Why AI Can't Tell You Why: The Causal Gap in Enterprise Analytics

The Smartest Student Who Never Learned a Thing#

There is a kind of intelligence that knows everything it has ever been taught, and nothing it has not.

Call it the frozen genius. Feed it ten trillion words. Train it on every dataset ever assembled. Ask it anything and it will answer with uncanny precision. Then deploy it into the real world, and watch.

It stops learning. Completely.

In March 2026, three researchers published a paper with an uncomfortable title: Why AI systems don't learn and what to do about it.

The names on it weren't obscure academics. Yann LeCun is Meta's Chief AI Scientist and one of the people most responsible for the deep learning revolution. Jitendra Malik is at Berkeley. Emmanuel Dupoux at the École des Hautes Études en Sciences Sociales. These are people who built the thing that's now running inside half the world's enterprise software stacks.

And they were saying, quietly but clearly, that it has a fundamental problem.

The paper isn't a doomsday argument. It's almost clinical. They lay out what current AI systems actually do: they observe patterns in data, they compress those patterns into weights, and then they stop. Once deployed, they learn nothing. They adapt to nothing. And they cannot tell the difference between a coincidence and a cause.

That last point matters more than most people realise.

Somewhere around 2023, a global logistics company started using AI tools to understand why vehicle accidents were happening. They had 12,000 incident records. Collision data, infringement logs, telematics feeds. More information than any team of analysts could process in months.

A modern language model, given that data, would do what it always does: find the strongest correlations. It might tell you that accidents cluster on certain routes, or that certain vehicle types appear more frequently in the incident log. It would present these patterns confidently, formatted beautifully, in seconds.

It would not tell you why.

The "why" turned out to be rest breaks. Specifically, the absence of them at the right times. That's a causal relationship, not a statistical pattern. Correlation might have pointed at the routes, or the vehicles, or the drivers. Causation pointed at the scheduling.

The difference between those two answers is the difference between buying new vehicles and changing a policy. One costs millions. The other costs a conversation.

RootCause.ai found the causal driver in three hours.

What LeCun and his colleagues are describing in their paper is a structural limitation, not a tuning problem. They call it the gap between System A and System B learning. System A is observation: you see data, you build a model of patterns. Every large language model ever trained is System A. It watches. It never touches. And because it never touches, it can never run an experiment, never see what happens when you change one variable and hold everything else fixed.

That's the only way to establish causation. You have to intervene.

The paper puts it plainly: because these systems are based purely on observation, they struggle to distinguish between correlation and causation. They don't say this as a feature request for the next version. They say it as a description of something that cannot be fixed by making the model bigger or feeding it more text.

This is worth sitting with for a moment, because the enterprise AI market is currently worth hundreds of billions of dollars and most of it is built on top of System A. Companies are making pricing decisions, operational decisions, strategic decisions, based on technology that its own architects are now describing as constitutionally unable to understand why anything actually happens.

Morgan Housel once wrote that the hard part of finance isn't the math. It's that the math looks right even when the reasoning is wrong. You can build a perfectly coherent model on a false premise and it will produce confident outputs until something breaks catastrophically.

AI analytics has the same problem. The outputs are fluent. The confidence is real. The causal reasoning is missing.

What Netflix and Amazon have spent years and significant money building internally isn't better pattern recognition. It's causal infrastructure. Teams that can run experiments, model interventions, and say with confidence: if we change this, here is what happens to that, and here is how certain we are. That's not a data science luxury. It's the difference between knowing your business and guessing at it.

The financial services firm that came to RootCause.ai had 142 KPIs tracked across 42 markets. A manual root cause analysis on one KPI in one market took two analysts three to four days. That's 5,946 variations. The math on human-hours is damning. The math on decisions made without that analysis is worse, because those decisions were still being made. They just weren't being made well.

LeCun's paper ends on a note of cautious optimism. The researchers believe that autonomous, adaptive AI systems are possible. They just think we're decades away from the full version, and that the current approach, more data, more compute, larger models, has largely run its course as a path to genuine understanding.

The gap they've identified, the causal gap, isn't waiting for some future architecture to fill it.

It's a gap that causal inference already fills, today, in the problems that actually cost businesses money.

The question worth asking isn't whether your AI can produce an answer. It's whether the answer tells you what to do, or just what happened. One of those is a story. The other is a decision.

Read the paper: Dupoux, LeCun, Malik - Why AI systems don't learn and what to do about it (March 2026)