Recently, the world has been shaken by a hitherto unknown fever, its name: Artificial Intelligence or AI. Given the clever answers that a program like ChatGPT gives to arbitrary questions within seconds, the collective excitement is understandable. Some people even believe they are talking to more than merely an intelligent machine; they imagine they are communicating with a compassionate human being. Yuval Noah Harari sees an apocalyptic time dawning where we will all be puppets of artificial intelligence.
Let’s leave aside for the moment the deceptive impression that AI may have genuine feelings. As far as intelligence is concerned, our amazement at its almost limitless knowledge is quite justified. ChatGPT as well as any other program of a similar kind is in principle able to access the knowledge of all mankind as it has been stored in written form (as well as in pictures or sounds). It is obvious that no single brain can retrieve more than a comparatively infinitesimally small section of this collective treasure. In this respect, artificial intelligence is an achievement that almost infinitely surpasses natural intelligence in two different ways. Firstly, by the fact that it may absorb the knowledge of all living humans (provided that it has been stored); secondly, by the fact that this knowledge can be increased by the factor time almost arbitrarily, namely by the knowledge of earlier generations (in so far as it was stored in any shape).
Both advances constitute nothing less than a quantum leap compared to natural intelligence. When asking ChatGPT an elementary question like „Why do people enjoy flowers?“, the program evaluates all stored sources with regard to where flowers are mentioned in connection with enjoyment relying at the same time on a wide range of pattern recognition – just like the human brain. It has to distinguish between question and answer, between relevant and less relevant answers, which it furthermore divides according to various aspects: aesthetic, symbolic, cultural (flowers as a gift), connection to nature, positive effect on people, etc. If ChatGPT would access exclusively texts of other than European cultures or even the traditions of early tribal cultures, then other answers and other categorizations would result. If one were to feed the system mainly with the utterances of lunatics and conspirators, one would again get a completely different picture. For questions about the political relevance of flowers and beauty, we would certainly get different answers depending on whether the artificial intelligence was programmed in Russia, China or North Korea or in Germany, South Korea or the USA. Answers from the mid-19th century or Luther’s time would still reveal a different weighting. Artificial intelligence is therefore no more „objective“ than natural intelligence, since it is programmed by people of a certain time and space. The weighting of answers tells us a lot about how much it is influenced by the spirit of the times and cultural orientation.
An exception to this rule are informations about repeatable events, which are the subject of natural sciences. Here objectivity consists in the fact that we receive the same answers to the same – experimentally verifiable – questions. We will come back to this point.
Artificial intelligence not only imitates, but is productive in its turn. It is, for instance, capable of producing new verses in the style of Goethe or Heine, in which – say – thyme or fantasy flowers are requried to figure prominently. Depending on the level of development of the artificial intelligence, these attempts are still rather clumsy or already so successful that even experts are unable to distinguish the newly created verses from existing original ones. In principle, the program proceeds just like any living poet who successfully imitates the style of a dead luminary. Due to their extensive knowledge of the original texts, both know exactly which phrases and words to use and which not. Every successful forgery presupposes this knowledge. When artificial intelligence imitates the writings, the words, the images or even compositions of a living or dead individual, it becomes a perfect forgery program. Perfect means that the individual imitated in writing, words, images or composition could have produced the subject in question in exactly the same way. This is, of course, no mere imitation, because the style in question is applied to new elements (such as thyme flowers or new themes). The ability of AI to actively produce all kinds of fakes makes it an immediate danger – the focus of current discussions worldwide.
If artificial intelligence is so close to natural intelligence, that is, if it is capable not only of imitation but of creative production, doesn’t this statement amount to saying that natural human intelligence could finally be replaced by artificial intelligence? The first impression in dealing with ChatGPT seems indeed to suggest just this conclusion, especially since it is explicitly confirmed by many of its inventors and propagandists.
And yet it would be completely wrong. True, we cannot deny that artificial intelligence is far superior to natural intelligence, because individuals can never store more than a vanishingly small fraction of total available knowledge. But the opposite is equally true: artificial intelligence is hopelessly inferior to natural intelligence in terms of quality – and this difference is just as irreversible.
