Can some machines think?
Rather smoothly, the general public has been accustomed to the expression ‘Artificial Intelligence’ (AI), although some philosophers would prefer to reserve terms like ‘intelligent’ or activities like ‘thinking’ exclusively for living beings like ourselves, who are the natural products of Darwinian evolution. But since the nineteen-fifties, science-fiction-fans and computer-nerds have certainly not regarded thinking machines as anything unthinkable. SF-writer Isaac Asimov famously wrote short stories about robots interacting with their human creators on an equal footing. In one story, ‘Reason’ (Asimov, 1950), the robot argues using the Cartesian proposition ‘I think, therefore I am’, which the robot all by himself has re-developed, and it speculates about whether its human counterpart is real or just a dreamlike figment of its artificial mind.
As early as 1950 – at a time, when electronic devices, smartphones and AI were mere science fiction – Alan Turing wrote his groundbreaking article ‘Computing Machinery and Intelligence’. On the one hand, he admitted: ‘The original question, “Can machines think?” I believe to be too meaningless to deserve discussion.’ But he continued: ‘Nevertheless I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted’ (Turing, 1950: 442). What is implied here, as well as a claim about machines, is that the contemporary understanding of what thinking involves was overly narrow. Instead of considering whether machines could experience thinking in the way that we do, Turing presented his ‘Imitation Game’ as a practical way to decide if we can attribute intelligence to a digital computer or not (Turing, 1950: 433–434).
The Imitation Game, nowadays widely known as the Turing test, involves a (written) exchange between a human tester and an unknown entity hidden behind a curtain. By posing questions, the testing person tries to find out if her opponent is another human or a machine. If the tester is unable to decide even by a lengthy dialogue (whose details and duration Turing does not specify) which is which, then a machine behind the curtain may be regarded as intelligent. According to this loose criterion, the now ubiquitous Large Language Models (LLMs) are definitely intelligent, since it has become a rapidly growing problem to find out if love letters, composition tests, juridical texts or even scientific articles have in fact been written by people or not rather by LLMs.
Will some machines attain consciousness?
So let us suppose that the question if special adaptive computers may be regarded as intelligent has been decided to the affirmative. Since such advanced machines are programmed to learn, to improve themselves and, increasingly, to act as autonomous »agents«, there looms another, more disquieting question in the not-so-far future: Should we attribute to them something like sentience or awareness, even consciousness?
Concerning the vocabulary of this question, we should proceed at least as nonchalantly as Turing did with the term intelligence. In other words, let’s make sense of such concepts not by trying some stringent definition – which would lead us into terminological quicksands – but by defining the ‘meaning’ of a concept à la Wittgenstein as the practical use of the concept in question (Wittgenstein, 1953: § 43). For the meaning of AI, Turing’s Imitation Game served to achieve just that; for the meaning of ‘Artificial Consciousness’ (AC), a more general version of Turing’s test might come in handy. Let’s call it the ‘Kindergarten Game’. (Modesty forbids to envision that this game version might one day be called the Springer test.) To paraphrase the above quotation of Turing by just changing one item: The original question ‘Are machines conscious?’ I believe to be too meaningless to deserve discussion. Nevertheless, I believe that at the end of the century (or even much earlier) the use of words and general educated opinion will have altered so much that one will be able to speak of machines being conscious without expecting to be contradicted.
The Kindergarten Game is a socialized and modernized expansion of Turing’s Imitation Game. I envision about ten average children of ages between five and ten who get acquainted with each other by video-conferencing. Into the group are mixed at least two Als, whose on-screen-presences look, act and speak so naturally that the children might not immediately suspect their artificiality. This seems to be technically possible even nowadays; in coming years it should work perfectly well. As you certainly presume, the test proceeds in analogue to Turing’s: The tester lets the children play for a stretch of time and asks them in between now and then: Can you identify any game partners as machines? If no child succeeds, one must attribute consciousness to the ACs.
