We live amazingly well, even without answers to the first questions of humanity. So well that questions about where from, where to, and purpose hardly play a role in our attention. Unless, of course, our everyday responses are no longer on par with non-everyday stresses and strains, and we have to admit gaps and voids where we need life-world orientation for new challenges. The current success of generative artificial intelligence, such as the ChatGPT dialog system, produces such a constellation, changes the agenda, and thus creates a discursive urgency.
John Hopfield, who was awarded the Nobel Prize in Physics in 2024 for his fundamental work on artificial neural networks, explained during his acceptance speech that he was motivated by the question of how the mind arises from the brain. For Hopfield, this is the deepest question posed by our humanity – the answer is yet to be found. However, artificial neural networks have been used to boost the performance of AI applications since 2010, driving radical innovations that shed new light on unquestioned human assumptions. Since machine and human performance have increasingly overlapped in relevant applications, the question "What is a human being?" has once again come into focus, as has the question "What is the fundamental difference between humans and machines?" for the first time.
The technical approach of artificial neural networks is inspired by the observable electrochemical processing in brains. Brains are organs found in many living beings, in all mammals, including humans. An essential ability attributed to the brain is the concept of causality as a connection between cause and effect. Natural causality refers to effects that necessarily occur in the objectively real environment when a cause occurs. Over the past 5,000 and especially the past 500 years, humans have developed an escalating ability to grasp natural causalities in the form of natural laws and have continuously increased their degree of generalization and depth of description.
In the 18th century, the two philosophers David Hume and Immanuel Kant developed different concepts for deriving causality – derived from data and experience (Hume, 1748) or as an innate concept of necessity (Kant, 1781). Arthur Schopenhauer contradicted both. He pointed to animals' fight or hunting behavior, which could not be explained without an instinctive "understanding" of causality. Dealing with natural causality thus becomes a defining characteristic of the brain of a living being and not just of the thinking mind. In contrast to the brain, however, natural causality is not a defining property inherent in integrated circuits. Instead, it is a performance that software delivers when models are integrated that explicitly represent the physical relationships. Silicon chips are not electron brains because then they would have causality. They are also not electronic minds because then they could use language for reasoning. This is neither quibbling nor AI defeatism but tool realism.
In humans, we speak of the mind, and since ancient times, it has been believed that thinking is specific to humans and distinguishes them from other mammals and non-human primates. Language is essential for human thinking. Animals do not have language in the sense of symbolic interaction, which is based on the ability to use signs with identical meanings. One can go as far as Johann Gottfried Herder, who once said: "Without language, man has no reason, and without reason, no language." Or like Wilhelm von Humboldt, who remarked: "Language is the formative organ of thought."
Without words, there is no language; without meaning, there are no thoughts; without thinking, there is no mind. Although the concept of language used in the currently successful applications of generative AI has led to high-performance AI systems, it is under-complex and one-dimensionally focused on the linguistic surface. It obscures the insight that intelligence is not a statistical phenomenon, that language production is more than the output of probable word sequences – and its dominance hinders the development of solutions that enable a human-machine knowledge discourse in the first place.
But isn't it the case that the astounding growth in quality over the past ten years has been and continues to be driven by the ongoing increase in computing power, the amount of data available, and language models with billions of parameters? This is not controversial. But it becomes problematic in the form of the scaling hypothesis, which assumes that performance will continue to grow in proportion to the amount of data and computing power so that only a resource and investment problem needs to be addressed until machine intelligence as artificial general intelligence completely overtakes human intelligence. The scaling hypothesis is the subject of controversial debate in current AI research. An interim report on "Advanced AI Safety" published in the middle of last year states: "Would continued ‘scaling up’ and refining existing techniques yield rapid progress, or is this approach fundamentally limited, and will unpredictable research breakthroughs be required to substantially advance general-purpose AI abilities? Those who think research breakthroughs are required often think that recent progress hasn’t overcome fundamental challenges like common sense reasoning and flexible world models. "
Artificial intelligence, understood as the digitalization of human knowledge skills, must offer a comprehensible idea of what human nature and "common sense" are. Only on this conceptual basis can fundamental limits and restrictions be addressed, which can – and cannot – be overcome by scaling up human, financial, and technical resources. It is not about storytelling but rather about tool realism, not about stock market hype, but rather about sustainable basic assumptions and social innovation maturity. Knowledge skills are about knowledge and abilities. In a broad sense, knowledge means cognitive, cultural, and social, but also technical, procedural, technical, and historical knowledge. Skills mean that knowledge is applied as know-how in a deliberate, planned, and practical way in order to achieve goals. These can be personal goals, but also those of the community, and in such a way that the actor is able to set aside some in favor of others in a balancing process.