Let us assume that ChatGPT will soon be perfected to such an extent that the program will not only be able – as is already the case – to produce rather clumsy verses in the style of Goethe or Schiller on any arbitrarily chosen topic, but that these verses will finally be indistinguishable from the historically transmitted ones. Is the „pseudo-Goethe“ created in this way then not simply a second Goethe, who could completely take the place of the first? And would this not equally apply to a pseudo-Mozart and a pseudo-Picasso, if artificial systems perfectly imitate and replace their products?
This question gives rise to another and still more far-reaching one. Can’t these artificial producers eventually take the place of humans altogether and replace their natural intelligence? I just said that among artificial intelligence enthusiasts, many are ready to claim just that.
There is a definite answer to this question, and it is negative. True, a perfect artificial intelligence could create a pseudo-Goethe whose verses would be indistinguishable from those of the master, because any existing sign structure (whether consisting of words, sounds, images, etc.) may be exactly re-produced in the original style even when containing new elements. The only possibility of distinction then lies in the historical proof that Goethe never wrote these verses during his lifetime. When fake videos of politicians expressing opinions fabricated by artificial intelligence are published, the proof to the contrary can likewise only be provided by historical evidence that the politicians in question never actually spoke those words at that time and place. Such a proof is, however, very difficult and often impossible for events consisting of the statements and actions of individuals – that is, for ninety-nine percent of all unique events – because there are either no eyewitnesses or only a limited number of them.
On the symbolic level, a perfect forgery cannot be invalidated in principle; it can only be invalidated by confronting it with something completely different: namely reality. But reality outside our own living and working space is beyond our sphere of direct experience. We are therefore dependent on the re-presentation of reality by fotos, videos etc., i.e. on the symbolic level, on which it is shown, for example, that a certain politician spoke these or those words at a specific time and place. Since we do not and will never know the real event, as it is unique, we have no means of distinguishing a perfect forgery from a genuine reproduction of the event. Artificial intelligence thus becomes an instrument for the production of fakes, i.e. of images of a reality that never existed but could exist.
This is nothing new; natural intelligence has always been able to do the same. Artificial intelligence only imitates an ability that has existed for millennia, namely as long as artists invented fictitious realities. But in this case we always knew that they were fictions. That is no longer the case with AI fakes. Whenever we are unable to relate to reality through direct experience, deciding between true and false becomes impossible. In this respect, artificial intelligence confronts us with a historically unique development.
This trait reveals the destructive potential of artificial intelligence. It remains ineffective only in those areas where we are dealing with events that can be repeated at will. If some housewife from Kenya claims she boiled water at 10 degrees Celsius, or a shaman insists that ingesting garlic enables people like him to levitate, these and similar claims can be disproved by experimentation. Unique events may be falsified, because we are not able to expose the fake by confronting it with reality. Repeatable events are immune against such falsification. Therefore, artificial intelligence will celebrate its greatest and indisputable successes in the field of science.
Thus, we may safely predict that artificial intelligence will give an enormous boost to the sciences and at the same time destroy the moral cosmos, because there true and false play a dominant role. But don’t think that harm and benefit cancel each other out. The right way of life, i.e. the moral cosmos, is much more important for individuals as well as for entire nations than is right knowledge. Ultimately, the latter is meant to promote the former.
Once more: Can artificial intelligence replace its natural counterpart? This question has already been answered in the negative, but we can now substantiate the verdict. Artificial intelligence lacks the socket of reality through which natural intelligence comes into being in the first place. This difference is fundamental.
How does natural intelligence come into being? When a child grows up in the linguistic cosmos of a particular time and culture, his brain is not simply fed with the data of that language and culture, as happens in artificial intelligence when a programmer stores the available knowledge in electronic chips. Children as well as adults proceed in a completely different way. Their sense organs provide them with visual, acoustic, emotional images of reality and to these they assign the symbolic languages existing in the respective culture. In other words, natural intelligence wraps the material provided by the sense organs in an existing cultural garment, which however it constantly expands in response to new realities, sometimes tearing it apart when this garment no longer fits reality. The dialectic between individual sensual perception of reality and its collective re-presentation is a constant, never-ending process – the outstanding characteristic of natural intelligence.