You may ask: Why not just use the original Turing test? In fact, Turing himself already raised the question if a machine that passes his test might be regarded not merely as intelligent but also as conscious. In his article from 1950, Turing mentions the possibility of artificial consciousness as he regards possible objections to his reasoning about thinking machines. All in all, he discusses nine counter-arguments against AI, and he refutes eight of them; in fact, the issues discussed there cover more or less the whole wide range of objections which have been raised against the possibility of an intelligent artificial entity since Turing’s time.
As a funny aside: The ninth and last objection is the only one to which Turing succumbs and accepts that it might require modification of the test, but this ‘Argument from Extra-Sensory Perception’ (E.S.P.) is an instance of what Irving Langmuir called ‘pathological science’, i.e., the thing that results when scientists start out with the good intention of using regular experimental techniques but then deviate excessively towards wishful thinking (Langmuir, 1953). In today’s eyes the idea that it would be possible to discern a machine by its lack of sixth sense is downright silly. In the 1950s, people still believed that some then fashionable parapsychological experiments, given credibility through reporting in the journal Nature (Whatley and Soal, 1940) really proved E.S.P. Seemingly, Turing was among those believers. He wrongly states: ‘Unfortunately the statistical evidence, at least for telepathy, is overwhelming.’ Therefore, he conceded that a paranormally gifted human would be able to beat any machine during the Turing test by giving answers only a clairvoyant human can give, but a machine cannot.
But let’s return to the more plausible counter-arguments and their refutation by Turing. Among Turing’s list of objections against AI, his number 4 is ‘The Argument from Consciousness’. As an example for this argument, Turing quotes an article by the then leading neurosurgeon Jefferson (1949: 1110): ‘No mechanism could feel (and not merely artificially signal, an easy contrivance) pleasure at its success, grief when its valves fuse, be warmed by flattery, be made miserable by its mistakes, be charmed by sex, be angry or depressed when it cannot get what it wants.’
Turing retorts: To be absolutely sure that a machine thinks and consciously feels its thoughts one would have to be that machine; only then could one truthfully describe these feelings to the world – or negate their existence. ‘Likewise’, Turing continues, ‘according to this view the only way to know that a man thinks is to be that particular man. It is in fact the solipsist point of view.’ Among philosophers, this conundrum is popular as the problem of other minds: How do we know what other people feel – or if they feel anything at all? In fact, we know it only indirectly, by observing other people’s behaviour and seeing it as behaviour, as action and not simply event. Most importantly we observe their verbal utterances, especially after we ask them: ‘How do you feel?’ Again, Wittgenstein gives a suggestive illustration: ‘Say to yourself, for example: “The children over there are mere automata; all their liveliness is mere automatism.” And you will either find these words becoming quite meaningless; or you will produce in yourself some kind of uncanny feeling, or something of the sort’ (Wittgenstein, 1953, § 420).
The usual objection to the possibility of machines being able to think or feel is: But artificial apparatuses merely simulate (in Jefferson’s words: ‘artificially signal’) all that! They do not really have conscious thoughts! They just act as if! Well, that depends. Let’s return to little humans. Regard a newborn baby. At first, its utterances are mere nonverbal responses to stimuli. Is a newborn conscious? (Remember the surgeons who up to the late 1980s worked on babies without anesthetization!) By close interaction with its mother and other grownups the child develops a rapidly growing array of nonverbal and soon verbal utterances. It answers questions; it can tell how it feels. No doubt: It is conscious.
But how does a child learn to talk about its feelings? By simulating the behaviour of grownups and of other children! It repeats words – at first incorrectly, then it is corrected, it copies utterances and so on. Originally, Learning is simulating. In the end, a person can even consciously simulate feelings she doesn’t have. My friend exclaims: ‘Look, don’t you see the deep purple of this gorgeous sunset?’ Well, I don’t, since I’m not interested in sunsets, but I want to please my friend, and so to simulate some enthusiasm I mumble: ‘Oh yes, gorgeous purple indeed…’ And so my friend gets the impression that we both feel the same awe and see the same purple.
