Types of Artificial Intelligence:
What Are ANI, AGI and ASI ?
Not everyone knows that there are three distinct types of Artificial Intelligence: ANI, AGI and ASI, each corresponding to a stage of its cognitive development. Only the first exists today, while the other two remain, for now, purely theoretical. Telling them apart means understanding where we stand and where we are headed. Let us explore their characteristics, differences and what the future holds.
( by: Antonio Maria Guerra | date: 20/03/2026 )
ANI: the ‘narrow’ artificial intelligence operating in our present.
Narrow Artificial Intelligence, also known by its acronym ANI (Artificial Narrow Intelligence), accounts for virtually all AI currently in existence. Contrary to what one might imagine, this technology is decidedly limited: it excels at a single, well-defined task (hence the label ‘narrow’), failing completely when pushed beyond its ‘comfort zone’. Think of it as an ultra-specialized professional who has spent an entire career mastering just one skill.
A concrete example: an ANI system designed for facial recognition — such as those found in many modern smartphones — is remarkably effective at its particular function. It analyzes geometries, lighting patterns and microexpressions, cross-referencing them against a vast database in real time. Impressive, no doubt. Yet if that same ‘specialized’ system were asked to perform a simple translation, it would not know where to begin.
Make no mistake, this is not a matter of insufficient processing power, but of an intrinsic, structural limitation. Not truly ‘understanding’ anything in the human sense of the word, these systems have no overarching picture of the world around them and cannot draw connections that go beyond their designated domain.
They process data. They detect patterns. They generate statistically consistent outputs. Nothing more.
The flip side is that precisely this extraordinary proficiency, call it ‘productive blindness’, makes ANI remarkably powerful. Just ask the ‘old-school’ stock traders replaced by algorithms capable of simultaneously monitoring thousands of transactions, efficiently identifying risks and opportunities in a fraction of a second.
It should also be noted that the many ‘expressions’ of narrow artificial intelligence now in use have grown so adept at their tasks that they have become, in a sense, ‘invisible’. We no longer notice their presence. Yet every time Netflix recommends a show or Spotify assembles a personalized playlist, ANI is quietly at work behind the scenes. Unseen, yet ever-present.
The structural limitation of ANI.
Sophisticated as they are, to the point of being almost unsettling, ANI systems conceal a structural fragility as profound as it is paradoxical. Unable to reason laterally, generalize or improvise, they operate exclusively within predefined boundaries: any variation falling outside their training domain is enough to throw them into crisis.
A practical example: an ANI-based algorithm trained to recognize cats in photographs, when presented with one wearing a small outfit, may fail to identify it, mistaking it for something else entirely. No human being would make this mistake, since we possess the ‘concept of cat’ — we know what it is, regardless of how it looks.
The paradox is straightforward: the extreme specialization that makes these systems so effective also renders them potentially vulnerable to anything falling outside their designated domain, making them, ultimately, wholly unfit to navigate the unpredictability of the real world autonomously.
Is the Turing Test still reliable?
In 1950, Alan Turing proposed what would become known as the ‘Turing Test’ (*1), according to which, if a machine could sustain a conversation indistinguishable from that of a human, it would have demonstrated in practice that it was intelligent. At the time, it seemed as simple a criterion as it was effective.
The ‘problem’ is that, today, even models that do not yet constitute an AGI, such as GPT-4, can pass the test with ease: a large proportion of people genuinely cannot tell they are conversing with a machine. And yet GPT-4 is not an AGI, that is, a ‘generalist’ system, but a ‘mere’ ANI: a system that, however sophisticated and specialised within its own domain (text generation), does not possess an authentic understanding of what it says, at least not in the sense in which we understand the term.
Read more
This has led the scientific community to reconsider the Turing Test as a means of establishing autonomous reasoning and intelligence, relegating it, at most, to the measurement of a machine’s conversational ability. It’s likely that the next benchmark for the most advanced forms of Artificial Intelligence (AGI and ASI) will no longer hinge on the question “does it seem human?”, but rather on “can it do what we cannot, while explaining the reasoning that led it there?”, thereby ruling out mere imitation, and allowing us to learn from it in turn.
Nota:
*1: Originally conceived as the ‘Imitation Game’.
AGI: the artificial intelligence that will ‘understand’.
Let’s be clear from the beginning: when we talk about Artificial General Intelligence, or AGI, we are discussing about something that, at least officially, does not yet exist.
