An attempt at an anti-disciplinary answer
“In the end, the rule is: computers compute, people live—with their scope for action and their dreams, but also with their failings and limitations.” Ulrich Hemel
Almost every day, reports of advances in artificial intelligence (AI) reach the general public. AI composes Beethoven’s 10th Symphony, AI-generated artworks fetch new record prices, AI solves the protein-folding problem, AI is revolutionising investing and the real-estate market... At the same time, awareness is growing that the density of information and work in the rapidly expanding digital space is pushing people to their limits. Depending on the degree of personal impact, social position, and educational background, people’s reactions to these trends vary widely, ranging from euphoria and bewilderment to fear. For this reason, the question posed by Ulrich Hemel—“What distinguishes humans from artificial intelligence” (Hemel, 2021)—is not only of philosophical interest, but of great relevance to civil society. Even now, people are making career decisions based on possible further advances in AI. In the United States, there is a shortage of truck drivers because it is widely assumed that autonomous driving will make this profession obsolete. Translators are being asked how they intend to earn money in the future given the superiority of translation algorithms. Even software developers are being told that the golden years are over. To the same extent, companies are asking themselves how far automation should be pushed. Does it make sense to replace call-centre staff with chatbots? What does a company gain—and what does it lose—when algorithms determine applicant selection?
Interestingly, in this discourse it is often the shortcomings and limits of humans that take centre stage, rather than those of digitalisation and AI. Even more serious is the often fluid transition from science to marketing and science fiction. Above all, however, it becomes clear that the discourse frequently remains confined to the boundaries of specialist fields and the paradigms prevailing there. Given the complexity of this task, I would like to attempt an anti-disciplinary answer. Dirk Brockmann describes complexity research as anti-disciplinary by its very nature. In my view, the same applies to research and reflection on digitalisation and artificial intelligence. Opposite the nearly 8 billion human actors, there are already more than 2 billion PCs, almost 4 billion smartphones, 20 billion connected IoT devices, and countless software instances acting more or less intelligently. The socio-psychological and technical complexity of this system is likely to surpass anything individual disciplines have analysed to date. Because here, too, the following applies: in the end, everything is connected to everything else.
In the first line of thought, an attempt will be made to sharpen the question. As a natural scientist and mathematician, one is trained to first ensure that the problem is well defined. Before turning to the difference between humans and artificial intelligence, the following questions should first be clarified:
- What is a human being?
- What is artificial intelligence?
The answer to (1) has, as is well known, been attempted often—and revised even more often. Originally a domain of philosophy and theology (one could say this question is their point of departure), today molecular biology, medicine, psychology, economics, law, computer science, chemistry, physics, and literature each have their own perspectives on the human being. Legal person, homo economicus, factor of production, user, agent, observer, complex system, heroine, patient are all facets of human existence, but each is insufficient on its own. Personally, I like what may be the historically first definition by Democritus (460–370 BC): “A human being is what we all know.” This already yields a first difference from artificial intelligence, because certainly not all people know AI.
Kant, who formulated the famous four fundamental questions of philosophy, was himself unable to provide a definitive answer. According to David Johann Lensing, “that is precisely the nature of all fundamental questions of philosophy: it is not about answering them conclusively, but about posing them again and again, addressing them, thinking them through, and answering them provisionally.”
It is important to note that the question “What is a human being?” cannot be answered definitively by anyone. Stanisław Lem writes in his future novel “Solaris”: “We are looking for humans, no one else.” Without wanting to anticipate what follows, this also seems to apply to the debate on artificial intelligence.
The answer to (2) is, surprisingly, no less difficult to find. If there is one thing all AI experts agree on, it is that the term artificial intelligence is an unfortunate choice. The first use of the term goes back to the initiators of the Dartmouth Conference in the summer of 1956: John McCarthy, Marvin Minsky, Claude Shannon, Nathaniel Rochester.
