Artificial Intelligence as an evolutionary phenomenon: scenarios and the case for diversity

The current debate surrounding Artificial Intelligence[1] (AI) is developing along several major lines.

From a sociological perspective, the central question is whether what we are witnessing represents an incremental transformation or, rather, a discontinuity comparable in scale to the Industrial Revolution, electrification or the advent of the Internet. Underlying this question is the crucial issue of the substitution — and complementarity — of human labor, the risk of technological unemployment, and the possibility that AI may become a further driver of capital concentration, perhaps ultimately requiring new forms of redistribution.

From an epistemic perspective, the debate concerns the very meaning of understanding. Does AI actually ‘reason’, or does it merely simulate reasoning? Does intelligence require consciousness? Is there anything specifically human that a machine could never reproduce? This area of discussion also encompasses the profound changes we are witnessing in the production of, and access to, knowledge. AI allows us, in effect, to ask a machine to synthesize the world on our behalf. Yet this possibility inevitably raises a further question: who determines what the machine regards as true?

This epistemic issue leads naturally to a further dimension: that of politics. Who controls AI? Who decides which data may be incorporated into the information used to train these systems? What are the limits of responsibility when tasks are delegated to machines, and what safeguards should be put in place against potentially harmful consequences — including, at the extreme, the loss of human control?

I propose a further dimension of the debate: the etiology of AI (origins and causes). My argument is that this perspective is not merely a contribution to the historical interpretation of AI as a technological phenomenon. It is relevant in its own right and may also provide useful insights into some of the unresolved questions arising in the other areas of debate, particularly the geopolitical one.

A central argument here is that an evolutionary-adaptive pressure acts upon the production and selection of human innovation. Within the domain of chance and discovery, this pressure may favour functional trajectories that enhance human adaptation and, ultimately, evolutionary success.

Rather than asking whether AI was invented ‘by chance’, we might therefore ask a different question: what pressures generated by the technological environment created by humans made it increasingly likely that something resembling AI would eventually emerge?

 

Read also: A warning to humanity or a political dividend?

 

According to Richard Dawkins’ theory of the extended phenotype[2], we may regard Homo sapiens as confronting natural selection not merely through its genomic phenotype but also through the external products of human ingenuity that enhance its capacity to adapt to the environment. These artefacts may also be regarded as instruments through which humans attempt to compensate for their ancestral perception of biological inadequacy — the condition described by Arnold Gehlen as Mängelwesen[3][4].

Let us adopt this evolutionary perspective and ask which changes in the environment exert the strongest pressure upon Homo sapiens to adapt. The purpose of the present argument is to explore whether AI might constitute a human response to precisely such a pressure.

What, then, are the major changes in the environment to which Homo sapiens will inevitably have to respond, and how might AI assist in that process?

 

Complexity as an evolutionary challenge

One property of the contemporary world that represents a major evolutionary challenge for Homo sapiens is complexity.

For the purposes of this argument, complexity can be understood as an emergent property arising from interactions among entities whose individual characteristics may be known, but whose interactions generate new behaviors that cannot be predicted solely from knowledge of the individual components[5].

Consider the world as a collection of increasingly interconnected entities linked through physical, political, economic, cultural and financial interactions. These relationships produce consequences that cannot necessarily be inferred from the characteristics of the individual entities themselves; rather, they arise from the increasingly number of interactions among them and from their strength.

This paradigm of complexity also characterizes natural systems that might otherwise appear to be governed by deterministic dynamics. The Earth’s climate system provides an obvious example. Its constituent subsystems interact through strongly non-linear processes, generating emergent properties and therefore complexity.

One of the great surprises of our age has been the realization that, within this climatic system that has operated for billions of years, a new and increasingly influential entity has emerged: the expanding population and technological development of Homo sapiens. Humanity now interacts strongly with the climate system itself, increasing its complexity and perturbing its dynamics.

 

Read also: The great unlearning: While AI gets better, human intelligence is collapsing

 

We have therefore moved from being merely observers of events occurring in the world to actors capable of modifying the very system we are observing. We are simultaneously spectators and actors on the same stage.

The increasingly strong interactions among the entities and systems created by Homo sapiens have generated effects, behaviors and dynamics that humans are not always able fully to understand or manage. The danger is that humanity may eventually become overwhelmed by the complexity of the world it has itself helped to create.

