909: Causal AI, with Dr. Robert Usazuwa Ness

Episode Summary

Executive Summary: The episode explains causal AI as the shift from predicting correlations to reasoning about interventions and counterfactuals. Robert Osazua-Ness argues that modern tools like PyTorch, Pyro, and PyWhy/DoWhy make causal modeling more practical by separating statistical machinery from causal assumptions, while LLMs can help propose graphs and code but still require careful validation.

Main Topics: Why causal AI matters beyond correlation: The conversation contrasts correlation-based machine learning with causal reasoning, emphasizing that real-world decisions require understanding what causes outcomes, not just what co-occurs with them. The ladder of causation: Pearl’s hierarchy—association, intervention, counterfactual—is used to explain what kinds of questions data can answer and what extra assumptions are needed for stronger causal claims. Bayesian networks, graphical models, and probabilistic programming: Robert traces modern causal AI back to Bayesian networks and probabilistic graphical models, showing how tools like Stan, PyMC, Pyro, and NumPyro operationalize causal inference. Causal workflow in practice: He outlines a practical workflow: define the causal question, encode assumptions as a DAG or structural model, test identifiability, estimate with an appropriate statistical method, then run sensitivity/refutation checks. LLMs as causal assistants and knowledge bases: The episode discusses research showing LLMs can help infer causal directions, generate DAGs, and synthesize domain knowledge, but they can hallucinate and omit crucial assumptions. Generative and multimodal causal AI: Robert describes emerging opportunities where causal structure could improve generative AI, especially for multimodal systems and simulated environments like video games.

Key Arguments: AI historically centered on correlation because traditional causal methods were rooted in classical statistics and scaled poorly to high-dimensional data. Causal inference is fundamentally about the data-generating process, not just the observed data; if the data are insufficient, no method can magically remove bias. Bayesian and causal approaches both inject assumptions, but Bayesian priors express uncertainty about parameters while causal assumptions encode relationships in the system. Modern libraries now make causal modeling more accessible by letting practitioners specify assumptions while delegating inference to software. The do-operator and intervention-based modeling let analysts estimate effects that randomized experiments would reveal, but only under explicit assumptions. LLMs can be useful causal copilots—suggesting DAGs, formalizing questions, and generating code—but they are not reliable enough to trust without checking. Counterfactual reasoning requires stronger assumptions than association or intervention because it asks what would have happened under a different action while conditioning on observed evidence. A robust causal workflow should include sensitivity analysis or refutation tests to see how conclusions change if assumptions are wrong. Domain knowledge remains essential; causal AI does not eliminate the need to understand the system being modeled. Causal structure can potentially improve generative AI by adding mechanistic constraints rather than relying only on surface-level pattern learning.

Data Points: Episode number: 909 - Super Data Science Podcast episode featuring Dr. Robert Osazua-Ness Book giveaway: 5 physical copies - Listener contest for commenting or resharing the LinkedIn post about the episode Contest deadline: Sunday, August 3rd - Deadline to enter the book giveaway mentioned at the start Causal hierarchy levels: 3 - Pearl’s ladder of causation: association, intervention, counterfactual Paper focus model: GPT-4.0 - Robert described experiments evaluating LLM causal reasoning, mainly on GPT-4.0 Audience alignment: Hands-on practitioners - Episode is framed as most relevant to data scientists, statisticians, and AI engineers Show notes link: superdatascience.com/909 - Provided at the end for episode materials and transcript

Pivotal Quotes: "a timely resource for building AI systems that generate and understand causal narratives" — Judea Pearl: Quoted by John Crohn as praise for Robert’s book Causal AI "the question is not what's true in the world, but the question is: how do we write algorithms that do what we think is happening in those people's heads or in those animals' heads?" — Robert Osazua-Ness: Explaining the cognitive science perspective on how humans and animals reason causally "causal inference doesn't care about what data you're stuck with" — Robert Osazua-Ness: Describing why causal conclusions require assumptions about the data-generating process, not just whatever data happen to be available

Implications: For practitioners, causal AI adds rigor to decision-making, experimentation, and policy analysis. For industry, it promises better product, medical, and game-optimization insights. For future AI, it suggests LLMs and generative models may become more reliable when grounded in explicit causal structure.