“We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves”
This already states what is meant by artificial: simulation with the help of a machine. Since the famous work by Alan Turing (Turing, 1936), it has also been clear that this machine can in turn be simulated by a Turing machine, a universal computer. Defining intelligence, however, is far more difficult. “Since individual cognitive abilities can be developed to different degrees and there is no consensus on how to determine and distinguish them, there is no universally valid definition of intelligence” (Wikipedia, 2021). The search for an appropriate definition of intelligence has thus become a research project in its own right (Shane Legg, 2007). A variant acceptable to the AI community was distilled from 70 (!) common but different definitions of intelligence: “Intelligence measures an agent's ability to achieve goals in a wide range of environments”
Overall, this shows that, faced with the simple question “What distinguishes humans from artificial intelligence?”, language fails. Mathematicians, computer scientists, physicists, economists, entrepreneurs, philosophers, theologians, sociologists, politicians, journalists, artists, and citizens of all ages take part in this discourse, each with their own images of the human being and their own understanding of intelligence. It is therefore unsurprising that, without a shared language, people often talk past one another and misunderstandings arise. With regard to digital ethics, Ulrich Hemel sees ethical language competence “as a core competence for the future if we want to live together peacefully in complex, digitally organised societies” (Hemel, Kritik der Digitalen Vernunft, 2020). Ethical language competence can be understood as the ability to disclose the assumptions underlying ethical judgements, and as the ability and willingness to change perspective when weighing ethical standards” (Hemel, Kritik der Digitalen Vernunft, 2020). I believe we likewise need a digital language competence that enables all participants in the discourse to disclose their respective assumptions for judgements in the human-versus-AI debate. The ability and willingness to change perspective helps not only in the context of ethical conflicts in the use of AI. It is a basic prerequisite for clear thinking and reasonable action in the digital space. Digital language competence should not be equated with digital literacy, which focuses on technical skills in using digital media. Indeed, the core of the problem described above is that people learn efficiently and quickly to function with digital technology, yet are not aware of the processes and contingencies beyond user interfaces. Just as ethical language competence becomes essential above all in (often dilemmatic) conflict and crisis situations, so digital language competence is needed to assess the significance, possibilities, and limits of artificial intelligence beyond the jargon of technology companies.
In the second line of thought, the required change of perspective will be illustrated using several use cases. The use of AI for employee recruitment is already relatively widespread in companies. Put simply, the aim is to select from the pool of all applications those that best fit the advertised position and the company. Traditionally, this task is handled by recruiters, hiring managers, and interviewers. What distinguishes these people from an AI (software) agent developed by data scientists, and what impact do the answers have on the self-image of everyone involved?
From the data scientists’ perspective, the focus is primarily on mathematical statistics and the question of how the fit of a particular applicant can be predicted on the basis of historical datasets. The details of how this is done are not important here. What matters, however, is that there is no magic or black box at work—only mathematics and computer code. As in theoretical physics, a problem is reduced to mathematics and solved using powerful algorithms and computers. The AI agent learns from the abundance of an ever-growing stock of applicant data, can sort applications based on the calculated predictive score, invite the top 10 to an interview, or even decide automatically. From this perspective, the following differences between humans and AI become apparent. Artificial intelligences follow rules; humans sometimes do. With identical data, the AI agent will deliver identical results at any point in time. The calculated fit is a function of the input data. A hiring manager’s vote, however, can vary and is unpredictable. Daniel Kahneman provides a number of examples of this in his book Noise. Statistical analysis of court rulings, for instance, has shown that judgements can depend, among other things, on the time of day. “If judges are hungry, they are tougher.” Even the same case can be assessed differently by people at different times, for whatever reasons. Kahneman calls this unintended variability noise. Incidentally, both artificial intelligences and humans can be affected by prejudice (bias). Amazon, for example, had to learn this when its AI agent systematically disadvantaged women in the recruiting process (Reuters, 2018). The reason was the training data. If a company predominantly employs men, the AI naturally learns that pattern as well. This brings us to the next difference. Artificial intelligences do not break new ground; humans do. AI agents are trained on historical data, find patterns, and optimise objective functions. In doing so, however, they always perpetuate the status quo. Normative decisions (for example: the share of women in the workforce should be increased) must be programmed in or controlled via parameters. This is a profound difference. Because AI consists 100% of mathematics and computer code, it cannot bring anything truly new into the world. This may seem surprising given AI that composes music and paints pictures. But left to themselves, all these agents end up in an endless loop. The true artists are the developers behind this kind of software. I will return to this point later, but it should already be noted here that AI has the character of a tool. Humans can certainly use AI to create something new—whether good or evil, beautiful or ugly. It would be a sleight of hand to attribute that to the tool.
