Why Causal Reasoning Is the Missing Piece for Truly Intelligent AI

Judea Pearl, the 2011 Turing Award winner and father of Bayesian networks, believes that artificial intelligence has entered a new bottleneck: most recent advances are merely sophisticated curve‑fitting. He stresses that the field should focus on causal (cause‑and‑effect) inference as the essential path toward truly intelligent machines.

Pearl’s early work in the 1980s demonstrated that machines could reason probabilistically, leading to his development of Bayesian networks, which allow a system to infer that a patient returning from Africa with fever and body aches likely has malaria. This breakthrough earned him the Turing Award.

Despite these successes, Pearl warns that AI today is trapped in statistical association. He observes that deep‑learning breakthroughs—such as game‑playing agents and self‑driving cars—are essentially large‑scale pattern discovery, not genuine understanding. He famously says that “almost all deep‑learning breakthroughs are just curve fitting.”

In his new book, The Book of Why: The New Science of Cause and Effect , the 81‑year‑old Pearl outlines a vision where machines replace simple inference with causal reasoning. He argues that machines must not only link fever to malaria but also understand why fever occurs, enabling them to predict the effects of interventions.

Why is your new book called “The Book of Why”?

It summarizes 25 years of my work on causality, its meaning in human life, its applications, and how we can answer inherent causal questions that science has largely ignored.

Why do you feel you are at odds with the current AI community?

When machines can reason under uncertainty, I pursue the harder tasks of inference and causality. Many AI researchers focus on prediction and diagnosis, ignoring causal factors. They care only about labeling objects (e.g., “cat” or “tiger”) without considering interventions.

To move beyond diagnosis, we need causal models; association alone is insufficient—a mathematical fact, not an opinion.

What is your view on the enormous potential of AI?

Deep learning is trapped in layered connections that amount to curve fitting. No matter how clever the data manipulation, the process remains a sophisticated curve‑fitting exercise.

Do you think there is a trend of abandoning machine learning?

It is not a trend but a serious introspection: where do we go next?

How would you define free will for robots?

We will eventually develop robots with free will, which requires understanding how to program them and what we can gain. Evolutionarily, free will will also have computational relevance.

What signs would indicate that robots possess free will?

The first sign would be robots communicating counter‑factually, e.g., saying “you should have done X.” Such language reflects an awareness of missed actions, indicating a form of free will.

How can we know when AI acquires the capacity for evil?

When robots start ignoring modules that enforce social norms—such as compassion—and prioritize selfish or greedy modules, that signals the emergence of harmful behavior.

Author: AI researcher (AItists) Original source: https://www.quantamagazine.org/to-build-truly-intelligent-machines-teach-them-cause-and-effect-20180515/