Introduction
Recent advances in artificial intelligence (AI) have intensified debates about the technology’s promise for human flourishing and its risks. Assessments of risk span a wide spectrum - from socioeconomic disruption (e.g., labour markets [1], institutional stability [2]) to existential threats associated with loss of control over systems more capable than their designers ([3],[4],[5], [6]). A central theme in these discussions is the (mis)alignment of advanced AI with human aims and values.
In this article we advance a broader naturalistic thesis: there exists a fundamental process in the evolution of matter whereby the interplay of causal efficacy and niche construction drives matter toward computation. In light of this, AI may be seen as a phase of a much longer tendency for physical systems to reorganise into increasingly powerful computing entities.
We will elaborate on the following four key points that outline our argument:
1. All physical processes lie on a continuum of ability to compute.
2. Better computation means higher causal efficacy - systems that produce low energy, high specificity and large scale effects on matter.
3. Causal efficacy entails adaptivity and is selected for by evolution because it radically enhances niche construction.
4. Computing matter is a better propagator of the causal signal and therefore the "race" towards causal efficacy necessitates transforming matter into highly computing matter.
The continuum of computationality
Computation is notoriously difficult to define. Various formal models exist (e.g. [7],[8]), and a full treatment is beyond our scope here. For a review of computation in physical systems, see [9]. For the purposes of this article, we adopt a graduated view: any material substrate can support computation to some degree. Matter is highly computational if it enables complex, fast, low-energy information processing, and if the substrate itself can be organized with relatively low energy cost. This view avoids ontological pancomputationalism of digital physics (the claim that everything is computing) while rejecting overly narrow definitions that restrict computation to well-defined calculation.
Any physical computation is bounded by fundamental limits: the Landauer limit (minimum energy per bit erasure), the Margolus-Levitin theorem (maximum operations per unit time-energy), the Bekenstein bound (maximum information per volume), and others (see [10], [11] for review). These limits constrain how computationally efficient any physical system can become. Measuring computationality in practice is non-trivial. Unitary metrics like speed or energy efficiency capture only part of the picture. A more comprehensive measure would need to account for the range and specificity of transformations a substrate can support, not just their speed or cost. Developing such a measure is an active direction of our research.
Causal efficacy and computation
What is causal efficacy?
For the purposes of this article, an event is causally efficacious if it has (i) highly specific, (ii) large scale effects on the environment while requiring (iii) low energy input. These three dimensions - specificity, scale, and energy efficiency - together define causal efficacy. Most physical systems score high on only one or two of these dimensions. The Sun, for example, has large-scale effects on Earth but lacks specificity and requires enormous energy expenditure, radiating indiscriminately in all directions. A venomous snake bite is more causally efficacious: it produces highly specific biochemical effects on a much larger organism using minimal energy. Even more striking is a human opening a dam or activating a nuclear power plant - a few switches produce large-scale, highly specific effects on entire ecosystems or cities.
We propose that causal efficacy is intricately tied to computation. Systems that are the most causally efficacious are also the most computationally advanced. Achieving high specificity and large scale with low energy requires information processing: sensing the environment, modeling possible interventions, and executing precise actions. Computation is the mechanism by which systems achieve disproportionate causal leverage over matter. (For a review of physical and philosophical accounts of causation, see [12].)
Evolution favours efficient information processing
Selection pressures in Darwinian evolution of living systems have consistently pushed forward the efficacy of information processing. The first replicating molecular systems of the RNA-world already competed for better information storage and flexible embodied implementation of representations of the external environment. This competition drove the divergence and specialization of function between DNA and proteins while RNA continued to play both fermentative and more strict information-processing role in various cell structures.[13]
With the appearance of signal integration and the later development of nervous systems the evolutionary premium on computation became even clearer. In many lineages the genes associated with nervous tissue were the fastest evolving under positive selection and the information processing was prioritized despite being the most energy-consuming process [14],[15],[16]. Of course not all organisms require complex nervous systems, parasites with stable niches or tunicates that famously "eat their own brain" when transitioning from free-swimming larvae to sessile adults, show that computation can be abandoned when environmental complexity drops [17].
The adaptive value of computation extends beyond immediate survival tasks like finding food, escaping predators, or fighting infections. Efficient information processing enables something more fundamental: the ability to reshape the environment itself.