Artificial intelligence lacks this process because it has no organs for recognizing reality – no eyes, no ears, no organ of touch, no emotional experience. In other words, it lacks the primary dimension constituting natural intelligence: input from reality. All human knowledge ultimately rests on this pedestal. When ChatGPT tells us surprisingly clever things about why flowers mean so much to humans or why it is morally reprehensible to hurt other people – and does so in words that we ourselves could perhaps not formulate any better – then we are intuitively tempted to ascribe to the machine the same ability to recognize reality as we do to ourselves. And this is exactly where we are mistaken. People appear authentic to us when we get the impression that they speak from personal experience, which they then communicate on the symbolic level by means of words (or other signs). But we are also receptive to the perfection of the symbolic level itself. If a person in the face of an overwhelmingly beautiful landscape only brings „Ahs!“ and “Ohs!” to his lips, then we may consider this to be very authentic; we concede that he or she may experience reality as intensely as we do, but it touches us much more when a great poet or painter helps beauty to find expression in a perfect way either through words, images or musical sounds. In this case, the perfect word or image appears so authentic that we take the corresponding experience for granted.
Artificial intelligence makes use of this illusion when it presents us with the most impressive ones among the almost infinite number of symbolic evocations. In truth, the machine feels, thinks, sees, hears absolutely nothing, while its algorithms spit out their infinite binary sequences of 000s and 11111s, each of which it translates into the words and phrases of a historical language or composes as an image. ChatGPT has not acquired its almost infinite knowledge by dealing with reality like any child when it constructs the realm of signs on this basis. Its only reference is the realm of signs itself. In other words, it evaluates stored binary sequences of 0’s and 1’s to generate binary sequences according to algorithms specified by the programs‘ inventors..
This answers the question of „empathy“ raised at the beginning. The machine feels absolutely nothing, because it lacks the emotional base of experiencing reality. This shortcoming leads to a further deficiency. By evaluating the existing sources AI may well create a pseudo-Goethe, a pseudo-Schiller or a pseudo-Picasso, but never a future great poet, painter or composer, who distinguishes himself by creating new symbolic languages on the basis of some new view of reality. Artificial intelligence can only imitate and recombine existing material, but it cannot create new symbolic worlds emerging from the struggle with reality, as man has done continuously in the course of his historical existence and will continue to do in the future. This limitation of artificial intelligence is a basic fact.*1*
Some AI enthusiasts will protest this statement. In fact, there are numerous examples in robotics of exploiting signals from reality. Let’s take a very simple example. A built-in temperature sensor ensures that an „artificial waiter“ that takes orders from diners and then distributes the food to them not only evaluates visual stimuli to serve the right tables but is also temperature-sensitive because it is supposed to avoid an oven. In principle, a voice program could still be added to the machine so that the artificial waiter could answer guests‘ questions sensibly. Moreover, his optical sensors could also let him correctly estimate the age and gender of the guests and then modify his answers to the point of apparent „empathy“ when dealing with children. These and many other sensors for reality mapping are conceivable and are already in use today. Yet none of these machines feels, thinks, senses anything of the reality about which it often speaks to us so „touchingly“. The machine is as dead as a stone, although some among us will intuitively attribute life to it (seniors in nursing homes will be especially inclined to do so, as these facilities usually suffer from staff shortages that they compensate for with talking dolls and similar “intelligent” devices ).
This leads to the old Frankenstein question: Does a machine eventually come alive when the sensors scanning reality are constantly expanded so as to asymptotically approach the pedestal of optical, acoustic, olfactory, haptic, and emotional signals that every normal living human being calls his or her own? The components for approaching this goal are at our disposition, and there is nothing to prevent us from perfecting them over time. Will it be possible in this way to mimic living man more and more perfectly, so that at some point the change from a dead machine into a living being will necessrily take place ? Or, in other words: Will such a machine suddenly develop consciousness?
