows ask what AI can do now, safety shows ask what happens if capability keeps rising, and policy shows ask what governments will do about it. Listening to only one group gives a distorted picture.

The safety shows also span a wide range of views. The 80,000 Hours Podcast and AXRP take catastrophic risk seriously and treat it analytically, Doom Debates stages arguments with people who disagree, and Your Undivided Attention focuses on present-day social harms. Choosing one show from each end is a fair way to hear the debate.

A balanced weekly stack takes about two and a half hours: one Dwarkesh or MLST episode, two short 80,000 Hours briefs or one interview, and the CSIS AI Policy Podcast on Thursdays. Add ChinaTalk when export controls or chip supply are in the news.

How did we choose these research and safety shows?

Each show had to cover AI research, risk or policy as its main subject and take evidence seriously. We excluded general science podcasts that only occasionally touch AI, and shows with no new episodes in the last 90 days, with one exception: AXRP, which publishes irregularly but has no substitute for paper-level safety discussion. Its listing says so plainly.

Audience numbers vary widely and do not track quality. Your Undivided Attention has 1.8K US Apple ratings, while AXRP has nine; Dwarkesh Podcast reports about 1.8 million views and listens per episode, while Interconnects is mostly short audio essays with a small rating count. We report each figure as the show or platform publishes it, with the date checked.

Contacts follow the same rule. Where a show publishes a business or feedback address, on its site, in its show notes or as the owner address in its RSS feed, we list it; lab-produced shows such as Google DeepMind: The Podcast publish none.

Listed here, or should be?

If an entry is out of date or something belongs on this list, email [email protected]. Listings are editorial and free. A featured, labeled Sponsored slot is available on request.

Related questions

What is the best AI safety podcast?

The 80,000 Hours Podcast is the most active AI safety show, publishing about twice a week with both short briefs and long interviews. AXRP goes deeper into individual safety papers but publishes irregularly.

Which AI podcast is best for understanding research?

Machine Learning Street Talk is the most technical, with long conversations about reasoning and cognition. Dwarkesh Podcast has the most prominent guests, and Interconnects offers short audio essays on open models and benchmarks.

Is there a podcast about AI policy?

The AI Policy Podcast from CSIS publishes every Thursday on regulation, national security and economic effects. ChinaTalk covers US-China competition, export controls and chips.

Explore More

Ready to Find the Right AI Tools?

Browse our data-driven rankings to find the best AI tools for your team.