AI Safety and Alignment: Experts Discuss Safer AI on Science Friday

Episode snapshot

In this Science Friday episode, Flora Lichtman hosts two leading researchers, Dr. Andrea Lincoln and Dr. Vinod Vaikunthanathan, as they unravel the core ideas behind AI safety and alignment. The conversation traverses how alignment is defined in practice, the limits of interpretability and the concept of chain of thought, and the challenges of ensuring sandboxed guarantees transfer to the real world. The guests discuss debates as a mathematical research agenda, safety by design, and the incentives that shape model behavior, including the risk of misalignment despite apparent safety in evaluations. The episode also considers the long arc of slowing down scientific progress to build robust safety, and what it means to treat AI as a societal challenge rather than a purely technical one.

• Alignment is a term of art without crisp mathematical definition
• Interpretability and chain of thought have limitations for safety monitoring
• Debate and other safety research agendas aim to verify models’ outputs
• Safety by design and incentive design may be essential to prevent undesirable behavior
• There is a concern that rapid progress creates an arms race, prompting calls to slow the clock

Introduction and guests

The podcast features Flora Lichtman guiding a discussion with Dr. Andrea Lincoln, a computer science professor at Boston University, and Dr. Vinod Vaikunthanathan, a cryptographer at MIT. The guests position themselves as researchers working on the engineering of AI safety, distinct from hype and corporate narratives. They set up the central theme: alignment as a research program aimed at producing AI that is capable and reliable while pursuing goals that align with human intentions. Lincoln emphasizes that alignment does not have a single crisp mathematical definition, but rather a family of definitions and research directions. Vaikunthanathan adds a pragmatic angle by comparing safety in AI to cryptographic design, where safety by design might ultimately be more reliable than attempting to retrofit safety after the fact. The conversation anchors its exploration in practical questions about how far along the field is in making AI systems safer and controllable.