Here is something that should bother you more than it does.
You can pay for the best AI model on the market right now, the one that writes cleaner emails than most people, passes the bar exam, debugs code, and sounds for all the world like it understands you. Ask it a question that hinges on the word “why,” and underneath all that fluency it is doing one thing and one thing only. It is guessing what comes next based on patterns it has seen before. It is not reasoning about cause. It does not know why anything happens. It has never known.
That is not a knock on the technology. I use these tools every day and they are genuinely remarkable. It is a statement about what kind of machine it actually is, and a man named Judea Pearl saw this coming decades before ChatGPT existed. He wrote the math that explains exactly where the ceiling is, and why no amount of bigger models and more data will get you through it on their own.
I want to walk through what he figured out, because once you see it, you cannot unsee it. And it changes how you should think about every AI decision you are about to make in your business.
Who Judea Pearl is, and why you should care
Judea Pearl is a computer scientist at UCLA. In 2011 he won the Turing Award, which is about as close as computer science gets to a Nobel Prize, “for fundamental contributions to artificial intelligence through the development of a calculus for probabilistic and causal reasoning.” That is the official language. In plainer terms, he is the person who gave machines a rigorous way to handle uncertainty, and then spent the back half of his career arguing that handling uncertainty was never going to be enough.
If you have heard of Bayesian networks, that is largely his work. If you have ever seen a diagram of bubbles and arrows that maps out how one thing causes another, the modern mathematical version of that traces back to him too. He built a lot of the foundation the current AI boom is standing on, which is what makes his criticism of that boom so interesting. This is not an outsider throwing rocks. This is the architect telling you which walls are load bearing and which ones are just painted on.
In 2018 he wrote a book for the rest of us called The Book of Why, co-written with Dana Mackenzie. If you only ever read one book to understand where AI is genuinely headed versus where the hype says it is headed, I would point you there. The dense academic version is his earlier book, Causality, but The Book of Why is the one you can actually enjoy on a weekend.
The central idea is something he calls the Ladder of Causation. Three rungs. Once you understand the three rungs, you understand the whole argument, and you understand why a business owner in the Comox Valley should care about an abstract piece of mathematics from a UCLA lab.
The three rungs
Pearl’s claim is that all reasoning about the world sits on one of three levels, and each level is genuinely harder than the one below it. You cannot fake your way up. A machine stuck on a lower rung cannot answer questions that live on a higher rung, no matter how much data you feed it.
Rung one is seeing
The bottom rung is association. This is pure pattern matching. You observe that two things tend to happen together, so when you see one, you expect the other. The classic shape of the question is, “if I see this, what is likely to also be true?”
A bird sees the sky darken and clouds gather, and it expects rain. Your inbox filter sees certain words and flags spam. Netflix sees that people who watched what you watched also watched this other thing, so it recommends it. Your sales dashboard notices that customers who buy product A often come back for product B.
This is enormously useful, and it is also exactly where today’s large language models live. When an AI model writes you a paragraph, it is, at heart, predicting the most probable next word given everything that came before. It has read a staggering amount of text and learned the patterns in it with incredible precision. Pearl’s blunt way of describing all of this is that “all the impressive achievements of deep learning amount to just curve fitting.” Fitting a very, very complicated curve to a very, very large pile of data.
That sounds dismissive, and people in the field have argued with him about it. But notice he is not saying it is useless. Curve fitting runs your spam filter and recommends your next show and drafts your emails. He is saying it is rung one, and rung one has a hard limit. It can tell you what tends to go with what. It cannot tell you what happens if you reach in and change something.
Rung two is doing
The middle rung is intervention. Now the question changes shape entirely. It is no longer “if I see this, what else is likely?” It is “if I do this, what will happen?”
This is the difference between noticing that your best months are the months you ran a particular promotion, and actually knowing whether running that promotion again will produce another good month. Those are not the same question. The first is association. You saw two things together. The second is intervention. You want to reach into the world, pull a lever, and predict the result.
Pearl gave this its own piece of mathematical notation. He writes it as do(X), the “do” operator, and built a whole system around it called the do-calculus for reasoning about exactly these kinds of questions. The reason he had to invent new math for it is the part that should make you sit up. You cannot get to “what happens if I do this” purely from observing data, no matter how much data you have. Two completely different causal stories can produce the exact same data. The numbers alone cannot tell you which story is true. You need something the data does not contain, which is a model of how things actually cause each other.
Here is the example everyone uses because it is so clean. Ice cream sales and drowning deaths rise together. The data shows a strong, reliable, year after year correlation. A rung one system sees that pattern and is perfectly happy with it. But nobody sensible would shut down ice cream stands to stop drownings, because we carry a causal model in our heads. We know there is a hidden cause sitting behind both, which is hot weather. Hot days drive ice cream sales up and they drive more people into the water. The ice cream is innocent.
That causal model, the thing that lets you ignore a real correlation because you understand the mechanism behind it, is precisely what a rung one machine does not have. It sees the correlation and has no way, from the data alone, to know it is a trap.
Rung three is imagining
The top rung is counterfactuals. This is the question that makes us human, and it is the hardest of all. “What would have happened if I had done it differently?”
What if we had not raised prices last spring, would we have kept that big account? What if I had hired the other candidate? What if this patient had not taken the medication, would they have recovered anyway? These questions are about a world that never happened. You cannot observe it. There is no data for the road you did not take. To answer them at all, you have to hold two worlds in your head at once, the one that happened and the one that did not, and reason about the difference between them.
