GPT cannot think, but even the toaster has its intelligence

Can GPT think? Reductionists answer yes, anti-reductionists exclude it. However, both agree on one point: the equivalence between intelligence and thinking. Let us question this equivalence and define intelligence as the clever use of reason. A use that, at least since the invention of writing, incorporates technological artefacts and involves our entire body. We can thus recognize the intelligence of the machine, without necessarily attributing to it the ability to think.

[This article was originally published in Italian on Paolo Costa | Media Arte Tecnologia]

The hypostatization of artificial intelligence

The problem with the recent debate on artificial intelligence, in which I occasionally find myself involved, is that it often posits a groundless premise: the idea that artificial intelligence can be abstracted from phenomenal reality and embodied in a reality that exists in itself. A reality in which utopias, nightmares and inconsistent ideological constructs materialize from time to time, which are not matched by the state of things. It is, in short, a form of hypostatization: an epistemological fallacy consisting in associating a mental construct with a magic word — artificial intelligence, in our case — and attributing real existence to that construct. It is thus that artificial intelligence becomes a new form of intelligence (here is the hypostatization), exhibited by machines and different from human intelligence. With respect to this new form of intelligence, of course, everyone feels obliged to take a stand.

The first question prompting the debate is whether artificial intelligence is a true form of intelligence. The answers are heterogeneous. Simplifying, we can recognize four:

1) Yes, artificial intelligence is a new form of intelligence, which has already surpassed or will soon surpass human intelligence, contributing to the advancement of society in many fields.

2) Yes, the artificial one is a new form of intelligence, so powerful that it will turn against its creator, a new Prometheus.

3) No, we are not dealing with a new intelligence, as the artificial one does not express any form of consciousness and therefore awareness of itself and the world, so there is no point in fearing it, speculating that it might escape our control or otherwise challenge human primacy.

4) No, we are not confronted with a new form of intelligence, for the reasons given in point 3); but, because of this, if we want to remain human, we must fight the rhetoric of big tech companies, which are interested in supporting a reductionist perspective — intelligence reduced to calculation — corresponding to the position expressed in point 1).

An operational definition of artificial intelligence

It will displease the enthusiasts of both parties, but I start from a different premise. Re-elaborating Gartner’s official definition, I adopt a perhaps more prosaic interpretation, which does not lend itself to nurturing theologies in one sense or the other, but which seems to me to correspond better to the reality of the facts: artificial intelligence consists — quite simply — of a heterogeneous set of applications in which advanced techniques of analysis and logic, including machine learning, are used to interpret events, support and automate decisions and take actions.

Which implications does such a definition bring with it? To proceed in my reasoning, I need to find a consensus with those who read me on the meaning to be attributed to certain words. In this case, we should perhaps decouple the value of terms that are carelessly often used interchangeably: intelligence and thinking. Recognising that machines incorporate a form of intelligence is not the same as saying that they can think. Unless one pretends to invert the relationship between proxy and world, that is, between that form of intelligence consisting of a set of calculations (see Gartner) and the metaphysical richness of what that set of calculations represents.

The reductionist school, from Leibniz to Turing

The attempt to reduce reason (ratio) to calculation (reor) has very noble fathers. The most important of which is surely Gottfried Leibniz, the great German philosopher who, in Dissertatio de arte combinatoria (1666) came to hope for a world in which mathematical perfection would make all discussion superfluous: ‘when disputes arise between two philosophers, discussion will no longer be necessary, as [it is] between two calculators. It will be sufficient, in fact, that they take up their pens, sit down in front of their abacuses, and (if they like, at the invitation of a friend) say to each other: let us calculate!’ Leibniz’s dream is mirrored in that of many contemporary thinkers and is, after all, the underpinning of the great season of artificial intelligence, inaugurated by Alan Turing’s famous article Computing Machinery and Intelligence (in ‘Mind’, LIX, 236, October 1950, pp. 433–460). But thinking is not only about performing calculations. Human thought also consists of making judgements. It is therefore dispassionate and reflective, based on ethical commitment and responsible action. Thus, while we deny that thinking is reduced to calculations, we exclude that computing machines can think (Brian Cantwell Smith, The Promise of Artificial Intelligence: Reckoning and Judgment, Cambridge MA, The MIT Press, 2019).