From the perspective of psychology and economics, human decision-making processes in applicant selection are comparable to the way other experts (judges, doctors, assessors) arrive at judgements based on presented facts. “The decision makers must consider multiple options and apply their values to make the optimal choice. But the decisions depend on underlying predictions, which should be value-neutral” (Kahnemann 2020). Experienced HR leaders know from experience which indicators and characteristics speak for or against a candidate, ask follow-up questions, enter into an open conversation, and form a judgement. Of course, this judgement is anything but objective and is not determined. Mood, room temperature, time of day, and countless other factors completely independent of the applicant can influence the judgement and the decision. That is why it is common practice to make hiring decisions based on the judgements of several experts.
Above all, it is the conscious or unconscious application of values that distinguishes humans from AI. In Max Scheler’s classic definition, values are “irreducible basic phenomena of feeling intuition.” Irreducible means that values cannot be mapped onto numbers or mathematics. Values simply elude any form of quantification and thus digitalisation. Above all, values are not natural constants, but highly situational. A job interview conducted between two people can perhaps be recorded, but it can never be fully captured or simulated digitally. Ulrich Hemel calls this the digital incompleteness theorem. Just as statistical physics cannot make statements about singular phenomena but only about ensembles, no AI can capture all the nuances of a particular person or situation. Humans can do this because they are not dependent on data, but rather experience the situation subjectively. Humans invented mathematics and science to overcome subjectivity and to be able to make judgements free of feelings and prejudices. Artificial intelligence and computers, as method and hardware, are suitable for delivering objective results in this sense—provided data quality is sufficient. Source code and the data basis can be disclosed so that the process is transparent for all involved. If objectivity and justice were the only standards, then humans should indeed leave all personnel decisions to artificial intelligences. In summary: Artificial intelligences are potentially objective; humans are always subjective.
A company leadership team considering the use of a recruiting AI will have to weigh the advantages and disadvantages of human judgement compared with machine decision-making processes, as well as the direct and indirect effects on the workforce and the company’s reputation. A central statement of this analysis is that the image of the human being held by decision-makers, and their ability to change perspective, are of the greatest importance. If one assumes that applicants can be fully described by data profiles and that HR experts do not, in principle, work differently from computers—only more slowly and with more errors—then a decision in favour of a high degree of automation is superficially rational. Another extreme could be that management guided by classical humanistic values fundamentally rules out the use of AI in HR due to ethical concerns. At first glance, this decision appears morally right, but it does not do justice to the potential of artificial intelligence to arrive at objective assessments. One aspect of a new digital humanism could therefore be to sensibly balance the interplay between humans and AI.