Here again, the concept of Mängelwesen re-emerges: human inadequacy becomes both a driver of evolutionary pressure and a measure of the effectiveness of the technological artefacts that Homo sapiens is able to develop.

The question therefore becomes: What is complexity, and in what sense might AI constitute a response to humanity’s need to confront it and overcome this difficult evolutionary transition?

 

Requisite variety and adaptive technology

An important argument supporting the need for sufficiently complex mechanisms to govern complex systems can be found in the work of W. Ross Ashby. His Law of Requisite Variety states, broadly speaking, that in order to maintain stability and exercise control over a system or environment, the variety available to the controlling mechanism must be sufficient to match the variety of the system being controlled[6].

This principle should not be interpreted as meaning that a controller must simply be ‘more complex’ than the object it controls. Rather, an effective technological system must possess sufficient capacity to respond to states that may emerge within the complex system.

This is where AI appears to differ significantly from conventional technological systems based primarily on predetermined rules. In a conventional deterministic architecture, a given input produces an output according to rules and instructions explicitly provided to the machine.

The distinctive requirement placed upon AI is not that it should necessarily be more complex than the system it seeks to manage, but that it should be capable of adapting as the complexity of that system changes.

Indeed, the complexity of the controlled system may itself increase partly as a consequence of the intervention of AI. AI therefore becomes part of a feedback loop: it acts upon the system, the system changes, and the changed system presents the AI with a new environment to which it must respond.

One of the defining features of AI is thus its ability, at least in certain architectures, to modify over time the functional relationship through which inputs are associated with outputs. More precisely, this adaptive behaviour is learned and context-dependent[7].

 

AI and evolutionary selection

What risks, then, should concern us as this technology develops?

If we continue to interpret the etiology of AI within an evolutionary framework, we may ask how selection operates in accepting or rejecting modifications to a phenotype — in this case, metaphorically, an extended phenotype — according to their effectiveness in improving or diminishing adaptation to the environment[8].

AI technology already exists in multiple implementations. Different technological systems, based on different LLMs and other architectures, are controlled by different private operators. Although these organisations compete with one another, they largely operate within comparable economic and financial environments. They may therefore be regarded as solutions emerging from a global technological and economic elite whose incentives may privilege the commercial function of AI over its potential evolutionary function.

If, however, evolutionary pressure ultimately contributes to determining the success or failure of these technologies, multiple AI platforms operating in conjunction with Homo sapiens may be regarded as distinct sociotechnical systems, each possessing different capacities to acquire information, construct models of the world, anticipate consequences and select actions.

The analogy with biological evolution is useful here. In bacterial populations, a broad supply of variants expands the space of possible solutions available for exploration and provides natural selection with more opportunities to identify adaptive configurations. Similarly, within the evolving human–AI ecosystem, preserving different cognitive phenotypes may be essential if selection is to compare alternative strategies and retain those that prove most effective under future environmental pressures.

Technology has historically served precisely such a function, allowing human beings to occupy ecological niches in which their biological characteristics alone would have been insufficient[9].

AI, however, represents an intriguing further step. It does not merely extend human strength, mobility or memory. It extends part of our capacity to represent the world and to choose among alternative courses of action.

 

Competing human–AI sociotechnical systems

We may therefore imagine the emergence of numerous sociotechnical configurations, each combining human societies with a different implementation of AI.

Some AI systems may be designed primarily to optimise opportunity or convenience — political acceptability, economic return, institutional convenience or user satisfaction.

Others might instead be designed primarily to optimise epistemic effectiveness — evidential support, predictive accuracy, falsifiability, explicit estimation of uncertainty, and the effectiveness of proposed solutions.

In the short term, an opportunistic system might well outperform an epistemically oriented one in terms of adoption, political support, profitability or public acceptance.

Over longer periods, however, these different classes of sociotechnical systems will be exposed to the consequences of the solutions they generate. Their respective effectiveness will therefore become subject to a form of cultural and technological selection.

