A provocative paper from Teppo Felin and Matthias Holweg argues that the biggest misunderstanding about artificial intelligence may be hiding in plain sight:
Prediction is not the same thing as thinking.
Their paper, Theory Is All You Need: AI, Human Cognition, and Causal Reasoning, published in Strategy Science, challenges the idea that artificial intelligence can simply absorb enough data and eventually replace human reasoning. The authors argue that AI's data-driven prediction is fundamentally different from human theory-based causal reasoning.
And their argument becomes particularly interesting when you ask a simple question:
Where does genuinely new knowledge come from?
The machine has billions of answers. The human asks a different question.
Modern LLMs are trained on extraordinary quantities of text.
Felin and Holweg estimate that leading models around early 2024 were trained on roughly 13 trillion tokens. At 150 words per minute, a human would need more than 164,000 years just to read that amount of text.
Yet the authors point to an apparent paradox.
A child receives dramatically less linguistic input. The paper cites estimates of roughly 20,000 words per day, while emphasizing that human learning is also multimodal, involving vision, touch, hearing and interaction with the physical world.
And somehow, children don't simply repeat the sentences they have heard.
They construct new sentences.
They form hypotheses.
They ask questions.
They imagine things that have never happened.
And eventually, some humans produce ideas that nobody has previously written down.
That distinction is at the centre of Felin and Holweg's argument.
The problem with learning from the past
An LLM learns statistical relationships in its training data and uses those learned relationships to generate new sequences of language.
The authors describe this as fundamentally backward-looking.
The system learns from what already exists.
Human theorizing can work differently.
A person can say:
"The evidence says this is impossible. But what if the evidence is wrong?"
That statement is dangerous when it becomes stubbornness.
But it can also be the beginning of a scientific revolution.
This is what the authors call a data-belief asymmetry: a situation where someone's belief runs ahead of the available evidence, or even contradicts prevailing evidence.
The crucial difference is what happens next.
The human can act on the theory.
Build something.
Run an experiment.
Generate new evidence.
And potentially change what the data says.
Imagine giving an AI all the knowledge of 1633
The paper uses historical examples to make the point.
Imagine an AI system trained on the scientific knowledge available in the early seventeenth century.
It would have enormous quantities of material supporting prevailing assumptions about the universe.
Then along comes a theory that challenges those assumptions.
The problem isn't that the machine has too little information.
It is that the new theory may initially have less supporting data than the established theory.
This is precisely where Felin and Holweg say human cognition becomes different.
Humans can entertain a theory before the evidence exists.
The theory can then motivate an experiment that produces the evidence.
The machine, by contrast, is being asked to infer from the evidence already available.
That is a fundamentally different loop.
The Wright brothers understood the difference
The paper's most powerful example is flight.
For centuries, humans watched birds fly without knowing how to reproduce it.
In the late nineteenth century, prominent scientific opinion remained deeply skeptical about heavier-than-air flight.
The Wright brothers nevertheless believed that powered human flight was possible.
That belief wasn't simply an answer extracted from a giant database.
It became a research program.
They broke the problem into components including lift, propulsion and control. They experimented, constructed equipment and generated new aerodynamic data.
In other words:
They didn't wait for the data to tell them flight was possible.
They developed a theory that told them what data they needed to create.
That distinction is central to Felin and Holweg's argument.
This is why "AI can't invent" is more complicated than it sounds
There is an important nuance here.
The paper does not establish a mathematical theorem proving that no AI system can ever invent anything. Its claim is more specific: the authors argue that data-based AI prediction, particularly as exemplified by LLMs, does not provide the same mechanism for genuine novelty and forward-looking causal reasoning that human theorizing provides.
That's a much more interesting argument than simply saying "AI is stupid."
Because the authors explicitly recognize what AI is extremely good at.
AI can process enormous quantities of information.
It can identify patterns humans miss.
It can make predictions.
It can automate repetitive decisions.
And those capabilities are enormously valuable.
The question is what happens when the future contains something fundamentally unlike the past.
The future isn't in the dataset
Suppose every historical dataset says a particular product will fail.
A human entrepreneur can still say:
"I think the market has changed."
Then build the product.
The market responds.
New data appears.
The original dataset is no longer the complete description of reality.
This is the deeper idea behind the paper.
Sometimes data is not merely something you consume.
Sometimes data is something you create.
And creating new data requires acting on a hypothesis about a world that does not yet exist.
That is why the Wright brothers matter so much to the argument.
They didn't discover a hidden flight dataset.
They created experiments that produced information nobody previously possessed.
And this is where humans remain extremely important
The most interesting implication isn't that humans should stop using AI.
It is almost the opposite.
Use AI for what it does exceptionally well.
Let machines search, summarize, classify, calculate, compare and predict from enormous quantities of existing information.
But don't confuse those abilities with the entire process of discovery.
The human contribution may increasingly become the ability to decide:
What if we're asking the wrong question?
What if the consensus is wrong?
What experiment should we run?
What doesn't exist yet?
What should we build even though the historical data says it shouldn't work?
That is a different kind of intelligence.
The human at the keyboard still matters
This is also why my experience using AI increasingly makes sense through this lens.
The model can give me ten thousand possibilities.
But deciding that the eleventh possibility is worth pursuing is different.
The model can summarize yesterday.
The human can decide to build tomorrow.
The model can identify patterns in existing information.
The human can decide that the pattern itself is misleading.
The model can generate an answer.
The human can decide that the question deserves to be challenged.
Felin and Holweg aren't arguing that AI is useless. They are challenging the assumption that increasing computational power and data automatically eliminates the need for human theory, causal reasoning and experimentation. Their conclusion is that theory-based reasoning plays a foundational role in generating novelty and new knowledge.
And perhaps that's the most important lesson.
The future doesn't always come from better predictions of the past.
Sometimes it comes from someone believing the past is wrong.
Then building something to find out.
LLMs can predict what has been written.
Humans can decide what needs to be written next.
LLMs don't think.
You do.
Comments
Post a Comment