Most people think Artificial Intelligence is about prediction. Feed a machine a million cat photos, and it learns to spot cats. Give it a million medical scans, and it tells you which ones look suspicious. Amazing, right? But here's the problem: prediction is only half of intelligence.
Let's step back. If you drop a glass and it shatters, even a child knows something simple but profound: the drop caused the break. That's not prediction. That's understanding. You could've predicted the glass will fall. But knowing why it broke, and what would've happened if you caught it in time, is a different level of reasoning.
This is where Judea Pearl draws his line in the sand:
An AI scientist is someone who builds machines that don't just predict, but also understand, through the lens of causality and counterfactual reasoning.
Why Prediction Isn't Enough
Prediction is like memorising answers to exam questions. If tomorrow's test happens to use the same questions, you’re golden. But change the numbers, change the subject slightly, and memorisation collapses.
That’s the trap of today's AI. Feed it enough data, and it will "look smart." But ask it something it hasn't seen, or something that requires imagination about what if things were different, and it often stumbles.
Humans don't just see patterns. We run experiments in our heads.
- If I turn the steering wheel left, the car will turn left.
- If I water this plant more, it might grow faster.
- If I hadn’t hit "send," that embarrassing email wouldn't exist.
That’s counterfactual reasoning — imagining worlds that never happened but could have.
The Scientist's Task
So what makes an AI scientist? It's not bigger GPUs, more data, or shinier apps. That's engineering. An AI scientist wants to equip machines with the missing ingredient: the ability to ask and answer "why."
Think of it as three levels of intelligence:
- Association: Spotting that smoke usually comes with fire.
- Intervention: Knowing that if I put water on the fire, it might go out.
- Counterfactuals: Wondering what would’ve happened if the fire station had arrived sooner.
Right now, most of AI is stuck at Level 1. An AI scientist's job is to push it up the ladder.
Why This Matters for All of Us
You might be thinking: "Cool theory, but why should I care?" Well, if we keep building prediction-only machines, we'll get clever parrots, systems that mimic but don't understand. They'll be powerful, but brittle.
But if we build causal machines, we get tools that can reason about medicine, climate, economics, or even human behaviour in a way that helps us act, not just observe. That's when AI shifts from being a glorified calculator to a true scientific partner.
The Bottom Line
An AI engineer builds machines that answer what is likely to happen.
An AI scientist builds machines that can also ask and answer why did it happen, what if it hadn't, and what can we change.
That's the difference between memorising history and making history.