Of course, it is not enough to define as reductionist the position of those who exhaust thought in calculation, understood as the determination of a truth based on mathematical operations. It is also necessary to try to clarify what there is that is ulterior, with respect to calculation, in thought. Perhaps this “furtherness” consists in the self-reflexive dimension of thought, in its capacity to question itself, questioning its own foundation and its own limits. This is what, in a very different way, two capital philosophers like Immanuel Kant and Martin Heidegger seem to suggest. Kant does so by embracing a criticist perspective, that is, by establishing a tribunal in which reason is judge of itself: ‘In a species of its knowledge, human reason has the particular fate of being tormented by questions which it cannot reject, because they are imposed on it by its very nature, but to which it is nevertheless unable to give an answer, because they go beyond all power of human reason’ (Kritik der reinen Vernunft, 1781). Heidegger, on the other hand, sees thought as a dynamic process, which we can only understand by continuing to question and test it: ‘Thought itself is a journey. We correspond to this journey only by remaining on the way’ (Was heißt Denken?, 1954)..

Agency without intelligence, or intelligence without thinking?

As for the computational ability of machines, the question remains whether it is — without giving itself away as a thinking ability — a form of intelligence. Luciano Floridi recently formulated a hypothesis that seems to rule it out (AI as Agency Without Intelligence: on ChatGPT, Large Language Models, and Other Generative Models, in «Philosophy and Technology», 36, 15, 2023). The suspicion is that this hypothesis, in turn, postulates the coincidence of intelligence and thought. Since machines cannot be said to think, Floridi seems to suggest, then we must rule out that they are intelligent. On the other hand, machines can do many things in which we are inclined to recognize intellectual capacities of a higher order, emulating human ones (the ability to produce cohesive and coherent linguistic expressions, conversing with a human being, for example). To get out of such a cul-de-sac, Floridi comes up with a brilliant definition: the artificial is a form of “agency without intelligence”. However, as brilliant as it is, such a definition would not be necessary if we accept to decouple the meanings of the expressions intelligence and thinking. We could even rephrase it as follows: artificial intelligence is “intelligence without thinking”.

The fact remains that human beings like to call themselves intelligent. Animal ethology is also quick to recognise intelligent manifestations in other species, which has curiously fuelled a form of racism applied to the non-human world. A friend of mine is willing to eat poultry (because, as is well known, the hen is not an intelligent animal), while he refuses to eat octopus’ meat (whose cognitive abilities are well known: the octopus can solve non-trivial problems, even using tools, loves to play and is endowed with a remarkable memory). Now, what relationship can we establish between artificial intelligence redefined as above and natural intelligence? Clearly, to continue the argument, a hypothesis must also be made about natural intelligence. It — I argue — should be understood as a set of mental processes, but which are always expressed with the involvement of the body and with the aid of some artefact. Intelligence is not something purely mental and therefore ‘inside’. It manifests itself in behaviors that are observable from time to time and therefore lie ‘outside’. Behaviors that are social in nature (one is intelligent with others and through others) and that incorporate objects, artefacts, tools.

Intelligence shows up with technology…

One of the earliest manifestations of the intelligence of the human species was the ability to handle a stone for defense and offence. This was followed by the ability to work that stone, to make it sharper and sharper and thus more suitable for the purpose. Around 5000 years ago, our intelligence evolved through the introduction of two mirror technologies: writing and reading. But the practice of writing and reading — this is the underlying motif of my lectures on the history of the book (in Italian) — is not an “anthropological invariant” (Roger Chartier, Storia della lettura nel mondo occidentale, Roma-Bari, Laterza, 1995). They have been embodied over the centuries in very different experiences from time to time, linked to the use of specific technologies (the clay tablet, the papyrus rotulus, the codex first made of parchment and then of paper, the typographical book, the personal computer, the e-reader, the smartphone), each of which implies a peculiar involvement of the body.

… and with the body

I repeat: writing and reading are not purely mental activities. They are things we do with the body, or — to be more precise — intelligent behaviors that manifest themselves through the coordination and cooperation of two elements: body and technology. When I speak of the body, I am not just alluding to the brain component (the brain, which we can also call mind, if that makes us feel more comfortable). I am really referring to the whole body.

Let’s think about how many parts of the body are involved in the act of writing. And let’s think about how this involvement can vary, depending on the technology employed. I still vividly remember my transition from the first to the second grade, in 1968, which coincided with the school’s decision to equip all children with a biro, in place of the fountain pen to be dipped in the inkwell. For me, being left-handed and therefore somatically disadvantaged, it was a turning point. The biro allowed me to reorganize my mental processes and made me, so to speak, more intelligent. Later, at university, it was the turn to switch to the computer. And it soon became clear to me that writing with the word processor invoked unprecedented body movements and mental processes.

