Friedrich Nietzsche once said, “Time, space, and causality are only metaphors of knowledge with which we explain things to ourselves.”

Causality finds mention in multiple fields from Mathematics, Psychology, Genetics, and Epidemiology to Philosophy. We humans too primarily use causal reasoning as a tool to understand the world. Now, causality is also finding relevance in the field of AI.

Simply put, causality refers to the relationship between what we do (cause) and what happens as a result of it (effect). Researcher Bernhard Schölkopf, in a paper titled ‘Causality for Machine Learning’, states: “Despite recent successes, if we compare what machine learning can do to what animals accomplish, we observe that the former is rather bad at some crucial feats where animals excel.”

AI currently can do some pretty interesting things. Large language models (LLM) like GPT3 have shown exceptional capabilities in generating and summarising text, reasoning, writing poetry, fixing bugs in software, and more. 

However, such models also have their limitations. Even though they generate responses that appear to be intelligent and relevant, they do not have a deep understanding of the meaning or context of the text they generate.

This is because these models learn from patterns in data and do not have a natural understanding of the world. They cannot learn about causes and effects from data alone. Hence, providing machines the ability to learn like we humans do could be the next step in AI.

Can causality overcome the current limitations of AI?

While LLMs thrive on data, it is also one of their biggest limitations. “Deep learning models are trained on millions and millions of data points. But in most settings, we don’t actually have that kind of data,” Rohit Bhattacharya, assistant professor of computer science at Williams College, Massachusetts, told AIM.

“So I think fundamentally moving towards a paradigm where machines try to learn more like us and this is where causality comes into the picture. We don’t require millions of data points in order to do the things that we do. Maybe reducing the amount of data needed to make these machines function the way they do is one big aspect where causal reasoning can play a role,” he said.

Another big limitation of current machine learning, including deep learning, according to Turing awardee Yoshua Bengio is the ability to properly generalise to new settings, like new distributions – what we call ‘out of distribution generalisation’. And humans are very good at that.

“If you have a good causal model, you can generalise to new settings. Current LLMs are like encyclopaedic thieves, they’ve read everything, but don’t understand it as deeply as we do. So naturally, they’re not able to reason with that knowledge as consistently as humans are,” Bengio told AIM.

He believes there are good enough reasons to think that humans excel at that because they have causal models of the world. “The angle that I’m currently exploring is related to causal machine learning, where we introduce notions of causality in our neural nets,” Bengio said.

Causality and LLMs

The most recent big breakthrough in AI has been LLM-based ChatGPT. While it’s intriguing to imagine a ChatGPT with reasoning powers, teaching LLMs to fully understand causality is a challenging task. “With the current way that large language models are trained, it puzzles me how you would incorporate causal reasoning into the actual current training and building of these models.

“I feel we need a fundamental shift in the way we build these models because right now, we’re just making predictions and that seems to have some amount of success. However, in order to move away from that, you sort of have to build from scratch,” Bhattacharya said.

Causality and AI isn’t new

According to Bengio, a bunch of these older ideas have the potential to make a comeback in modern-day AI, like old wine in a new bottle. He believes causality could be transformative. Although discussions about causality in AI have been ongoing, there has been a significant surge in interest from the community in recent years.

Much of the credit for causality in AI goes to celebrated computer scientist Judea Pearl. He is of the opinion that for machines to be truly intelligent, we need to teach them cause and effect. “In the 80s, Pearl was working on Bayesian networks, which was touted to be the way forward, and AI.  And so, causal Bayesian networks have always been a big part of AI. In some sense, these kinds of models have been deployed for a long time with some success, especially in the medical setting,” Bhattacharya said.

In his book ‘The Book of Why: The New Science of Cause and Effect’, Pearl argues that causal reasoning could provide machines with human-level intelligence.