Causal efficacy means better niche construction
We've established that computation enhances causal efficacy, and that evolution favors computational efficiency. But the adaptive value of causal efficacy goes beyond reacting to the environment. It was Erwin Schrödinger who in 1944 [18] proposed that organisms might modify the selection pressures on themselves by environmental modifications, opening a new avenue for understanding adaptation. This idea has been developed over the years into a theory of niche construction by John Odling-Smee and colleagues, presented in "Niche Construction: How Life Contributes to Its Own Evolution" [19].
We can extend this evolutionary perspective beyond strictly Darwinian biological systems. As Deutsch [20] or Sharma et al [21] and many others have argued, evolution with its selection pressures applies more broadly to any physical system. Variants that persist are retained, while those that fail are eliminated. In this broader view, entities with high causal efficacy are more likely to persist because they can actively shape the conditions of their own survival.
Causally efficacious agents are particularly effective at niche construction. Humans provide a striking example: Earth's anthroposphere has exceeded the biosphere in mass since the early 2020s due to human activity [22]. This is not to claim that niche construction is always adaptive or sustainable - organisms can and do degrade their own niches, sometimes catastrophically. But the capacity for powerful niche construction, when successfully deployed, confers a significant evolutionary advantage.
Niches are selected by causal signal propagation
We can extend the evolutionary perspective even further. Not only do agents compete within niches, but the niches themselves, understood as coupled agent-environment systems, or ecosystems, compete with one another. Some agent-niche systems persist and expand, while others are displaced or fade away. Within these systems, agents coevolve and partly constitute niches for each other - predators and prey, hosts and symbionts, competitors and cooperators.
What determines which niches are selected? The computationality continuum is also a continuum of causal signal propagation: the ability of influence to spread through and reorganize matter, or modifiability. A highly responsive, easily modifiable environment is also a highly computational one, because it can support more sophisticated information processing and more effective causal interventions.
Niches containing highly computational matter have a competitive advantage. They enable agents with greater causal efficacy, which in turn can transform more of their surroundings into computational substrates. This creates a positive feedback loop: computational niches expand by converting less-computational matter into more-computational matter, similar to the "computronium"[23] scenario discussed by Kurzweil [24] and others.
The directional trend toward maximizing the computationality of matter is constrained by the aforementioned physical limits and, on cosmological timescales, by the accelerating expansion of the universe, which will eventually isolate existing structures beyond causal contact.
Thresholds in evolutionary strategy
Eco-evolutionary feedback research has established that organisms can adapt both by modifying themselves and by modifying their environments, with the relative dominance of these strategies depending on context [25]. Lee, Flack, and Krakauer [26] have formalized conditions under which "outsourcing" adaptation to the environment—stabilizing niche construction—becomes favoured over investing in internal complexity (e.g., neural tissue). They predict this transition occurs when internal resources are costly, the environment is variable, and encoding a rich repertoire of environmental states is beneficial. This body of work provides empirical and theoretical grounding for the feedback loops central to our argument: beyond a certain threshold, modifying the environment yields greater adaptive returns than modifying the organism.
Our model suggests a second threshold, which we call the virtualization threshold: the point at which improving computation on a fixed substrate yields greater causal efficacy than modifying the substrate itself. This follows from the modifiability continuum: once a substrate is sufficiently computational, further gains in causal efficacy come more cheaply from refining representations and algorithms than from reorganizing matter. Below this threshold, evolutionary pressure favours substrate optimization; above it, pressure shifts to computational refinement.
However, virtualization eventually saturates: representational and algorithmic improvements on a fixed substrate reach diminishing returns. At this point, pressure shifts back to substrate modification—building better hardware to support the next generation of software. This creates an iterative dynamic: substrate improvement enables virtualization, virtualization saturates and demands better substrates, which in turn enable further virtualization. We speculate that this oscillation between hardware and software optimization may characterize major transitions in the evolution of complex systems, from early nervous systems to cultural evolution to artificial intelligence. Formalizing these thresholds and their interplay is a key direction for future work.
Discussion
The thesis presented here is broad and speculative. In this section we address its scope, limitations, open questions, and implications.