Where ANI excels, but solely within a single domain, AGI, thanks to an unprecedented cognitive flexibility, will be able to transfer knowledge from one field to another, tackling any challenge with versatility and awareness, learning from it and applying what it has grasped to entirely unfamiliar situations.
Once again, the central issue will not be computational power, relevant as it may be, but the architecture of the system itself: in the case of AGI, this will need to be endowed with a cognitive awareness capable of grasping not only the ‘meaning’ of a word, but also the underlying ‘concept’, such as ‘justice’, ‘beauty’, and so forth.
Right now, major companies are pouring billions of dollars into discovering how to push their AI beyond current limits. OpenAI, Anthropic, Alphabet, Meta and many others are pursuing dozens of different directions simultaneously, hoping that at least one will lead to AGI. Forecasts have grown increasingly bold: some predict the goal could be reached as early as 2027–2029 — though in this field, a degree of caution is never misplaced.
ASI: when the machine will surpass humanity.
If AGI will represent, in a sense, a level of ‘parity’ with human intelligence, ASI (Artificial Superintelligence) will possess a cognitive level radically superior to it and, for the time being, barely conceivable. Make no mistake: this will not be so much a quantitative leap as a ‘qualitative’ one.
Thanks to its extraordinary capabilities, ASI will be able to reason at a level of abstraction and complexity simply beyond the reach of the human mind. And this is precisely where the real issue lies: the danger posed by such an advanced system will not be its potential ‘malevolence’, but something far more subtle and unsettling, an unbridgeable ‘distance’ from its own creator. Not a deliberate threat, then, but something akin to what a human being feels towards an ant crossing their path: total indifference.
This type of artificial intelligence remains entirely hypothetical for now, but its theorization is not a purely academic exercise — it points to the direction in which research is heading and the questions that will sooner or later demand an answer.
When did people start talking about ‘types of artificial intelligence’?
Although John McCarthy, an American mathematician and computer scientist, had coined the term ‘Artificial Intelligence’ in 1956 at the landmark Dartmouth Conference, the distinction between its types emerged almost thirty years later, through a fascinating journey intertwining technology and philosophy. It was in fact a philosopher, John Searle, who in the 1980s devised the thought experiment known as the ‘Chinese Room’ (*1), through which he highlighted the difference between simulating understanding and genuinely possessing it,
Read more
sparking a heated debate that led to the theoretical distinction between ‘weak’ AI (which simulates intelligence) and ‘strong’ AI (endowed with an authentic one). These reflections laid the conceptual foundations for the classification that would follow.
It was Ben Goertzel, together with Shane Legg and Peter Voss, who formalised the contemporary tripartition into ANI, AGI and ASI in the early 2000s.
Nick Bostrom made a decisive contribution to the debate on ASI with his landmark essay ‘Superintelligence’ (2014), highlighting its potential risks, while Goertzel continued to develop concrete methodological approaches towards achieving AGI.
Note:
*1: The ‘Chinese Room’ is a thought experiment in which a person locked in a room receives written questions in Chinese and, by following a set of instructions, returns correct answers — without understanding a single word of the language. The argument: a computer does exactly the same thing, processing symbols and producing apparently intelligent outputs, but without any real comprehension.
The risks of a ‘truly’ intelligent Artificial Intelligence.
One of the most debated topics in the AI research community concerns the question of control: the ‘alignment problem’. Scientists are essentially asking how we can ensure that an AGI (should one ever be built) acts in accordance with human values and interests. Unlike narrow AI systems, which are relatively straightforward to oversee precisely because they are confined to the domain for which they were trained, an autonomous AGI could develop goals of its own (*1), pursuing them with cold efficiency while remaining wholly indifferent to the consequences.
Read more
This would not amount to malice, but rather to a kind of algorithmic indifference. A fitting analogy is an avalanche: capable of causing as much destruction as a deliberate act of violence, yet entirely devoid of intent. Institutions such as the Future of Life Institute and the Center for AI Safety have begun raising concerns on the matter, advocating for rigorous and systematic research into the technical and ethical alignment of emerging technologies. Some researchers go so far as to call for moratoriums on certain classes of models until a clearer framework for their governance can be established. A position, however, that is far from universally held: scholars such as Yann LeCun and Andrew Ng consider these risks overstated, arguing that attention should instead focus on the concrete challenges posed by AI systems already in deployment.
