Major Issues in Causal AI dismantled by a Quantum Physicist at Spotify

The Uprising of A Miraculous Magic - Causal AI

Throughout the history, humanity has been adapting and evolving relentlessly from an early primate to the dominant, powerful, and transformative force of the earth. In that journey, the Homo sapiens had been through four different industrial revolutions (IRs), driven by constant and incessant curiosity about the world. And recently, the transformative power of Large Language Models (LLMs) marks what many consider the Fifth Industrial Revolution – a subtle initiative fundamentally reshaping economic dynamics and individual livelihoods worldwide. Within this transformation, Causal Artificial Intelligence (AI) emerges as a breakthrough in the field of Machine Learning (ML), addressing what Pearl(1) criticized: “Machines’ lack of understanding of causal relations is perhaps the biggest roadblock to giving them human-level intelligence.” Overlooking such a revolutionized knowledge would be undoubtedly oversight.

About The Guest – A Former Quantum Physicist

Accompanied by Associate Professor Ciarán M. Gilligan-Lee, we will unravel key issues in this revolutionary concept in the second episode, season two of Causal Bandits podcast. Associate Professor Gilligan-Lee formerly studied quantum physics, which later on brought him to a transformative journey in the field of Causal AI, is now head of Causual Inference Lab and Senior Research Manager at Spotify. His recent work explores “Foundation Models for Causal Inference,” positioning him at the forefront of integrating modern AI with causal reasoning. His strong academic rigor and industry expertise makes him a perfect guest for this summary.

Core Causal Inference Concepts Discussed in The Episode

This episode discusses a wide range of topics, ranging from theoretical methods like synthesis control to practical challenges and applications as medical diagnosis. In the first part, Lee explained systhesis control as an intuitive method particularly useful when treatments have precise timing and when identifying confounder is challenging. He then introduced a new approach to go beyond traditional linear factor model assumption – using DAGs for identifying what assumptions are needed to detect untreated “similar units” to construct a valid counterfactual

In the next part, he shared about the Causal AI engineering complexities of confounder selection, methodology selection, and variance control at Spotify. He specifically highlighted that naively adding all condounders can worsen estimates. According to Lee, correlation rankings can completely flip when controlling for confounding, which makes naïve approach fail. Therefore, the engineer must choose an appropriate method: double machine learning, inverse propensity score, etc. to manage the uncertainty and get the best out of the data at hand.

While sharing about his previous experience at a healthcare startup, he addressed the major automated symptom-to-disease mapping issue in medical AI. The right approach Lee appraised is to use counterfactual reasoning: “what would happen to symptoms if we cured this disease?” This approach outperformed the traditional correlation approach and matched the performance of top 25% doctor.

Knowing The Cure Does Not Guarantee Being Treated

Conincidently, while travelling to Italy this summer, I met a Ph.D. student working in this exact field at University of Pisa. We had interesting conversations about how medical AI, or AI/ML more broadly, is essentially yet silently shifting how the world operates. Regardless of the excitement surrounding this new technological wave, my friend emphasized that building trust between humans and machines has now become a major challenge, which is also agreed by Gilligan-Lee. I learned that simply inventing a new technology that can theoretically diagnose and treat diseases rapidly does not guarantee its successful implementation. Instead, fostering trust in the new technology for the public, especially in such a sensitive domain as healthcare, is one of the top priority. We thus need to develop comprehensive and trustworthy frameworks such as sensitivity analysis, bounds analysis, etc. that can both integrate expert domain knowledge and enhance credibility.

From this particular podcast, I also came to understand that successful causal inference in industry demands careful consideration of underlying assumptions, appropriate methodology selection, and robust validation, rather than simply applying black-box solutions. As the saying goes, knowing the cure does not mean the disease will be treated. The doctor must also persuade the patient to participate in the treatment – that is an undisputable important task. If Only I Have A Time Machine

I had just completed a condensed yet informative course on Causal Inference and I realised how naïve I was previously. Structural Causal Models (SCM) are powerful tools for diagnosing the causal relationships between variables, but accompanied by causal graphs (DAGs), they can together guide the choice of identification methods (backdoor, IV, etc.), which is the most important phase in the entire causal inference pipeline. The strong correlation without a causal link between chest pain and diabetes demonstrated in the podcast is a perfect example of how I learned how a confounder like being overweight distorts the correlation analysis in a misleading way. By using SCM to reframe this as a counterfactual reasoning task – “what would happen if a specific disease was cured, given the presence of the symptoms?” – followed by a causal DAG that identifies being overweight as a fork, the true causal structure becomes clear. I wish I had more time to delve into the practical aspects of applying SCMs and DAGs for identification in a causal pipeline.

References

1. Pearl, Judea (2019) The book of why: the new science of cause and effect.