In a special episode of the INSEAD Knowledge podcast, we shine a spotlight on a sister podcast series, The Age of Intelligence. 

Hosted by Theodoros Evgeniou, Professor of Technology and Business at INSEAD, and Tim Gordon, co-founder of Best Practice AI, the series features insightful conversations with notable guests from a range of different fields. Its aim is to look at how AI is rebalancing our world – from disrupting national powers and influencing business competitiveness to impacting individual lives. 

In this episode, Evgeniou and Gordon speak with computer scientist and MIT professor Pattie Maes. Their discussion centres on Maes’ pioneering work in AI and her unique perspective that technology should be used to augment human intelligence, not replace it.

A pioneer of early recommendation systems back in the 1990s, Maes argues that we need to move away from the current race to achieve AGI (Artificial General Intelligence). Instead, AI must be engineered to improve various aspects of human existence, serving as a layer that promotes learning, critical thinking and creativity.

To achieve this, we need to overcome specific obstacles. The first is the crisis of trust vs. competence. Although foundational models have largely solved the ability to automate tasks and possess knowledge (the competence issue), the hype around AI means we are too trusting of systems that still hallucinate, carry implicit biases and suffer from sycophancy.

At the same time, there is also a danger that we use current AI tools, such as ChatGPT, to outsource our thinking, instead of enhance our learning. Maes insists that the AI interface must change to function more like a Socratic tutor. Instead of providing long, definitive answers, it should ask very pointed questions, thereby encouraging curiosity, critical thinking and greater engagement with the topic. 

Maes also raises concerns that constant reliance on AI for advice (on relationships, mental health, and so on) weakens human social networks. This is because AI is always available, never disagrees and "looks up to" the user. She suggests developing benchmarks to test how prosocial AI models are, specifically evaluating the extent to which they push people towards seeking out real human interaction rather than simply replacing this interaction.

We can only hope to develop a future where AI improves and enhances our lives by changing current benchmarks. Instead of simply assessing the competence of each new model, we should be asking how it interacts with people.

You can access the full episode of the podcast to hear more from Evgeniou and Gordon on the topics raised in conversation with Maes.