Here are three paradoxes, according to Virginia Dignum, my guest today: 1. The more capable AI becomes, the more it reveals the richness and complexity of human intelligence. 2. Less bias in AI does not necessarily create more justice. 3. The pursuit of artificial superintelligence may ultimately reveal that humanity's greatest intelligence is collective, not artificial.

We discuss the limits of computation, the dangers of confusing data with reality, why AI ethics often misses deeper social problems, and what it would mean to build technology that genuinely serves human flourishing rather than replacing it. Our conversation is grounded in the book “The AI Paradox”, authored by Virginia, who is a professor in Responsible Artificial Intelligence and the Director of the AI Policy Lab at Umeå University in Sweden.

Jonathan Schaeffer thinks we're building AI the wrong way.

While large language models have produced remarkable results, he argues that hallucinations, bias, and unreliability aren't bugs that can be fixed—they're consequences of the underlying architecture itself. In his view, LLMs are an important stepping stone, but not the path to the kind of AI we can truly trust.

We discuss whether current AI systems are "good enough," automation bias, AI regulation, data centers, environmental costs, and the race toward AGI. We also debate whether society should slow down long enough to put meaningful guardrails in place before deploying increasingly powerful AI systems at scale.

Jonathan Schaeffer is a Professor of Computing Science at the University of Alberta and a pioneer in artificial intelligence research.

My guest, Mona Sloane, author of Predicted: How AI Is Restructuring Social Life, argues that AI has become part of our social infrastructure. Its predictive systems increasingly shape how we work, find information, build relationships, and navigate society.

Mona worries that as prediction becomes embedded in more areas of life, we risk becoming less willing to deliberate, challenge assumptions, and shape our own futures. I push back on whether AI really should be understood as infrastructure and whether predictions made by AI are fundamentally different from the predictions humans have always made.

We also discuss democracy, power, regulation, and what happens when prediction becomes the dominant way of understanding the world.

Book: https://a.co/d/04GwwuFR

Science depends on more than just results. It depends on researchers asking questions, testing hypotheses, challenging assumptions, and scrutinizing evidence.

My guest, Emily Sullivan, Senior Lecturer in Philosophy of Science and AI at the University of Edinburgh, argues that AI is beginning to influence every stage of the scientific process—from deciding which questions get asked to how papers are written, reviewed, and published.

We discuss algorithmic monocultures, scientific de-skilling, AI-generated research, and whether the pressure to accelerate discovery risks undermining the very process that makes science reliable in the first place.

I'm sympathetic to the promise of AI in science. Emily is concerned that, if we're not careful, we may end up optimizing for scientific output at the expense of scientific inquiry itself.

My guest, Fabio Tollon, a postdoctoral researcher on the BRAID programme at the University of Edinburgh, argues that answering that question is more difficult than it first appears. Traditional theories of moral responsibility suggest that people should only be blamed for actions they understand and control. But AI systems seem to challenge both requirements.

We discuss responsibility gaps, the problem of many hands, whether AI developers are more like parents or engineers, and Fabio's distinction between moral responsibility and moral answerability. Along the way, we explore whether answerability can help us make sense of AI harms when blame is difficult to assign.

Aligning an AI traditionally looks like a matter of giving it rules to obey. But my guest, Gillian Hadfield, Professor of AI Alignment and Governance at Johns Hopkins University, thinks that’s the wrong approach. She argues that we need to think about what it means to have a normative capacity - an ability to categorize behavior as (un)acceptable in a given context by observing that context - and then think about what it would mean to give that capacity to an AI. Lots to dig into here, including especially our disagreement about whether she’s focused on an ethical normative capacity vs. a prudential normative capacity.

Technologist’s are racing to create AGI, artificial general intelligence. They also say we must align the AGI’s moral values with our own. But Professors Ariela Tubert and Justin Tiehen argue that’s impossible. Once you create an AGI, they say, you also give them the intellectual capacity needed for freedom, including the freedom to reject your given values. Originally aired in season 2.

The Thomson Reuters Foundation recently conducted a global survey and found that most companies are lagging in their attempts to govern AI. For me the most surprising stat is that 85% of companies don’t have any training on AI risks for their employees. I think that’s just insane. Today my guests are Antonio Zappulla, CEO of the Thomson Reuters Foundation, and Katie Fowler, Director of Responsible Business. We talk about how they conducted their research, their results, and what incentives there are for businesses to do better.

To get in touch with Thomson Reuters Foundation to participate in their next survey, please go to: https://www.trust.org/newsletter/

My guest, Carissa Véliz, is author of the new book “Prophecy: Prediction, Power, and the Fight for the Future, from Ancient Oracles to AI.” Her thesis is that when leaders in AI say things like “AI adoption is inevitable,” they’re not making a prediction, but rather giving us a command and attempting to legitimize their power. Is she right? Have a listen!

Chapter Six: Dream Teams for Ethical Nightmares

• Three Types of ENC Teams
• ENC Teams as Emergency Response
• Tools for Teams
• ENC Teams in Bloom

Chapter Seven: ENC: An Approach So Flexible It Makes Simone

• Biles Look Like C-3PO
• Hands Off!
• You Do You
• Marrying ENC to Existing Practices
• Folding Existing Resources into ENC Teams
• Folding ENC Teams into Existing Resources
• The Ethical Nightmare Challenge for... Everyone

Chapter 4: The Standard Approach to Responsible AI Is Crumbling

• The Standard Approach
• The Madness in the Method
• Turn That Smile Upside Down
• Cats and Tigers, Oh My!

Chapter 5: Why I Like Nightmares and You Should, Too

• The Power of Nightmares
• What Good Nightmares Look Like
• And Now the Moment You've Been Waiting For