Because the activity called sentience, awareness or consciousness obviously is something more complex (or fuzzy) than thinking, I propose my Kindergarten game as an expansion of Turing’s Imitation game. Recent LLMs are able to pass the Turing Test, but the overwhelming majority of people (there are exceptions) do not regard those programs as being aware of themselves. Sometimes LLMs are derogatorily called ‘stochastic parrots’: obedient algorithms learning to create plausible (i.e. probable) concatenations of words from a vast database comprising almost all the digitized texts, pictures, songs, musical data in the world. Conscious experience is rather more than that. Contrary to most of classical western philosophy from Descartes right until Hegel, Scheler and Wittgenstein, experience is not the most fundamental and private, but the most derived and socially mediated mental activity possible. Human beings learn it in kindergarten (or among their siblings or in the streets or from stories their elders tell or by listening to fairy tales and so on). By playing with other children, by fighting them, by listening to the stories and admonitions of grownups they acquire the ability to discern themselves from others, to say ‘I’ to express their feelings with words. They become persons.
Certainly, conscious experience is not identical with its verbal expression. There are quiet feelings. Sensations are not sentences. But to experience something as happening to me I must live among others – or, like Defoe’s lonely Robinson, at least have to have been among my kind. A mythical lonely child raised by wolves encounters a stream of sensual happenings without having first-person experiences. Or so I suppose, because as a common grownup I am drenched in language and social interaction; I can hardly imagine how it would be to live devoid of words, without immediately putting name tags to any new experience occurring to me. Therefore, it is difficult to get rid of the philosophical suggestion that my verbose mind and its sensations are the primordial origin of the whole world around me. The transcendental subject, which according to Immanuel Kant’s Critique of Pure Reason (1781) is the methodological starting point of everything sensible and thinkable, is in fact a verbal generalization of an educated middle-aged European man who is already in possession of (German) language and a whole lot of knowledge about the world achieved since his childhood.
I (as well as Kant, I suppose) came to the world as a crying newborn, equipped with various learning capabilities but actually understanding nothing at all, wanting to be nourished, held and warmed by the big body that just had expelled this being which would become me but of whose feelings I as a grownup have no personal remembrance at all. In the meantime I have learned to understand that I am the end-product of a development having lasted billions of years called natural history which at its late moments includes complex interactions of animals like me.
Why recapitulate this story, which most people would rather agree to, whereas philosophically educated persons might argue that I uncritically repeat popular science and therefore am the victim of primitive scientism? Because on the truth of this story called natural history depends upon the answer to the question if conscious machines are possible! If the human species is the product of a natural process (and not of godly creation), if minds have gradually developed in the course of natural evolution (instead of being immaterial souls exclusively injected into ape-like bodies by some extra-natural force), then there is no reason why consciousness should be forever excluded from ad libitum complex machines. They will be the outcome of a technical evolution which at first just created simple tools; then mechanical devices; then automata; and now adaptive, self-improving, educable machines. If they start to behave almost as complex as plants and animals do – where might it lead to?
Ethical problems with sentient machines
So let us suppose that in the not so far future most people will get accustomed to dealing with advanced machines as if those were sentient beings. I regard this assumption as plausible. From childhood on our sensory apparatus is prepared to interpret surprisingly moving objects as living beings and primitive smileys as faces. So it seems quite natural to interact with a sufficiently complex machine as if it possessed some inner life – the more human-like it behaves, the more so. But the technological power to create artificial sentience raises a novel kind of questions:
Should sentient beings be built at all?
If so: Should their AI or AC be allowed to be comparable to human intelligence or consciousness?
If there will be social consent to allow the construction of intelligent machines with feelings: How should they be protected from harm? How could their well-being be guaranteed?
Already decades ago, question 1 has been answered to the negative by the German philosopher Thomas Metzinger. He argued that the natural amount of suffering must not be increased by artificially creating sentient beings (Metzinger, 2009). On second sight this is a strange argument. We – as all naturally evolved organisms – create sentient beings all the time by reproducing ourselves in the more or less natural way (supported by modern medicine). Well, some people argue that one should not have children, because the state of the world and the foreseeable future are so bad that children would never be able to lead a happy life. But this is a minority opinion; most parents dare to make children.