The artifice of writing and reading

It is worth adding that no writing experience can exist in the absence of a technical device. Writing is, strictly speaking, an artifice. Technology is part of this experience, making it possible and at the same time limiting and circumscribing it. So, we ask ourselves: when does the natural experience end and the artificial one begin?

Likewise, reading is also a practice in which we engage our bodies. Italo Calvino dwells on this circumstance, not without a certain irony, in the prologue to If On A Winter’s Night A Traveler (Philadelphia PA, Mariner Books Classics, 1982). The same Calvino, incidentally, who predicted the advent of literary machines. Moreover, reading is a social practice, even if modernity has taught us to consume it in solitude (which was not the case in the ancient world and the Middle Ages). Which is to say that the reader’s intelligence proceeds through another ‘artifice’: the instrumental use of other people’s intelligence. We take advantage of this instrumental use in many, all in all, trivial circumstances. We do it every time we consult an encyclopedia or a dictionary (I will come back to this), a map or a cooking recipe, or rely on the advice of a friend. Intelligence is the shrewd use of reason, which involves the use of some artefact.

A literary theory for robots

This hypothesis seems close to the idea suggested by Dennis Yi Tenen in his provocative essay Literary Theory for Robots. How Computers Learned to Write (New York NY, Norton & Co., 2024). A few years ago, I had the opportunity to appreciate another work by Tenen that is of some interest to those who deal with issues straddling language and computation: Plain Text. The Poetics of Computation (Stanford CA, Stanford University Press, 2017). Also in Literary Theory for Robots, Tenen focuses his reflections on what he calls ‘textual technologies’, which include chatbots, automatic translators, named entry recognition or NER systems (capable of identifying predefined categories of objects within a text corpus), speech recognition technologies, entity-spam filters, search engines, auto-completion functionalities, automatic correctors, text or text summaries generators.

In short, we are in the realm of NLP (natural language processing), which, however broad, excludes many other categories of technologies now classified as artificial intelligence systems. In any case, Tenen introduces his ‘theory’ (we must recognize that, in fact, it is much less than a theory) by proposing an idea of intelligence as ‘artifice’ or ‘skill’. It is the virtue that in Latin is called ars and in ancient Greek τέχνη (téchne). By decoupling intelligence from thought, Tenen achieves a twofold result. On the one hand, he can speak of artificial intelligence without the need to attribute to machines the ability to think. On the other hand, he can also qualify as ‘intelligent’ technologies with a suggestive power undoubtedly inferior to the — almost magical — power of ChatGPT; a power, however, that can be traced back to the same principle: taking upon oneself and thus automating decisions based on calculation.

The intelligence of the automatic transmission

Take, for example, the automatic gearbox that equips most cars today. This is a subsystem that ‘decides’ when to change gear depending on a series of parameters (speed, acceleration, car set-up and road surface conditions) on which it is constantly informed by special sensors. We can say that this is intelligent behavior, in some ways more intelligent than human behavior, without having to postulate that our car’s gearbox is capable of thinking. But let us also consider, trivializing further, the most basic of the many smart appliances with which we fill our homes: the ‘smart’ toaster. It monitors the progressive heating up of slices of bread and ‘decides’ to interrupt this process at the right moment, delivering us fragrant slices without being burnt. Is this not also, if you like, a form of artificial intelligence?

Between metaphor and anthropomorphisation

My insistent use of inverted commas is to mark the somewhat metaphorical value of attributes such as intelligent and predicates such as to decide. In fact, in the introduction to his essay, Tenen himself speaks of ‘intelligence as metaphor’. In short: let machines be intelligent, without imagining that they are pretending to think. It is we who, by anthropomorphizing technology, give it such a claim. My appeal is addressed to three categories of interlocutors with whom I have often been confronted for some time now. There are those who take for granted the ability of machines to think, at least in perspective, assuming that this ability is the inexorable outcome of their intelligence. Then there are those who denigrate artificial intelligence or snub it, since it does not express any form of thought and therefore does not deserve our interest. And finally, there are those who predict dystopian scenarios, in which a deterrent form of thinking — which one? I ask myself — will supplant human thinking.

