AI EDUCATION: What Is Causal AI?

Which is fine, because the questions we ask of generative AI are always what- or how-type questions. “What is causal AI?”, for example—which is what we’re going to be talking about—or “How do I write a riveting column about causal AI with a great introduction that just sucks everyone in?” 

There is significant demand for intelligence that can tell us why things happen, or why they might happen in the future, and that’s what we’re going to call “causal AI.” We also want to talk about why causal AI is an important step in the development of artificial intelligence, and that’s going to take us back to one of the early AI Education columns we wrote for this newsletter on explainable AI. 

Think of explainable AI as kind of a cousin to causal AI. Explainable AI is artificial intelligence that can not only give us the answers to our what or how questions, that can not only generate content like images and blocks of relevant and well-written text, but can also tell us how and why it made every decision along the way that led to the content it generated. Why it included some information, but excluded other information. Why it shaped a sentence or a feature of an image one way but not another. As generative AI continues to make inroads into regulated spaces, explainability remains a significant concern. Causal AI is significantly different, but important for similar reasons. 

Correlation Versus Causality 

It’s one thing to look at data and make correlations—noticing correlations is just noticing that some sort of relationship might exist between groups of different data points. It doesn’t even mean a relationship exists.  Over the past two centuries, there’s been a strong correlation between rising global temperatures and declining piracy on the seas, as an internet meme points out—that doesn’t mean that climate change is somehow destroying global demand for piracy.  

Knowing a correlation exists doesn’t mean we know anything about the nature or direction of the relationship—A and B are clearly related, but we don’t know with any certainty if A is being caused by B or if B is being caused by A, or if there’s an external C that is influencing both of them simultaneously. This is something human beings, in our centuries of developing and honing science, struggle with all the time, largely because of limitations on our abilities to observe.  

We looked at the sky and thought that the Sun, Moon and stars revolved around Earth, and the Earth’s centrality in the Cosmos moved the universe, not the other way around, because it appeared so to us. More recently, when we discovered organic chemicals, we believed that they were a byproduct of living creatures and could only exist as the byproduct of life, and not the other way around—that life came to be some time after the natural formation of organic chemicals, and that organic compounds aren’t at all rare throughout our solar system and the observable universe.  

Finding correlations usually belongs in what Daniel Kahneman might call System 1 thinking—given datasets within certain sizes, our brains can do it with such ease it often feels automatic. Understanding causality, on the other hand, more often falls in the slower realm of System 2 thinking—it requires some active, analytical thought. We’ve trained our machines to be experts at system 1 thinking at a massive scale, AI as we commonly understand it works by finding the correlations in huge data sets and learning from them, and it performs the magic trick of appearing to understand very complex relationships based on what is actually mere pattern recognition.

Causality Comes to AI 

Yes, our technology resembles its creators. It already has been finding correlations for us for some time, and most generative AI—even most predictive AI—is, behind the scenes, trained on recognizing the correlations between words or numbers, not necessarily understanding the causal relationships between them. This is because it’s easier to create a machine that can do the massive work of ingesting data and learning about millions of correlations in hopes of some day, in its inference phase, coming up with a good match for what its users are looking for, than it is to create a system that understands the directionality involved in causation—it’s like adding a third face to a coin flip. 

Allowing a machine to understand causal relationships versus mere correlations permits new dimensions of reasoning—AI can be trained to identify root causes behind problems or events, or to understand counterfactual “what-if” scenarios. While explainable AI is concerned more with showing how the AI got from a cause, an input, to an effect, the model’s output, causal AI is concerned more with how possible or probable inputs—variables, if you will—can lead to probable outputs—or conclusions. This is AI that doesn’t just make predictions,  

This means that causal AI systems, beyond being explainable and more regulation-friendly by design, can also create autonomous troubleshooting and problem-solving workflows, where one system identifies and diagnoses a problem, proposes interventions, tests the interventions itself to identify which one has the highest probability of success, and then deploys. In other words, this is AI that doesn’t merely make predictions—it makes decisions. 

The gory details of causal AI are not just found in the algorithms and frameworks being used by developers to train more sophisticated models, which we’re not going to dive into today. They’re also found in the intellectual, statistical and structural frameworks around causal inference—which we’re also going to spare you from today. As we said above, we take for granted that much of what we call reasoning is really just Kahneman’s System 1 thinking leading us to quick, easy, and mostly right insights based on identifying correlations because that’s enough for us in our day-to-day lives. That’s not enough to determine causality, and it’s not enough for regulated industries. 

Applications of Causal AI 

Obviously, causal AI’s ability to not just help make decisions, but to actually make decisions themselves, has major implications for business in general. There are already brick-and-mortar stores being piloted that are managed entirely by artificial intelligence agents. In some cases, AI agents which employ human subordinates.  

In finance, causal AI can not only help identify and respond to fraud, it can be applied to fraud prevention. It can also be used to map out, test and plan for scenarios. 

In healthcare, causal AI can help give caregivers a better understanding of the causes and pathologies of disease, both in individual patients and in populations, and similar to potential financial applications, AI can be used to test different medical scenarios and interventions.    

In education, causal AI can help personalize lesson plans on a student-by-student basis instead of a classroom basis. Causal AI can also help model and select interventions. 

Causal AI can also be turned on AI itself to help identify potential bias in AI models, or understand where certain types of bias may come to be introduced into AI models.