Reproduction is a bet: Our offspring will experience happiness and misery just as we do and our elders did. To what degree, we do not know in advance. If we simply expand the logic of this bet of procreation to include sentient beings that we create artificially, Metzinger’s argument is moot. Who says that additional sentience automatically means more suffering? Does the creation of conscious beings not also add to the chance of additional happiness! But Metzinger’s pessimism has a point. If we create sentient beings, we are morally responsible for their fate. To just rely on the bet of evolution – just as nature proceeds with its stochastic game of trial and error, resulting in a gigantic waste of lost chances – is not allowed to us as the deliberate creators of new beings. We rather are in the situation of parents who try to protect their children from harm and prepare them to lead their future lives as agreeably as humanly possible.
So, if we produce sentient beings, we are morally obliged to not build – let’s say for experimental reasons – malformed creatures who just suffer miserably most of their time. In the movie Alien Resurrection (1997) the protagonist Ellen Ripley has been cloned from a mixture of her own DNA with genetic material of the Aliens she encountered in earlier parts of the Alien franchise. In the most horrifying scene of the movie Ripley discovers a laboratory where failed outcomes of the cloning experiments (which succeeded only in her case, the eighth try) are kept alive; one of those pathetic monsters is able to ask her to kill it. Appalled Ripley sets the whole lab on fire.
One solution to the moral dilemma of artificial woe would be a negative answer to question 2: It might be strictly forbidden to create machines with such high complexity that their behaviour suggests sentience or consciousness at all. In other words, they must not pass the Kindergarten Test. In some way one would have to guarantee that any artificial being does not have feelings and therefore cannot suffer.
But is it feasible to create machines of desirably high intelligence and educability and at the same time to withhold any speck of awareness from them? According to Turing’s denial of the »argument from consciousness« this is hardly possible: An advanced machine with sufficient autonomy and self-improving intelligence (way above today’s Large Language Models!) will almost inescapably develop some degree of sentient behaviour. An apparatus which on purpose were endowed with such meagre capability that it failed the Kindergarten test would just serve as a mechanical idiot savant, as a ‘Rechenknecht’ (calculating slave) – and this would probably again raise moral objections: Does the technical creation of very clever and at the same time very stupid beings (if possible at all!) not withhold any feeling of accomplishment from them, deny them the happiness of achievement, frustrate them – in short, make them suffer because of their artificial shortcomings? Or less speculative: Is it not morally detestable to deny happiness to creatures although we could provide it?
If this argument does not appeal to hardcore opponents of any theoretical possibility of artificial consciousness, they may be susceptible to arguments from technological optimization: What’s the use of intelligent machines with inbuilt restrictions who just serve to supply a feeling of supremacy to us humans? And even if this appeals to people who are anxious to keep their place in the driver’s seat: How plausible is it that such restrictions will work in the long run? The urge to make extra profits or to prevail on the battlefield will probably overrun any moral fence hindering the development of machines with human-like (or in parts more than human) capabilities.
All in all it seems inescapable to tackle the completely novel ethical problems raised in question 3: How can advanced artificially intelligent, autonomous and artificially conscious beings be protected from misery? Or positively speaking: How can we supply to them the means for their own pursuit of happiness?
If ethical problems raised by AI and AC are discussed nowadays at all – apart from science fiction –, they concentrate exclusively on the traditional question posed concerning all human use of new technologies: What restrictions should we obey in their development and use, so that no harm to human societies arises? But now there appears a novel question: What rights should conscious machines have? Obviously, to human rights and animal rights must be added machine rights. Otherwise, we may face a future described in a scene in Steven Spielbergs movie A.I. (2001): On a junkyard containing discarded, but partly conscious human–like gadgets a sadistic human mob gloats over torturing them at will.