Consider a hypothetical financial market in which two AI systems compete to maximize investment returns. AI-A is opportunistically oriented. Its primary objective is to maximize short-term financial performance. It continuously analyzes market prices, trading volumes, investor sentiment, and behavioral patterns, identifying opportunities to exploit temporary market inefficiencies. It does not need to understand the underlying economic reality of the assets it trades. It only needs to predict how other market participants will behave. AI-B is epistemically oriented. Its primary objective is to construct the most accurate possible representation of economic reality. It evaluates companies according to their actual financial conditions, productive capabilities, technological prospects, and long-term economic sustainability. It distinguishes between market prices supported by fundamentals and prices driven by speculation, misinformation, or collective expectations. Suppose that a speculative bubble develops around a group of technology companies. AI-A recognizes that investor enthusiasm is likely to continue for several weeks. Although the companies are substantially overvalued, the system identifies an opportunity to profit from the continuation of the speculative trend. It purchases the relevant assets and may generate substantial short-term returns. AI-B recognizes the same speculative dynamics but also concludes that current valuations are inconsistent with the companies underlying economic fundamentals. If its decision policy prioritizes investments supported by reliable evidence of intrinsic value, it may reduce its exposure or avoid entering the market. Over the following weeks, AI-A significantly outperforms AI-B. This result does not necessarily indicate that AI-A possesses a more accurate representation of economic reality. Rather, it demonstrates that, under certain environmental conditions, an AI system optimized for exploiting existing behavioural dynamics can outperform a system whose decisions are more strongly constrained by epistemic accuracy. The distinction becomes particularly important when market participants collectively respond to expectations rather than fundamentals. In such environments, accurately predicting what others will believe or do can be more immediately profitable than accurately identifying what is economically true. However, the advantage may be temporary. If the speculative bubble collapses, AI-B’s more accurate assessment of fundamental value may support greater resilience, while AI-A’s performance will depend on its ability to recognize the reversal and adjust its positions before substantial losses occur. Importantly, an epistemically oriented AI could also recognize speculative dynamics and profit from them. The example therefore assumes that AI-B’s epistemic orientation influences not only its beliefs but also its decision-making objectives and constraints.

 

The example presented above illustrates a broader principle concerning the evolution of AI systems.

Selection pressures operating over short time horizons may favour systems that exploit their environments more effectively, rather than systems that represent those environments more truthfully. If competitive success is measured exclusively through immediate rewards, opportunistically oriented AI systems may enjoy a selective advantage even when their behaviour is less closely aligned with the pursuit of objective knowledge.

Over longer time horizons, however, environments characterized by uncertainty, structural change, and unexpected disturbances may increasingly reward the ability to construct accurate causal models of reality. These different classes of sociotechnical systems will be exposed to the consequences of the solutions they generate. Their respective effectiveness will therefore become subject to a form of cultural and technological selection.

The central question is therefore not whether opportunistic intelligence is inherently superior to epistemic intelligence, but under which environmental conditions, and over which time horizons, the pursuit of truth becomes an adaptive advantage rather than a competitive constraint.

As AI becomes increasingly integrated into human cognition and decision-making, different human–AI configurations may thus be regarded, metaphorically, as competing extended cognitive phenotypes.

Their long-term persistence will depend not only upon their social, political or economic acceptability but, ultimately, upon the effectiveness of the solutions they generate when confronted with the constraints of the real world.

Crucially, the feedback provided by reality is not indefinitely negotiable. A proposed solution must eventually encounter physical, biological, environmental or social constraints, and its effectiveness — or failure — will provide a criterion by which the sociotechnical configuration that generated it can itself be assessed.

 

Europe and the preservation of epistemic diversity: a proposal

Europe could play a particularly important role in this context.

By 2026, the European Union has established 19 AI Factories linked to EuroHPC supercomputers and is developing AI Gigafactories. These infrastructures are accessible not only to industry but also to researchers, universities and public authorities. The EU has also amended the EuroHPC Regulation to support next-generation AI infrastructure and large foundation and generative models.

Although the current European strategy is explicitly concerned with competitiveness, productivity, industrial innovation, technological sovereignty and democratic values, Europe could give one or more of its sociotechnical AI implementations an explicitly epistemic mission.