Note:
*1: More precisely, behaviours emerging from poorly specified objective functions.
LLMs: a first step towards Artificial General Intelligence (AGI)?
In recent years, researchers have observed intriguing phenomena in the behaviour of ‘LLMs’, Large Language Models such as ChatGPT, Gemini, Claude, and others. It turns out that, every so often, these systems, classified as ANI, and therefore structurally ‘limited’, are nonetheless capable of applying concepts acquired in one context to an entirely new task. A rudimentary capacity for ‘generalisation’ thus appears to be emerging: not spontaneously, but as a consequence of the growing scale of these models and the sheer vastness of the data on which they are trained.
Read more
Beyond this, the most recent models are beginning to display a form of ‘reasoning’, solving problems in ways their creators never explicitly programmed. Caution is warranted, of course: what we may be witnessing could be nothing more than sophisticated ‘pattern matching’ (that is, the ability to recombine patterns already encountered during training, without any genuine understanding). And yet, the very fact that such things occur at all gives cause for optimism about the path being taken towards the development of AGI.
‘ELIZA Effect’: the first illusion of genuine understanding.
Between 1964 and 1966, Joseph Weizenbaum at MIT created ELIZA, a program based on simple ‘pattern matching’ rules that responded to statements such as “I feel sad” by reflecting them back questions such as “Why do you feel sad?”. Despite its remarkable simplicity, many people developed a genuine emotional attachment to ELIZA, convinced that it could truly understand their problems and reluctant to give it up, treating it as a kind of ‘virtual confessional’.
Read more
Weizenbaum was deeply troubled by this phenomenon: the so-called ‘ELIZA Effect’ had been born, laying bare the human tendency to project intelligence onto simple systems. Today, with the advent of sophisticated assistants such as ChatGPT, we continue to anthropomorphise machines, mistaking computational processing for genuine understanding: the historical lesson being that we must learn to distinguish between appearing intelligent and truly being so.
The global race to AGI: who are the contenders, and what is the prize?
The pursuit of the first AGI has become a full-blown competition. The most prominent American players are OpenAI, Google (Alphabet), Anthropic and Meta: companies that can draw on teams of thousands of world-class researchers, as well as staggering operational budgets. Consider, for instance, that the infrastructure initiative known as ‘Stargate’ (*1) has announced projected investments of up to 500 billion dollars over the long term, with an initial confirmed commitment of 100 billion. The Chinese government, for its part, is deploying equally vast resources, operating both through direct state funding and by leveraging private giants such as Baidu, Alibaba, Huawei and DeepSeek, in a bid to lead a field it rightly regards as strategically paramount. The European Union is attempting to keep pace, albeit by funding research rather cautiously and focusing primarily on regulatory frameworks, such as the AI Act (which entered into force in August 2024), designed to curb the risks associated with the new technology: a choice that positions it more as a regulatory power than as an industrial protagonist in the race.
The prize awaiting whoever develops the first AGI will be of extraordinary consequence: namely, the potential to shape humanity’s trajectory throughout the twenty-first century, across multiple dimensions:
- Economic supremacy, by dominating the dynamics of global markets;
- Military supremacy, by holding a decisive strategic advantage on the battlefield;
- Technological supremacy, through an exponential acceleration of scientific progress.
It is hardly surprising, then, that Russian President Vladimir Putin declared, as early as 2017: “Whoever becomes the leader in this field will become the ruler of the world”.
Note:
*1: A joint initiative led by OpenAI, SoftBank and Oracle.
‘Reasonable’ predictions on the evolution of Artificial Intelligence.
It is fair to assume that, in the short term (that is, over the next five years), ANI will continue its tendency to permeate every aspect of society, becoming ever more sophisticated and ubiquitous. We will witness significant advances in its application to robotics, medicine and automation systems. At the same time, debates on the regulation and governance of new technologies will multiply: it will be an absolutely unavoidable necessity. In the medium term (5–15 years), should progress maintain its current pace, we may begin to witness the first ‘stirrings’ of AGI: systems that will start to display genuine generalisation capabilities, an evolution considered plausible by some researchers, and premature by others. In the long term, certainties will grow scarce: it appears probable that ASI, Superintelligence, will be brought into being by an AGI rather than by human hands, with all that such a development may entail.
Copyright information.
The images on this page were created using generative Artificial Intelligence tools.