Zairjah, Ars Magna ans chatbot

Tenen notes how some devices, conceived as early as the Middle Ages to generate ideas mechanically, present interesting conceptual analogies with modern chatbots. This is the case of the zairjah (زايرجة), the wheel described by Ibn Khaldûn in chapter 6, section 28 of his Muqaddimah (مقدّمة, ‘introduction’), one of the most important treatises on the history, geography, political theory and proto sociology of Islamic culture. The work, which dates to 1377, is in English (The Muqaddimah: An Introduction to History — Abridged Edition, edited by Nessim Joseph Dawood, translation by Franz Rosenthal, Princeton NJ, Princeton Classics, 2015). The zairjah was used, according to Ibn Khaldûn, to generate the answer to virtually any question, based on the connections between the letters contained in the text of the question itself. The device found use — again according to Ibn Khaldûn, who does not hide a certain skepticism about it — in divinatory practices.

An evolution of the zairjah, which has had greater influence in Western culture, is the Ars Magna of the Catalan-Majorcan Ramon Llull (1305). Sometimes referred to as the first mechanical calculator, the Ars Magna is in fact a table structure designed to derive — in its author’s words — a ratio (we would say the ‘sense’, but also the ‘quantitative relationship between two dimensions’) from each cell (in Latin, camera). In essence, the Ars Magna is a machine that produces meaningful postulates about the world through a combinatorial mechanism. A bit like ChatGPT, one might say. It is no coincidence that Llull’s ‘invention’ finds an important place in the work of psychologist Bruce H. Hinrichs, Mind as Mosaic: The Robot in the Machine (Minneapolis MN, Ellipse Publishing Company, 2007), who subscribes to the school of those who see the human mind as nothing more than a deterministic machine.

All the magic of a table

More widely, there is no need to profess to be a reductionist to recognize in any table structure, conceptually, a ‘machine’ for thinking. What is the table if not a device that presents certain information in a systematic manner? Every table, Tenen observes, bases its function on limiting the possibilities of interpretation. To this end, it applies a taxonomy, i.e. a scheme that expresses specific relationships between the objects of the world mentioned in it.

Recently, there have been those who have insisted on another ‘intelligent’ property of more advanced, predictor-based language models such as GPT-4: the ability to determine the value of certain mental states — beliefs, intentions, desires, and knowledge — in order to draw conclusions regarding the truth/falsity of specific propositions in which these mental states are expressed (Michal Kosinski, Evaluating Large Language Models in Theory of Mind Tasks, 4 February 2023–17 February 2024, in ‘Arxiv’). This is the same capacity that theory of mind trials attempt to measure in human beings. A capacity intended, in this case, as the ability to attribute mental states to oneself and to others, and to understand that others have mental states that differ from one’s own (David Premack and Guy Woodruff, Does the chimpanzee have a theory of mind?, in ‘Behavioral and Brain Sciences’, special issue: Cognition and Consciousness in Nonhuman Species, 1, 4, December 1978, pp. 515–526). But, again, I see no need to recognize this capacity as a form of thinking, nor — conversely — to reduce thinking to this capacity..

Perhaps the textual technologies that, more than others, help us understand the difference between thinking and intelligent machines are vocabularies and encyclopaedias. On closer inspection, here too we are dealing with technical devices, i.e. ‘artifices’ that enable the automation of certain operations inherent in reasoning. There is, however, a substantial difference between the accumulation of machine data and the encyclopedia. Cloud computing and artificial intelligence base their power on the transfer of knowledge in a virtually unlimited space (even if economic reasons intervene to set a limit). The congestorium (‘granary’) of memory, evoked in the 16th century by the Dominican theologian Johannes Romberch in his handbook Congestorium artificiosae memoriae (1520), fades in the face of the irrepressible dimension of big data. Today’, observes Carlo Ossola, the new president of Treccani, “we are increasingly witnessing a transfer of data, operations, skills and memory from our minds to other bodies outside us, databases, whose governance is delegated” (in “Domenica — Il Sole 24 Ore”, 1 September 2024, III). Another thing is the encyclopedia, which — etymologically — circumscribes, i.e. defines the measure of what is knowable. In fact, the ancient Greek expression ἐγκύκλιος παιδεία (enkyklios paideia) indicates a ‘sum of knowledge gathered in a circle’. This circle includes a space that is limited and to our measure, but for that very reason cannot be delegated.