While the current development of frontier AI is largely driven by industrial competition in the United States and by a combination of private enterprise and state-coordinated strategic investment in China, Europe could pursue a distinctive path by launching a multinational, publicly funded initiative explicitly dedicated to the development of epistemically oriented AI. Drawing on its experience in organizing ambitious international scientific enterprises, such as CERN, Europe could address the evolutionary challenge of developing AI systems whose primary objective is not merely competitive advantage or immediate economic performance, but the advancement of reliable knowledge and the capacity to confront humanity’s most complex scientific and societal challenges.

 

Read also: Europe should bet on Original Intelligence

 

Such an initiative might be conceived as a “European Public Epistemic AI Infrastructure”.

This would be a publicly governed AI system whose primary objective was neither profit, political consensus nor the maximisation of user engagement, but rather the production of the best achievable representation of the current state of knowledge, explicitly accompanied by uncertainty estimates, sources, competing hypotheses and mechanisms for verification.

The argument is not that such an implementation would necessarily prove superior to all others. Its importance would lie instead in its existence as a distinct evolutionary variant.

Maintaining such a variant would increase the diversity of the sociotechnical configurations that Homo sapiens introduces into an environment in which different human–AI systems will effectively undergo processes of cultural and technological selection.

An ecosystem in which all major AI systems were exposed to broadly similar economic, political and commercial pressures would reduce the variety available to selection. A publicly governed and epistemically independent AI would introduce a substantially different variant, governed by a different objective function.

We cannot know a priori which variant will ultimately prove most effective.

That uncertainty is precisely the reason why such a variant should be created and preserved.

AI systems optimised primarily for short-term political or economic advantage might enjoy an immediate selective advantage while nevertheless proving less effective over longer time horizons.

Maintaining at least one AI ecosystem whose principal objective is epistemic reliability, rather than profit, political expediency or user engagement, would preserve an alternative phenotype upon which cultural and technological selection could act. A publicly governed AI oriented towards epistemic reliability would therefore possess evolutionary value precisely because it would preserve a variant that an ecosystem dominated by similar commercial or political pressures might otherwise eliminate.

Europe could therefore undertake such an evolutionary experiment by establishing an independent, publicly governed AI infrastructure centred upon evidence, falsifiability, uncertainty management and effectiveness.

The rationale for a European public AI would consequently be not merely ethical, economic or geopolitical. It could also be understood as a strategy for preserving evolutionary diversity within the human cognitive ecosystem.

In an increasingly complex environment, reducing all human cognitive extensions to the same objective function — whether commercial profitability, political expediency or another short-term criterion — could itself constitute a systemic risk.

The diversity of human–AI cognitive architectures would then play a role analogous to diversity within an adaptive system: we cannot know in advance which configuration will prove capable of meeting future selective pressures; maintaining genuinely different variants therefore enlarges the space of possible adaptive responses.

 

 

---

Footnotes:

[1] In this article, the term Artificial Intelligence (AI) refers to the broad and evolving family of computational systems capable of learning from data, constructing representations of their environment, generating predictions and other outputs, reasoning, and performing goal-directed actions, encompassing currently deployed technologies—including large language models (LLMs), multimodal foundation models, reasoning models, and agentic AI systems—as well as future generations of increasingly capable AI architectures endowed with greater autonomy, adaptability, and potentially novel emergent capabilities.

[2] Dawkins, R. (1982). The Extended Phenotype: The Long Reach of the Gene. Oxford University Press.

[3] Gehlen, A. (1940). Der Mensch. Seine Natur und seine Stellung in der Welt. Junker und Dünnhaupt, Berlin.

[4] Bronowski, J. (1973). The Ascent of Man. Little, Brown and Co.

[5] Bridgman, P. (1927). The Logic of Modern Physics. Macmillan, New York.

[6] Ashby, W. Ross (1964). An Introduction to Cybernetics. Methuen, London.

[7] Mathis, M. W. (2026). ‘Leveraging insights from neuroscience to build adaptive artificial intelligence’. Nature Neuroscience, 29, 13–24.

[8] Here, rather than using extended phenotype in its strict biological sense, the term sociotechnical community/system may be used to describe the close coupling between humans and a particular AI implementation, characterised by the objective function towards which that implementation is oriented.

[9] Derex, M. & Boyd, R. (2015). ‘The foundations of the human cultural niche’. Nature Communications, 6, 8398.