A new study challenges the growing belief that artificial intelligence will ultimately replace human decision-making. Teppo Felin of Huntsman School of Business and Matthias Holweg of Saïd Business School, University of Oxford, argue that AI’s reliance on past data differs fundamentally from human theory-based causal reasoning. The researchers highlight a critical distinction, framing it as where AI predicts based on probability while humans employ forward-looking logic.
They illustrate this with the example of heavier-than-air flight, an innovation unlikely to emerge from a purely data-driven system. Current AI models also outperform more than 90% of humans in various professional qualification exams, such as the bar exam in law and the certified public accountant exam in accounting (Achiam et al. 2023).
AI’s Data-Driven Prediction Differs From Human Causal Logic
Human cognition prioritizes forward-looking theorizing, a process fundamentally distinct from the backward-looking, data-dependent approach of artificial intelligence. Unlike AI systems focused on identifying patterns within existing data, humans employ causal reasoning to generate novel data and explore hypothetical possibilities, a capability critical for genuine innovation.
This distinction manifests as human thought being driven by theories and beliefs that may initially lack extensive supporting evidence. The limitations of a purely data-driven approach are starkly illustrated when considering the history of flight; as one observer noted years ago, “There probably can be found no better example of the speculative tendency carrying man to the verge of the chimerical than in his attempts to imitate the birds.” The development of heavier-than-air flight wasn’t simply an extrapolation from observations of bird flight, but a result of developing a theory of aerodynamics, a conceptual leap that enabled the creation of something fundamentally new.
The Input-Output Analogy Limits Understanding of Human Minds
The prevailing view of cognition as computation overlooks a fundamental asymmetry between artificial intelligence and human thought; humans readily generate theories to guide experimentation, while AI primarily extrapolates from existing data. This distinction, detailed in recent work, challenges the notion that AI will fully replicate, or even surpass, human decision-making capabilities, particularly when facing genuinely novel situations.
Researchers argue that framing the mind as simply an information processor limits understanding of how new knowledge emerges, a process driven by forward-looking causal reasoning rather than backward-looking prediction. This emphasis on prediction, central to many AI systems, contrasts sharply with the human capacity for envisioning possibilities not directly supported by past experience.
Data-Belief Asymmetries Distinguish AI and Human Cognition
Human cognition actively seeks data to validate pre-existing theories, a process sharply contrasted with the predictive approach of artificial intelligence systems. This distinction explains why humans can generate genuinely novel ideas while AI often remains constrained by existing datasets. The study highlights that a rational decision maker does not simply weigh beliefs by available data, but actively shapes those beliefs through experimentation and the pursuit of confirming evidence, even in the face of contradictory information.
The capacity for forward-looking theorizing, central to human innovation, necessitates this asymmetry; humans don’t passively receive information, but formulate hypotheses and then design experiments to test them. As an example, the paper references Lord Kelvin, who, lacking belief in the possibility of human flight, dismissed any supporting evidence, while the Wright brothers actively pursued their despite limited initial data.
Forward-Looking Theory Drives Human Experimentation
Forward-looking beliefs actively shape the data humans seek, a process fundamentally different from how artificial intelligence operates. Unlike AI systems prioritizing prediction based on existing information, human cognition is driven by theorizing and causal reasoning that creates the need for new evidence, rather than simply processing what already exists.
This proactive stance means beliefs often precede and motivate data acquisition, a concept illustrated by the Wright brothers’ pursuit of heavier-than-air flight. Their work wasn’t simply extrapolating from bird flight; it involved developing a theory of aerodynamics and then designing experiments to prove, and refine, that theory.
Human Cognition Generates Novelty Beyond Imitation
Human cognition distinguishes itself from artificial intelligence through a fundamental capacity for forward-looking theorizing, a process not merely extrapolative but generative of entirely new data points. The ability to forge strong connections with the latest ideas from computer science, machine learning, and statistics does not equate to genuine novelty, the researchers contend. Large language models, while adept at assembling language stochastically, demonstrate imitation rather than linguistic innovation when compared to human children acquiring language skills.
The paper points to a critical limitation: LLMs, despite their power, are fundamentally constrained by the data they’ve been trained on, unable to independently generate truly new knowledge. This capacity for theory-driven experimentation is central to human learning and innovation, extending far beyond language acquisition. The researchers suggest that indicating that even seemingly simple acts of perception are underpinned by pre-existing frameworks of understanding.
AI’s Probability-Based Approach Relies on Past Data
Large language models generate text by predicting the most probable next word, a process fundamentally rooted in the patterns encountered within their training data. This reliance on past inputs distinguishes them from human language acquisition, where predictive processing operates alongside a capacity for theory-driven reasoning, not merely statistical likelihood. While both systems aim to anticipate what comes next, the human brain actively weighs the source of information, prioritizing expertise and consensus when forming beliefs, a nuance absent in current AI architectures.
Consider the historical pursuit of heavier-than-air flight; a rational assessment of existing data, observing birds and insects, might have suggested limitations based on size and wing structure. However, focusing on reliable, scientific sources and a developing understanding of aerodynamics allowed for a fundamentally new approach, one not simply extrapolated from past observations.
This highlights a critical distinction: AI models, even sophisticated ones, struggle to move beyond the boundaries of their training data, while humans can formulate beliefs based on underlying causal structures and hypothetical possibilities, Wright brothers says. The authors write, emphasizing the role of conceptual frameworks in human thought. The emphasis on prediction, central to both large language models and many approaches to cognitive science, is not inherently flawed, but its limitations become apparent when making forward-looking decisions.
Data, by its nature, represents a snapshot of the past, and relying solely on past data risks overlooking novel possibilities or unforeseen circumstances, according to Wright brothers. Agrawal et al. (2022) have articulated this concern, and the current work builds on it by demonstrating how this reliance on past data impacts the ability to generate genuinely new knowledge.
The authors contend that data is not necessarily the best source of information when navigating uncertainty, a condition inherent in most real-world scenarios. The claim that machine learning is “theory-free” is a misnomer; the architects of these systems make numerous top-down decisions regarding algorithm design, learning processes, and valued outputs. These choices inherently embed theoretical assumptions, even if they are not explicitly articulated. The researchers note that “learning of languages would be the most impressive, since it is the most human of these activities,” suggesting that true intelligence requires more than just pattern recognition.
They frame this as a difference between AI’s data-driven approach and human behavior, which can be understood as solving a statistical inference problem, but one informed by pre-existing beliefs and causal models. “It is wrong always, everywhere, and for anyone to believe anything on insufficient evidence,” they state, underscoring the importance of critical evaluation and informed judgment.
Will there be anything that is reserved for human beings? Frankly, I don’t see any reason to set limits on what AI can do…And so it’s very difficult to imagine that with sufficient data there will remain things that only humans can do.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
