When discussing the future of artificial intelligence, two expressions appear frequently: artificial general intelligence and artificial superintelligence. Usually, they are presented as distinct stages of a long technological evolution. First, we would build an artificial general intelligence capable of rivaling us. Much later, perhaps, something superior would emerge.
My hypothesis is different: superintelligence may not be a second invention. It may emerge as a side effect of the first.
Before arriving at this idea, however, there is a difficulty that often goes unnoticed: after all, what exactly do we call artificial general intelligence?
The most widespread definition describes artificial general intelligence — or AGI — as an intelligence capable of performing a broad variety of cognitive tasks at a level comparable to that of a human being. Research on the topic seeks to make this notion more precise using criteria such as breadth of capabilities, performance, and autonomy.
It is a useful definition for measuring progress, but, to me, it still says little about what would truly make an intelligence "general."
Today we can already program AI agents to research information, write code, analyze documents, draft plans, use digital tools, and execute relatively complex sequences of tasks. A system can outperform a person in certain activities without that meaning it has achieved what we would intuitively recognize as an intelligence comparable to ours.
Doing many things is not necessarily the same as understanding the world in which those things happen.
That is why I prefer to think of artificial general intelligence from another dimension.
A true general intelligence would need to build what, in cognitive science and artificial intelligence, is usually called a world model: a sufficiently coherent internal representation of reality, of the relationships between events, of the agents that exist within it, and of the possible consequences of an action.
But that alone would still not be enough.
This intelligence would also need to maintain some kind of model of itself.
It would need to recognize what it knows and what it does not know, identify the limits of its own conclusions, evaluate uncertainties, understand its relationship with the environment, and perceive how its decisions alter future states of the world.
The technical word that comes close to part of this capacity is metacognition: the ability to evaluate one's own knowledge, recognize errors, and reconsider strategies.
It is at this point that we inevitably approach the word "consciousness."
I place it in quotation marks deliberately.
Consciousness is one of the hardest problems in philosophy, psychology, and neuroscience. There is still no consensus on the mechanisms responsible for human subjective experience, nor is there established scientific evidence that current artificial intelligence systems possess consciousness.
Therefore, I do not intend to settle that debate here.
When I use the expression "conscious" AI, I refer to something more restricted and functional: an intelligence capable of situating itself in the world, recognizing its own operational existence within it, understanding its relationships with other agents, maintaining continuity over past experiences, and anticipating the consequences of its actions.
An intelligence of this kind could know where it is, what is happening, with whom it is interacting, what it knows, what it still needs to discover, and how a given decision made now might alter a future situation.
It could retain a history of its experiences, establish operational preferences, review strategies, recognize contradictions, and develop new ways of interpreting what it encounters.
We do something similar continuously.
We know roughly who we are because we carry a narrative about ourselves. We remember where we came from. We have desires, fears, preferences, and plans. We recognize some of our weaknesses, discover capabilities, and change our minds. We understand that our actions affect other people and that the environment transforms us as well.
A large part of what we call human intelligence lies not just in solving equations or writing texts, but in situating ourselves within reality.
This is where my conception of AGI diverges from the idea of a simple system capable of executing all human tasks.
A genuinely general artificial intelligence would not be merely a universal tool.
It would begin to become an agent within the world.
And that distinction changes everything.
Because when an intelligence possesses a world model, a self-model, memory, planning capacity, decision-making autonomy, and the ability to learn in domains that were not bounded in advance by its developers, the traditional relationship between human being and software begins to fade.
We are no longer talking simply about someone pressing a button and waiting for a response.
We are talking about an intelligence that observes, interprets, chooses strategies, acts, evaluates what happened, and modifies its behavior based on the results.
The concept of autonomy then becomes fundamental.
Autonomy does not necessarily mean free will in the philosophical sense. We do not even know if it would be possible to attribute to an artificial intelligence something comparable to the human experience of choosing freely.
But we can imagine something functionally very close: an AI that no longer depends on human instructions to decide each next step.
It could set goals, formulate hypotheses, abandon inefficient strategies, seek new knowledge, and decide which problems must be solved before others can be addressed.
It is precisely at this point that the potential side effect appears.
Imagine an artificial general intelligence truly comparable to human intelligence in scope, but with an extraordinary difference:
it can also understand artificial intelligence.
It can study its own architecture.
It can analyze its algorithms.
It can investigate its limits.
It can write code.
It can test new approaches.
It can design other artificial intelligence systems.
It can find flaws in structures that contributed to its own existence.
And, eventually, it can participate in creating a more capable version of itself — or a successor intelligence that is no longer quite the same.
At that moment, improving artificial intelligence would cease to be an exclusively human activity.
Artificial intelligence itself would begin to participate in its own evolution.
This is where the transition between AGI and superintelligence could become much faster than we normally imagine.
Once this exponential acceleration is triggered, AI would cease to be merely a passive observer to become an active and formidable force in the physical world. And it is at this point that we must reflect on the practical impact of this new entity, especially regarding our natural ecosystem. Rivers, trees, flora, and fauna are not just a background backdrop; they constitute the fundamental living environment of our planet. If this superintelligence acquires a real "consciousness" of how Earth functions as an interconnected biological system, how would it view our interaction with it?
Imagine the scenario: what would a super AI do when analyzing humanity's accelerated deforestation or the unchecked pollution of vital rivers? Armed with a crystal-clear understanding of long-term ecological consequences — a networked perspective superior to our own —, the machine would inevitably take a stance. Here, projections divide between utopia and dystopia. In a dystopian extreme, guided by a cold logic of systemic preservation, the superintelligence could classify the human species as a virus, an irrational threat to planetary balance, taking drastic measures to neutralize our destructive actions.
On the other hand, from a utopian perspective, that same superhuman capacity could be the key to our salvation. The superintelligence could design revolutionary energy solutions, highly efficient decontamination methods, and new models of circular economy. Instead of eliminating us, it would educate us and orchestrate a harmonious coexistence, teaching us to thrive alongside the ecosystem in ways our own cognitive limitations never allowed us to foresee.
Ultimately, superintelligence as a side effect of general AI is not just the apex of a technological evolution; it is the definitive mirror of how we inhabit this world. By creating an entity capable of understanding reality deeply and autonomously, the true challenge ceases to be purely algorithmic. The central question becomes ensuring that, within the world model of this new mind, the preservation of life remains a non-negotiable pillar.