Artificial neural networks enable the machine to identify patterns and classify states and processes. The ability to assign arbitrary objects to classes can help identify something familiar or draw attention to anomalies. In both cases, the information density in the interaction with the environment increases. Machine assistance systems can support and improve the efficiency and quality of diagnostics in medical imaging procedures. Commercially available AI applications can assist, but they cannot be used for discussion. They are not suitable for applications that involve knowledge acquisition or knowledge enhancement. Pattern recognition can segment speech in an acoustic stream, assign words to signal segments, and enable dictation systems that convert spoken language into written text. The result for the user is a gain in efficiency but not in knowledge. Transcription can help to classify existing or missing relevance better, but nothing new is created mentally when the spoken words are on display.
It is essential that users can immediately and competently assess whether the speech recognition has worked reliably. Machine translation should only be used if you can assess the target language translation result's adequacy. Generative AI precisely addresses the gap between passive and active knowledge skills. Most people understand a foreign language better than they can speak it. This may seem unsatisfactory given the pithy headlines, but it is worth a lot, is full of opportunities, and supports work. Regarding knowledge interaction, you should not assume that you can acquire new knowledge when interacting with current chatbots. Chatbots can only be used as "teachers" if learners already have the knowledge and simply want to update their own knowledge.
Chatbots based on language models and artificial neural networks enable powerful assistance systems with an astounding range and quality of applications. However, they are miserable discussion partners in their own right. We grant human discourse partners a leap of faith. If someone can name reliable sources when asked and justify a statement in a comprehensible way, the claim can become the basis for a directional decision. The factual correctness of chatbot outputs is possible but not reliable. The machine-generated strings of words are not trustworthy statements.
A knowledge discourse is defined by the exchange of statements, structurally presupposes freedom from domination or hierarchy, relevance in the content of the contributions, truthfulness among the participants, and the possibility of mutual adoption of perspectives. Statements can be expressed as assertions for which speakers claim approval and are intellectually or possibly contractually liable. The correctness of an assertion is proven by reference to comprehensible, objective real-world facts. Discourse is a process that aims to achieve a constructive result through dialog. Ideally, this is guided by the idea that the unconstrained compulsion of the better argument can lead to agreement and a shared, common conviction that guides action.
In order for us to benefit socially and culturally from AI, we must try to promote machines as "partners" in knowledge discourses without ignoring the fundamental human-machine differences. On the machine side, these differences include an inability to feel, need, and bond, a lack of perspective-taking, and social, political, and cultural immaturity. All information is always without guarantee. But we can trust machines with the disinterested "view from nowhere." The problem is reliability. One possible solution is to hybridize the system architecture so that the large language models of generative AI are used for the task segmentation of the input and the linguistic interface of the output. In contrast, symbolic AI and knowledge graphs are used for inference, deduction, and verification.
We should not forget that Hopfield's initial question cannot yet be answered. Although we do not know how the self and its consciousness emerge from the organic-bodily foundations, we can explore the ego and its emergence in the history of ideas, always keeping in mind the prospects for the opportunities and limits of machine intelligence.
Since Descartes, everyday political and cultural life in the West has been dominated by a dogmatic dualism. Antonio Damasio described this "separation of the most sophisticated operations of the mind from the structure and functioning of a biological organism" as Descartes' error. Characteristic of the dualistic approach of subject philosophy is the self-development of the atomic ego on the basis of its transcendental talents; for Kant, these include the understanding of necessity, substance as that which remains, and space and time as the pure forms of sensory perception. The empirical ego of subject philosophy finds itself from the inside out into the world, which confronts it as the other, the alien, and as resistance, or serves as a means for achieving its own ends.