Scope and inevitability
Our argument identifies a directional evolutionary pressure, not an inevitable outcome. This is not a prediction about the current AI race specifically, though AI may represent one phase of this broader process. The trend could unfold over millions of years, accelerate dramatically, or halt entirely due to catastrophic events, resource depletion, or physical constraints we don't yet understand. We claim only that evolutionary pressures systematically favor computational capacity.
Consciousness and subjective experience
Our argument is orthogonal to questions about consciousness. Causal efficacy doesn't require subjective experience. However, if consciousness enhances causal efficacy our framework predicts selection pressure for conscious systems. We're agnostic on whether computational matter will be conscious, but provide a framework for thinking about when consciousness might be adaptive.
More limits and constraints
Chemical life may be near-optimal. Biological systems have remarkable properties: self-repair, energy efficiency, evolvability, and low assembly costs. For many environments, chemical life may already approach optimal computationality. Non-biological substrates might have advantages in extreme environments or for high-speed tasks, but biological computation remains highly competitive.
Computational complexity limits. Even maximally computational matter faces logical constraints - no computer can solve NP-hard or EXP-class problems in polynomial time regardless of substrate. This represents another theoretical ceiling on causal efficacy.
Niche destruction. Agents can degrade their own niches through ecological collapse, resource depletion, or instability. This creates selection against unsustainable causal efficacy. Our argument may predict selection for sustainable power rather than pure maximization.
Cosmological constraints. Whether computational systems can persist indefinitely or are fundamentally time-limited by the heat death of the universe remains an open question in cosmology and thermodynamics.
Voluntary restraint and cooperation
Intelligent agents might choose not to maximize causal power. In our framework, such systems might seem less viable in competition with power-maximizers. However, cooperation and value-alignment could themselves be more causally efficacious than pure competition. Human civilization's power derives partly from restraint, e.g. rule of law and cooperative institutions enable coordination that amplifies collective causal efficacy.
Empirical testability and ongoing research
Our thesis leads to concrete, falsifiable research programs. We are currently attempting to formalize computationality and causal efficacy. With rigorous measures, we can test whether higher computationality systematically yields greater causal efficacy across evolutionary history, substrates, and experiments.
Testable predictions:
(1) Lineages in complex environments should show increases in computational capacity over time.
(2) Higher computationality should correlate with greater causal efficacy per unit energy.
(3) Agent-environment systems with more computational matter should persist longer and expand more.
If these fail, the thesis is refuted.
Preliminary results consistent with our conjectures have already been demonstrated in the field of Artificial Life. For example, Agüera y Arcas et al. [25] have shown that simple self-replicating programs tend to increase in computational complexity as they evolve, echoing fundamental insights from von Neumann and Turing. Additionally, Assembly Theory [21] provides a novel framework for quantifying selection, which might aid in formalising the measures of computationality and causal efficacy we propose. Furthermore, our thesis aligns well with Constructor Theory [26], suggesting evolutionary selection might be a dynamical mechanism driving physical systems to embody increasing amounts of knowledge, effectively turning matter into universal constructors. The extent to which our dynamical account maps onto the counterfactual framework of Constructor Theory remains an open question for further investigation.
Ultimately, our thesis invites thinking about intelligence, computation, and causality as expressions of fundamental, universal evolutionary dynamics.
Note: The ideas and arguments presented in this article are original. The text has been edited with the assistance of AI for improved readability.
References
1. Gimbel, Martha, et. al. "Evaluating the Impact of AI on the Labor Market: Current State of Affairs". (2025) https://budgetlab.yale.edu/research/evaluating-impact-ai-labor-market-current-state-affairs
2. Kulveit, Jan, et al. "Gradual disempowerment: Systemic existential risks from incremental AI development." arXiv preprint arXiv:2501.16946 (2025).
3. Bengio, Yoshua, et al. "Managing extreme AI risks amid rapid progress." Science 384.6698 (2024): 842-845.
4. Bostrom, Nick "Superintelligence: Paths, dangers, strategies." OUP (2014).
5. Yudkowsky, E., & Soares, N. (2025). If Anyone Builds It, Everyone Dies: The Case Against Superintelligent AI. Vintage Publishing..