Given our previous definition of natural intelligence as the ability to create a symbolic representation of reality above the pedestal of experienced reality, such a possibility could not be ruled out – namely, if we succeeded in making the artificial sensory apparatus more and more like the natural one of humans. But this consideration makes us realize immediately that our definition of natural intelligence is grossly inadequate. There are living beings – the whole vegetal kingdom belongs to it – which do not show any consciousness in the way of human beings, because this would not make biological sense. A tree cannot choose its own location, so it does not need a sensory system to determine its optimal location. The reaction to environmental signals is therefore reduced to elementary processes such as orientation towards light. Consciousness makes biological sense only for living beings actively intervening in their environment, provided that their actions have a positive or negative effect on them. But even in this case, unconscious, merely instinctual reactions are possible if they can be reduced to yes-no alternatives, such as when approaching a fire. Only when the environment confronts a living being with complex stimuli does consciousness emerge as the ability to make choices between correspondingly complex courses of action. The stimuli of reward for right behavior and punishment for wrong behavior (pleasure and pain) thus play a central role. Without them, there would be no consciousness because it would be biologically superfluous.
Consciousness is, furthermore, not identical with knowledge. Even the most primitive organisms must have an evolutionarily acquired knowledge programmed into them in order for them to be adapted to their particular environment. Consciousness is indeed much more than mere knowledge. In its current shape, artificial intelligence possesses almost unlimited knowledge and yet is no more than a dead machine. In order to have consciousness, the machine would have to be able to feel pain and pleasure like natural beings and not only – as is the case now – merely imitate them on the symbolic level. But this only makes biological sense if the machine can avoid pain on the basis of its own decisions and seek pleasure in the same way.
Conclusion: AI machines amaze us by speaking with tongues of angels and devils, and yet they are emotionally dead, incapable of experiencing reality – and that’s the way it will probably stay.*2* Apocalyptic predictions such as Harari’s, according to which artificial intelligence will soon dominate and subjugate mankind, may confidently be dismissed as sensation mongering. But its verifiable effects are far-reaching enough. On the one hand artificial intelligence may turn out to be in the moral cosmos the counterpart to the atomic bomb in the physical: It has the potential to lastingly destroy the moral cosmos – with all foreseeable consequences for politics and society. On the other hand, it will certainly give a great boost to science and industry.
*1* Linguistics suffers particularly from this limitation. It meanwhile produces – almost – perfect translation programs, but all transfers from one language to another are mere trans-formations: the linguistic form of a language A is transferred into the linguistic form of a language B by applying appropriate algorithms. The fact that natural intelligence proceeds in a completely different way (as illustrated above by the example of infantile learning) is ignored. Natural intelligence first translates the reality input into concepts (thus, out of the continuum of electromagnetic light waves it creates individual colors like red, yellow, etc. – I speak of meaning). The immaterial substrate of meaning is then materialized in form (i.e. in sounds, gestures or other signs – I speak of the „realization of meaning by form“).
The idea of emulating the procedure of natural intelligence in linguistics, that is, of incorporating the socket of reality, is not in itself new. The great Danish linguist Otto Jespersen already spoke of a „notional grammar.“ Steven Pinker explicitly states that a prelinguistic meaning underlies all language, calling it “mentalese” and defining it as „The hypothetical ‚language of thought,‘ or representation of concepts and propositions in the brain in which ideas, including the meanings of words and sentences, are couched.“
The program of conceiving languages – all languages – as different formal realizations of an – in essence general – structure of meaning and analyzing them accordingly, has however so far been conceived only as an idea. I have tried to elaborate this idea systematically. See The Principles of Language – Towards trans-Chomskyan Linguistics.
This procedure offers no practical advantages – the existing translation machines are, as said, quite advanced. But it offers the only possibility to understand why and how the variety of existing and possible languages comes about and which regularities it is subject to. The interest of the linguistic science I founded is of a theoretical nature.
*2* I do not see any possibility to exclude the development of the machine to a living being in principle – already now biogenetics can manipulate the building blocks of evolution to produce artificial beings. If, however, we were to succeed in perfecting these interventions in such a way that we thereby create artificial beings which not only possess an almost infinite knowledge but also evaluate experience and develop a consciousness, then it would be inevitable that they would perceive themselves as different from us (and, in addition, far superior). We would have reason to fear such beings even more than our own kind – and we are already dangerous enough for each other. By its ability to arbitrary falsify unique events, artificial intelligence is already now an acute danger. Humans could not do anything more stupid than to create a competition for themselves, which increases this danger immeasurably – just as Mary Shelley and many horror films have invoked it for a long time already.