Can Artificial Intelligence substitute the generation of knowledge through human intelligence? Or will human driven scientific intelligence keep the lead in creating novel insights? These are basic questions when it comes to the potential of AI.
Undoubtedly, AI technology offers significant advantage to the processing of bulk data, what enables new ways of pattern recognition and their utilization. However, as it is generically the forward projection of empirical data, it lacks the creativity of unprecedented thinking. On the other hand, scientific intelligence, as the established method in human research, builds upon the capability of human thinking with its inherent principal element of ideas, what requires the ability to imaginate. Of course, this human intelligence is reliant on empirical information about reality while it uses theoretical foundation.
The well-known mutual dependency of empiricism and cognition is less a dichotomie, but more a fruitful combination. And so can the two questions above, in the context of powerful AI technology, become the starting point for new methods of generating new knowledge. That is what we earlier suggested as hybrid intelligence.
A most valuable exemplification is presented by an actual work in the context of climate change, „Physics-based models outperform AI weather forecasts of record-breaking extremes“. It compares a physics-based numerical model HRES from the European Centre for Medium-Range Weather Forecasts with state-of-the-art AI models and demonstrates that forecast errors in AI models are consistently larger for record-breaking weather than in HRES across nearly all lead times. Among others, the authors suggest that promising further research should involve two-sided hybridization: hybrid models where specific parametrizations in physical climate models are replaced with AI components. Or physics-informed neural networks that are trained while respecting specific physical laws described by nonlinear partial differential equations.
Norbert Jastroch, https://orcid.org/0000-0002-4046-450X
- Jastroch, N.: Sustainable Artificial Intelligence: In Search of Technological Resilience. Springer 2023. https://doi.org/10.1007/978-3-031-25182-5_31. PLM2022, Grenoble/France
- Physics-based models outperform AI weather forecasts of record-breaking extremes. Sci. Adv.12, eaec1433(2026). DOI:10.1126/sciadv.aec1433
Note: This post was written without using A whatsoever
- Oznake
- AI development
- Za objavu komentara morate se prijaviti
Komentari
The question of whether artificial intelligence can substitute human-driven knowledge generation is often framed too dichotomously. In practice, the trajectory observed across scientific domains suggests convergence rather than substitution.
Current AI systems excel at extracting structure from large-scale datasets, enabling advances in pattern recognition and probabilistic forecasting. However, their epistemic foundation remains anchored in statistical inference over observed distributions. This makes them highly effective within the bounds of known regimes, but comparatively less robust when extrapolating toward rare, extreme, or previously unobserved phenomena.
This limitation is not incidental; it is structurally consistent with the design of such systems. In contrast, human scientific intelligence integrates empirical observation with abstraction, hypothesis formation, and conceptual reasoning. It is precisely this ability to formulate models that are not strictly derived from prior data distributions that underpins major scientific breakthroughs.
From this perspective, the emergence of hybrid intelligence frameworks is not only expected but necessary. The example cited—comparing AI-based weather forecasting systems with physics-based numerical models such as HRES from the European Centre for Medium-Range Weather Forecasts—illustrates this point clearly. The observed superiority of physics-based models in predicting record-breaking extremes reflects the importance of embedding first-principles constraints, particularly when addressing non-linear and boundary-case dynamics.
This outcome is entirely consistent with longstanding practices in climatology. Institutions such as Météo-France have, for decades, developed and refined numerical weather prediction systems grounded in physical laws, supported by high-performance computing infrastructures. The computational paradigms underpinning contemporary AI models—large-scale optimization, distributed processing, and numerical approximation—are themselves a continuation of this broader lineage of advanced scientific computing.
In that sense, AI does not represent a conceptual rupture but rather an extension of an existing methodological continuum. Both approaches are rooted in the same fundamental intuition: that complex natural systems can be modeled through formalized representations and computational iteration.
The critical question, therefore, is not whether AI will replace human scientific reasoning, but how these complementary paradigms can be integrated effectively. The concept of hybridization—whether through AI-enhanced parametrization within physical models or through physics-informed neural networks constrained by governing equations—points toward a structurally sound path forward.
A more strategic question also emerges at the governance level: should these advanced computational models, which increasingly shape public understanding and policy decisions, be developed within open frameworks? The case for open source is not merely ideological. It directly impacts auditability, reproducibility, collective validation, and long-term resilience. In domains such as climate science, where societal stakes are high and trust is essential, openness may well become a prerequisite for legitimacy.
In conclusion, the evolution toward hybrid intelligence reflects a natural progression in scientific methodology. Rather than competing paradigms, AI and human-driven science form an interdependent system—one that, if properly structured and governed, can significantly enhance our capacity to generate reliable and actionable knowledge.