Why AI cannot create new scientific knowledge

In his work Reflections on Language, published in 1975, Noam Chomsky argued that children learn to speak not just by imitating what they hear, but by constructing a theory from sparse and disordered information. According to the linguist, a child infers knowledge that goes far beyond what they have heard, allowing them to produce new sentences with no direct connection to previous experiences. In other words, children do not merely reproduce patterns — they create original knowledge.

For researchers Teppo Felin, from Utah State University, and Matthias Holweg, from the University of Oxford, this is the central point that differentiates human learning and reasoning from artificial intelligence (AI). In their academic essay Theory Is All You Need: AI, Human Cognition, and Causal Reasoning, published in late 2024 in the journal Strategy Science, the authors describe AI language generation as “backward-looking and imitative,” whereas human cognition is “forward-looking and capable of generating genuine novelty.”

“There are researchers who have analyzed how babies process their environment. And it turns out that they not only absorb data, but they are constantly making conjectures or formulating hypotheses,” explains co-author Felin, 52, via video call from Utah. “If I drop my cup on the table, I learn something about the world around me. And it turns out that’s precisely the crux of the matter: the ability to formulate conjectures, to want to experiment, or to formulate hypotheses.”

The researcher, who is also the founder of the Institute for Interdisciplinary Studies at Utah State University, states that one of his goals is to “debunk all the hype surrounding AI” and highlight how the human mind is unique in its causal and theoretical reasoning. The study emphasizes how the mind is not merely an information processor and that humans not only predict the world but also intervene in it and transform it. This, according to the authors, dismantles the mind-machine analogy.

Galileo and the Wright brothers

To illustrate the limitations of large language models, Felin and Holweg use the example of how an AI trained on the “corpus of thousands of years of geocentric texts” up to 1633 would deny Galileo Galilei’s heliocentric model. This is what the authors call “data–belief asymmetry” — that is, while AI will accept something as true if most texts assert it, humans can believe in something that contradicts the data.

This asymmetry is what allows human cognition to form beliefs that may initially seem delusional or contrary to existing knowledge, but which can eventually lead to new discoveries.

Felin states that large language models are, for now, translators or reformulators that reflect past patterns. “In Galileo’s time, the data indicated that the Earth was stationary. And if you look around, you see that the Earth is not moving and that the sun appears to move from east to west. Therefore, an AI with a limit of 1633 will accept that model as correct,” he explains.

To illustrate this point, the authors use the example of the Wright brothers and how, in the 19th century, scientists considered it impossible for objects heavier than air to fly. But while the scientific consensus ruled out human flight, the Wright brothers conducted experiments to solve the problems of lift, propulsion, and steering, demonstrating that flight was indeed possible.

“In uncertain environments, only human theoretical thinking has the advantage because creativity depends on theories that challenge data, not algorithms,” Felin argues. “AI extrapolates data from the past to predict what will happen in the future, but that only works when the environment doesn’t change and there is no uncertainty.”

The world is not a database

For Felin, human reasoning, despite its limitations, is the only thing that can accurately reflect a world that is “constantly changing.” “Human beings can process a limited amount of data, we are biased and make bad decisions, but it turns out that we live in a very dynamic environment and AI has no way of dealing with that,” the researcher explains.

Furthermore, the author argues that people have to make decisions without data every day. “In a sense, we’ve placed too much importance on data because we don’t always have the right data in front of us. So we have to think about how to obtain that data, and that’s what leads to creativity,” he continues.

Felin also warns against the “panic” that some specialists have fostered regarding AI. In the study, for example, the researchers cite the example of Geoffrey Hinton, one of the “pioneers of AI” and winner of the 2018 Turing Award, who speculated about the possibility that large language models could eventually exhibit forms of intelligence or consciousness. The authors reject this view and argue that equating the mind with these computational devices is “conceptually flawed and philosophically reductive.”

The Finnish academic argues that AI is a “technological wave with limitations, especially in areas that require true creativity, problem formulation, and strategic, forward-thinking decision-making.” Felin compares large language models to “a kind of dynamic Wikipedia” and argues that AI must be seen “for what it is: statistics and machine learning in action, with nothing mystical behind it.”

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition