Olivier Toubia is a Columbia Business School professor working at the frontier of AI, marketing, consumer behaviour, and digital twins. This was recorded late last year but his arguments are relevant and clear: LLMs drive towards the average which means its hard to mimic the diversity of human existence. 

• Digital twins are promising — but not ready. Synthetic consumers for research purposes (e.g. virtual focus groups) are fast and cheap, but Toubia’s research suggests they are still too biased and inconsistent to trust too far. Learn what happened when they compared digital and human twins' responses to the same questions.

• Politics is the worst domain to deploy synthetic focus groups: The base model exerts a powerful pull. Twins often reflect the worldview of the model as much as the individual they are meant to represent, especially in politics, where performance is weakest. AI is too sensible - compared to humans it tends to be more moderate: in one study 45% of humans wanted to deport illegal immigrants and only 5% of digital twins agreed.

• Synthetic personalities reflect a clear pro-tech bias. Twins tend to be more comfortable with algorithms, less concerned about privacy, and more trusting of technology than humans actually are. AI twins thinks AI is smarter than humans....

• Generative AI has entered its monetisation phase with advertising as the likely endgame Subscriptions and enterprise sales matter, but the bigger prize is becoming the interface for search, discovery, decision-making, and purchase.

• Business adoption remains uneven but labour anxiety is rational Many pilots are failing to deliver meaningful returns; senior leaders are often enthusiastic, while middle managers and employees are more sceptical.

• AI can be creative, but not at the edges. AI tools can recombine ideas - "its a fallacy to think that AI will never be able create new things". However, it still misses the diversity of lived experience that gives humans their edge. The push is towards the average.

We usually worry about AI driving polarisation - Olivier's work shows that current LLM tools trend towards the opposite. Humans can provide the spikiness and AI the padding.