Pearl’s framework is one of the only rigorous ways anyone has found to do this with mathematics rather than just intuition. And it is the rung where regret, responsibility, credit, blame, and genuine learning all live. When you say “that decision was a mistake,” you are making a counterfactual claim. You are saying a different choice would have led somewhere better. A machine that cannot climb to rung three cannot truly reason about its own decisions, cannot explain itself in the way we mean when we ask a person to explain themselves, and cannot really learn from a mistake the way you and I do.
Why this matters more now, not less
You might think that as the models get bigger, they naturally climb the ladder. They do not. That is the whole point, and it is the part people miss in all the excitement. A bigger pile of data and a bigger model give you a better fit to rung one. They make the curve fitting astonishing. They do not move you to rung two or rung three, because those rungs require information that simply is not present in observational data. You have to bring the causal structure to the problem from the outside.
This is why a model can write a flawless essay about why correlation is not causation and then, in the next breath, fall straight into a correlation trap when it is reasoning about your actual numbers. It learned the words about causation. It did not learn causation. Those are different rungs.
Pearl has been consistent and a little impatient about this. His view is that if we want machines that genuinely reason, that can be trusted with real decisions and can explain themselves, we have to build causal understanding in deliberately. It will not emerge on its own from scale. Plenty of brilliant people disagree with how strongly he puts it, and the honest state of things is that this is a live argument at the frontier of AI research. But almost nobody disagrees with the underlying map. The ladder is real. The rungs are real. The question is only how far pure pattern matching can carry us before the structure has to be added on purpose.
What this actually means for a business owner
Strip away the math and here is the practical version, the one I would give you over coffee.
The AI tools you can buy off the shelf today are spectacular at rung one. Summarizing, drafting, classifying, spotting patterns in your data, answering questions in plain language, handling the repetitive work that eats your week. If that is what you need, and for a lot of businesses it genuinely is, then the off the shelf tools are a gift and you should be using them.
But the decisions that actually determine whether your business grows almost all live on rung two and rung three. Should I change my pricing. Which marketing actually drove those sales and which just happened to be running at the same time. If I move this person off the front desk, what breaks. What would have happened if we had not done the thing we did. These are causal questions. Doing questions and what-if questions. And a tool that only ever did curve fitting will answer them with total confidence and no actual idea whether it is right, because it cannot tell the difference between a pattern that means something and a pattern that is a coincidence dressed up in good numbers.
This is the trap I watch people walk into. They see an AI give a fluent, confident answer about why their numbers moved, and they treat that confidence as understanding. It is not. It is rung one wearing rung two’s clothes. The model is very good at sounding like it knows why. Knowing why is a different and much harder thing.
So the real skill, the thing that separates a business that uses AI well from one that gets burned by it, is knowing which rung your question lives on. Pattern questions, hand them to the off the shelf model and move on with your day. Cause and decision questions, those need a system that has the causal structure of your business built into it on purpose, or at the very least a human who understands the difference standing between the model and the decision.
Where Entoura sits in all of this
This is not abstract for me. It is the center of how we think when we build.
What we do at Entoura is build custom software and engineer AI into businesses. When we wire an AI layer into how a company actually runs, the ladder is the first thing on my mind, even if I never say the words “Judea Pearl” to a client. Because there is a real and expensive difference between standing up a model that spots patterns in your data and building a system that can be trusted to reason about cause and feed a real decision. The first is a weekend with an API. The second takes actually understanding your business, mapping how the pieces cause one another, and being honest about where the machine should hand the decision back to a person.
That honesty is the part I care about most. The easy money in AI right now is selling rung one and letting people believe they bought rung three. Build a confident chatbot, point it at a hard question, let the fluency do the rest, and most people will not notice it is guessing until a decision goes sideways. I am not interested in building that. When we put AI into a business, I want it doing the pattern work it is genuinely great at, and I want the cause and decision work either built properly or handed clearly back to a human who can see the whole board. Plain language, real numbers, and a straight answer about what the machine can and cannot actually do. That is the whole pitch.
It is also, not by accident, why we build the way we build. You own and control the software we make for you, and we manage it so you do not have to. When a system is reasoning about real decisions in your business, you should know how it works and you should own it outright, not rent your own logic back from someone else on a subscription.
The part I keep coming back to
Pearl tells a story about his own three year old grandson, who could already do something effortlessly that the most advanced machine on earth still cannot. The child could ask why. Not parrot the word, actually reach for the cause behind a thing and imagine how it could have gone differently. We treat that as the most ordinary thing in the world because every toddler does it. It might be the most sophisticated thing any of us do.
So the next time an AI gives you a smooth, certain answer about why your sales dipped or why a customer churned or what will happen if you change your pricing, I want you to hear a small voice in the back of your head asking which rung that answer is really standing on. Most of the time the honest answer is rung one, a beautiful pattern with no understanding underneath it, and most of the time the tool will not volunteer that. Knowing the difference is not pessimism about AI. It is the thing that lets you actually trust it where it deserves trust and keep a hand on the wheel everywhere else.
That is the whole reason a forty year old idea from a UCLA lab ended up mattering for a business owner on Vancouver Island. The machines got fluent faster than almost anyone expected. They have not gotten wise. The gap between those two things is exactly where the good decisions, and the good software, still have to come from people. Judea Pearl just gave us the map for telling them apart, and once you are holding it, the whole conversation about AI gets a lot clearer.
Further reading: Judea Pearl and Dana Mackenzie, The Book of Why: The New Science of Cause and Effect (2018). Pearl’s Turing Award citation and the ACM announcement. His more technical Causality. And an interview where he discusses large language models and causal reasoning directly.