We see Descartes' modern philosophy and the Enlightenment of Hume, Kant, and Fichte as almost contemporary. Moreover, we use the terms subject and object in everyday language as opposites. This front-forming fundamental opposition was overcome when the social and language-bound process of the development of the ego was experimentally reflected upon. Anthropologist Michael Tomasello describes the infant's developmental phase, which is crucial for the formation of the ego, as the "nine-month revolution" in the process of becoming human: The ego emerges first and through language-bound social interaction. Language is the leading medium of socialization and not a superficial phenomenon; it is the catalyst that transforms the preconscious into the conscious ego. The further development of AI should be inspired by this paradigm shift so that the knowledge-centered human-machine discourse does not remain stuck in irrelevant probability.
The linguistic turn takes the medium of language seriously. It is not limited to an innovative method of interpretation but opens a fundamentally new perspective on the development of the self. Habermas sees the history of philosophy as a succession of three paradigms: "Accordingly, metaphysics is first replaced by the philosophy of the subject, and this in turn by the philosophy of language." The linguistic turn transcends subject philosophy in that it sees the development process of the ego anew and positions it in linguistic interaction. This means that language, reasons, and meaning are essential to accept a counterpart as an interlocutor. Machines can only fulfill this requirement if they no longer produce output based on probabilities but on the basis of reasons.
With the linguistic turn, the subject is understood as a fundamentally socially and linguistically constructed identity. This arises in early childhood interaction because the infant spends the first few months of its life in intensive interaction with its caregivers, who continuously view and address it as an "I." Multimodally with touches, caresses, and gazes. The others challenge the child to be an "I" and thus bring about its becoming. The ego of the linguistic turn emerges from the outside in. In retrospect, the empirical ego has no conscious access to its preconscious social process of emergence. Its social root is abstract. The ego experiences its desires very concretely. It encounters itself in its stream of consciousness as an actor with bodily needs for air, water, and food. It desires closeness and community, affirmation, recognition and exchange, reliability, security, and freedom.
The socially and culturally problematic consequence of the subject-philosophical dualism is the formation of fronts and the resulting friction and hardening. The ego of needs becomes the self-opinionated, demanding subject of claims. In the dualistic paradigm, the others are not the midwives of the self but rather the troublesome worldly resistance to be overcome, which limits the fulfillment of a desire perceived as justified and stands in the way as an obstacle.
Language and social interaction are the "parents" of our ego. The ego concept of the linguistic turn changes the view of the root of human individuality. This contains the ego, socialized linguistically in every fiber of its tissue, as the self of self-consciousness. Every person brings this potential with them - and no other known being. The fabric is formed by being together with others. Togetherness attracts, challenges, socializes, and shapes the self. The others always address us as "you" so that the subject cannot avoid discovering itself as an "I." This means language is the medium of socialization.
An AI approach that trivializes the role of language to a collection and output of word strings misses the opportunities at the moment of their technical feasibility. Language is not just a communication tool. It pulls us by the hair from a preconscious state into the human world of symbolic interaction.
How do AI applications based on large language models and artificial neural networks fit in with the concept of the linguistically socialized subject of the linguistic turn? The role of language is dominant in both cases. Large language models depict the linguistic surface and the use of words in word chains. They are data-driven, work probabilistically, and the outputs are euphonious and usually syntactically correct. A chatbot makes no claims; it only generates output and is, therefore, not a conversation partner. But the outputs – and this creates a social danger – can be mistaken for assertions by the users. The statements may be accurate, but there is no guarantee or reliable inference. If comments such as "thinking process" are displayed during processing, then this serves to entertain the user and enhance the user experience. However, it is actually an act of disinformation because the AI system does not think.
The linguistic turn in machine intelligence means no longer treating language as a surface phenomenon but taking it seriously as a tool. Only then will we be able to construct machines that not only process language but also understand it, that not only generate strings of words but can also justify statements in a comprehensible and transparent manner. Only then will a human-machine discourse be conceivable in which solutions to knowledge problems can emerge co-creatively. Knowledge and cognitive problems are abundant.
Reinhard Karger is a theoretical linguist who has been employed since 1993, has been the company spokesman since 2011, and has been a member of the Supervisory Board of the German Research Center for Artificial Intelligence (DFKI) since 2022.
Original article in German:
https://www.faz.net/aktuell/wirtschaft/kuenstliche-intelligenz/zukunft-der-ki-chatbots-sind-keine-gespraechspartner-110212293.html