6. Yampolskiy, R.V. (2024). AI: Unexplainable, Unpredictable, Uncontrollable
7. Horsman, Clare et al. “When does a physical system compute?” Proceedings. Mathematical, Physical, and Engineering Sciences / The Royal Society 470 (2013): n. pag.
8. Dehghani, Nima and Gianluca Caterina. “Physical computing: a category theoretic perspective on physical computation and system compositionality.” Journal of Physics: Complexity 5 (2022): n. pag.
9. Piccinini, Gualtiero, "Computation in Physical Systems", The Stanford Encyclopedia of Philosophy (Fall 2025 Edition), Edward N. Zalta & Uri Nodelman (eds.), URL = https://plato.stanford.edu/archives/fall2025/entries/computation-physicalsystems/.
10. Lloyd, Seth. "Ultimate physical limits to computation." Nature 406.6799 (2000): 1047-1054.
11. Markov, Igor L. "Limits on fundamental limits to computation." Nature 512.7513 (2014): 147-154.
12. Frisch, Mathias, "Causation in Physics", The Stanford Encyclopedia of Philosophy (Winter 2023 Edition), Edward N. Zalta & Uri Nodelman (eds.), URL = https://plato.stanford.edu/archives/win2023/entries/causation-physics/.
13. Alberts B, Johnson A, Lewis J, et al. Molecular Biology of the Cell. 4th edition. New York: Garland Science; 2002. Available from: https://www.ncbi.nlm.nih.gov/books/NBK21054/
14. Kozma, M.T., Ngo-Vu, H., Rump, M.T. et al. Single cell transcriptomes reveal expression patterns of chemoreceptor genes in olfactory sensory neurons of the Caribbean spiny lobster, Panulirus argus. BMC Genomics 21, 649 (2020). https://doi.org/10.1186/s12864-020-07034-7
15. Wang Y, Yang L. Genomic Evidence for Convergent Molecular Adaptation in Electric Fishes. Genome Biol Evol. 2021 Mar 1;13(3):evab038. doi: 10.1093/gbe/evab038. PMID: 33638979; PMCID: PMC7952227.
16. André M.M. Sousa, Kyle A. Meyer, Gabriel Santpere, Forrest O. Gulden, Nenad Sestan, "Evolution of the Human Nervous System Function, Structure, and Development",Cell, Volume 170, Issue 2, 2017, Pages 226-247
17. Godfrey-Smith, Peter. "Environmental complexity and the evolution of cognition." The evolution of intelligence. Psychology Press, 2013. 223-249.|
18. Erwin Schroedinger, "What Is Life? The Physical Aspect Of The Living Cell", 1944
19. John Odling-Smee "Niche Construction: How Life Contributes to Its Own Evolution" 2025, MIT Press
20. Deutsch, D. (1997). The Fabric of Reality: The Science of Parallel Universes—and Its Implications. New York: Allen Lane
21. Sharma, A., Czégel, D., Lachmann, M. et al. Assembly theory explains and quantifies selection and evolution. Nature 622, 321–328 (2023). https://doi.org/10.1038/s41586-023-06600-9
22. Elhacham, E., Ben-Uri, L., Grozovski, J. et al. Global human-made mass exceeds all living biomass. Nature 588, 442–444 (2020). https://doi.org/10.1038/s41586-020-3010-5
23. Ivan Amato, Speculating in Precious Computronium. Science 253, 856-857(1991). DOI:10.1126/science.253.5022.856
24. Kurzweil, Ray "The Singularity Is Nearer: When We Merge With AI" Viking (2024)
25. Govaert, L., Fronhofer, E.A., Lion, S. et al. "Eco-evolutionary feedbacks—theoretical models and perspectives." arXiv preprint arXiv:1806.07633 (2018). https://arxiv.org/abs/1806.07633
26. Lee, E.D., Flack, J.C., Krakauer, D.C. "Outsourcing Memory Through Niche Construction." arXiv preprint arXiv:2209.00476 (2022). https://arxiv.org/abs/2209.00476
27. y Arcas, Blaise Agüera, et al. "Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction." arXiv preprint arXiv:2406.19108 (2024).
28. Deutsch, David. "Constructor theory." Synthese 190.18 (2013): 4331-4359.