Artificial intelligence currently dominates conversations about transformational technologies, but measurable progress beyond a few notable achievements remains limited. Peter Coveney and Roger Highfield argue that the relationship between AI and physics is severely unbalanced, and that physics has much more to offer for current AI development than the other way around. They demonstrate that existing AI architectures, including large-scale language models, often rely on vast numbers of parameters without providing true understanding or grasp of fundamental scientific principles. This research highlights the limitations of current AI and proposes a new path forward, termed “Big AI,” that integrates established theoretical frameworks with the adaptability of machine learning to create more robust and insightful artificial intelligence.
While current AI is good at identifying correlations, it often lacks a deeper understanding of the underlying physics, chemistry, and biology that it models, simulating rather than owning intelligence. This leads to reduced reliability, especially in critical applications such as science and medicine. These models tend to produce counterfactual information, struggle with complex systems, and accurately predict rare but significant events. Big AI proposes combining the power of large datasets with established scientific theory to build models constrained by fundamental laws through techniques such as physics-based neural networks and hybrid approaches that combine AI with traditional numerical simulation.
This emphasizes understanding causation beyond simple correlation, enabling more accurate predictions and interventions, and ultimately creating realistic and reliable human digital twins for personalized healthcare, accelerating drug discovery, and fostering a deeper understanding of human biology. Applications of this approach include designing molecules with specific properties based on chemical principles, improving the accuracy of weather predictions, especially for extreme events, and designing new materials with desirable properties based on physics and chemistry. Big AI also promises to advance personalized healthcare through the creation of digital twins of patients to predict disease risk and optimize treatment plans, and through the use of quantum-informed machine learning to improve predictions in chaotic systems. Reliability is paramount for critical applications, and the integration of scientific theories is essential to building truly intelligent and reliable systems. Because the future of AI is not just about building large-scale statistical models, but about building systems that understand the world around them, guided by the laws of nature.
Basic models fail to generalize physical laws
This study highlights the limitations of current artificial intelligence architectures and suggests a path toward more robust and interpretable systems by integrating principles of physics. The researchers demonstrated that while the underlying model is good at pattern recognition, it lacks the ability to generalize the underlying physical laws, even when presented with rich data. To illustrate this, the team tested the underlying model's ability to learn Newtonian mechanics using trajectory trajectory data. As a result, even though the model accurately predicted the movement of celestial bodies, it failed to grasp the fundamental laws of gravity and instead relied on task-specific shortcuts. This study took a rigorous approach, focusing on the model's ability to extrapolate beyond its training data and apply learned principles to new physics tasks.
When the researchers specifically tested whether the model could derive Newton's law of gravity by analyzing orbital data and calculating second-order derivatives, they discovered a major flaw. In other words, the model reflected the historic Ptolemaic epicycle system, which prioritized accurate predictions over understanding the governing laws. Furthermore, this study highlights that reliance on Gaussian distributions within machine learning algorithms is widespread even when applied to non-Gaussian real-world data, which can lead to inaccurate predictions and unstable outputs when dealing with complex nonlinear systems. To overcome these limitations and build AI systems that are truly understandable and generalizable, this research advocates a move toward “big AI” that combines the rigor of theory with the flexibility of machine learning.
AI models lack scientific understanding
This study reveals that, despite considerable hype, current artificial intelligence, aside from a few specific successes, has only modest measurable impact and often lacks a true understanding of the systems it models. Researchers have demonstrated that existing large-scale language and inference models rely on vast numbers of parameters, yet fail to capture even elementary scientific laws, fail to provide mechanistic insights, and, while good at identifying patterns, remain “black boxes” that do not provide explanations for the patterns. why Those patterns exist. Experiments have shown that basic models trained on extensive datasets consistently fail to apply Newtonian mechanics when adapted to new physics tasks, even when the data itself contains the information needed to discover the laws. Rather than generalizing principles, the model learns task-specific shortcuts and produces illogical results.
This is because current AI often assumes that data follows a normal Gaussian distribution, which is a simplification that does not take into account the nonlinear and often discontinuous nature of real-world phenomena. A comparative study of six large-scale language models and their inference-optimized variants shows that inference-tuned models consistently outperform non-inference models on scientific computing and machine learning tasks, but even these advanced models are prone to ambiguous or inaccurate outputs, especially when tackling problems of medium to high complexity. Further analysis reveals that the inference model outperforms the standard model only on tasks of moderate complexity, indicating clear limits to its ability to reliably solve complex problems.
Physics-based AI offers new avenues
This study shows that although artificial intelligence is currently receiving considerable attention, its measurable impact remains limited, especially when compared to the potential of physics to inform and improve AI development. The authors say that current AI architectures, including large-scale language models, often rely on over-parameterization, lack a fundamental understanding of the underlying scientific laws that govern the data they process, and highlight issues such as distributional bias and lack of quantification of uncertainty that impede the reliability and interpretability of AI results. This research proposes a path forward through the development of “big AI” that combines theoretical rigor with the flexibility of machine learning. This approach emphasizes the importance of building AI on established scientific principles, allowing for more robust, generalizable, and insightful models. The authors acknowledge that current AI systems can exhibit bias and inaccuracy, and that these limitations must be carefully considered when applying AI to critical areas. Future work should focus on integrating physical laws and theoretical frameworks into AI architectures to overcome these shortcomings and unleash the full potential of this technology, and on continuing to explore the inductive biases of the underlying models to better understand the internal representations of the world.