Similar questions arise wherever AI, from a business and technical perspective, offers potential to increase the degree of automation. Image recognition is already widely used in practice. One use case is the so-called shelf audit in the consumer goods industry. The aim is for manufacturers to regularly check whether their products are presented in supermarkets in accordance with agreements with retailers. Traditionally, this task is carried out by sales representatives, who visit stores, check shelves, note prices, and so on. Today’s state of the art, however, is that sales representatives simply photograph the shelves, and an AI determines in real time whether there are deviations and where action is needed. This problem is very well understood by data scientists. There are millions of images that can be used to train the models. Objectivity, precision, and scalability make a shelf-audit AI a valuable tool for the sales representative, who ideally gains more time for tasks that only humans can perform. Just as astronomers can capture more detail in the night sky with a telescope, this AI enables sales representatives to analyse more shelf space faster and more accurately. Image recognition is also used in radiology to produce cancer diagnoses based on high-resolution X-ray images. From the data scientists’ perspective, similar methods are used here as well. Here too, radiologists and oncologists have a tool at hand to access millions of findings in a very short time and identify patterns. However, it is far-fetched to assume that such image-recognition software could or should take on the role of a doctor. The evaluation of an X-ray image represents one data point within a patient’s spectrum and requires classification and assessment by a human being. Here again, it becomes clear that the image of the human being held by decision-makers makes the difference. If one assumes that patients can be fully described by data profiles and that doctors do not, in principle, diagnose differently from computers—only more slowly and with more errors—then the medical profession appears at risk. If one acknowledges that patients cannot be reduced to data and that statistical analyses of clinical data deliver great added value, then both AI and the medical profession have a shared future. Thus, it becomes clear once again that the fundamental question is not the difference between humans and artificial intelligence, but rather “What is a human being?” Being able to state one’s personal stance on this question as a prerequisite is one of the consequences of digital language competence.
In the third line of thought, the question of the difference between humans and artificial intelligence will be considered from the perspective of the history of science and technology.
The methodology of data science goes back to Galileo Galilei, who recognised that nature is formulated in the language of mathematics. Conducting experiments, recording measurement results (data), and mathematical modelling have been the recipe for success in the exact natural sciences since Galileo. It is this scientific method that connects achievements such as Carl Friedrich Gauss’s discovery of Ceres’s orbit with the intelligent recommendations of Amazon, Netflix, or Apple. In all cases, it is about finding patterns in more or less noisy datasets. Judea Pearl, one of the pioneers of machine learning, sums it up as follows: “At the end of the day AI is nothing but fitting data to a curve” (Pearl). In fact, it was Gauss who developed the method of least squares in connection with calculating Ceres’s orbit. Based on observed orbital positions (data) and the known equations of motion (curve), he was able to optimally adjust the parameters (fit) and thus predict positions (predict). To develop a recommender AI, data scientists adjust the parameters of their mathematical model (curve) until the error (deviation from the objective function) with respect to the training data is minimal (fit). At its core, then, it is about using mathematical statistics to get as close as possible to the truth (ground truth). Or, put differently: learning from data. Both machines and humans (Gauss had nothing but paper and pencil at hand) are capable of this.
However, Daniel Kahneman’s work makes the difference between machines and humans clear. By nature, humans tend towards storytelling and heuristic judgement (causal, fast thinking) rather than mathematical analysis (slow, statistical thinking). “Our sense of understanding the world depends on our extraordinary ability to construct narratives that explain the events we observe (Kahneman, 2021). “Relying on causal thinking about a single case is a source of predictable errors. Taking the statistical view, which we will also call the outside view, is a way to avoid these errors”.
This tendency towards causal thinking and storytelling strongly influences the AI discourse and its effects on people’s self-image. Both laypeople and experts all too easily succumb to the illusion that intelligent machines have the character of a subject. The growing arsenal of specialised artificial intelligences resembles a puppet theatre. Only when the puppeteer steps forward do we awaken from the illusion. Only when we understand how an AI works and discover the human actors holding the strings does the feeling of facing something magical disappear.
While science has traditionally sought to expose and abolish myths, the AI community today sometimes does the opposite and invents new myths. Speculation about a possible singularity, superintelligence, or strong AI is no longer science, but (more or less good) storytelling. There are parallels to this in the history of physics as well.
In 1814, Pierre-Simon Laplace devised the so-called Laplacian demon. “We must therefore regard the present state of the universe as the effect of an earlier state and as the cause of the state that is to follow. An intelligence which, at a given instant, knew all the forces by which nature is animated and the respective positions of the beings that compose it, and which were moreover vast enough to submit these data to analysis, would embrace in the same formula the movements of the greatest bodies of the universe and those of the lightest atom. Nothing would be uncertain for it; the future and the past would be present to its eyes.” (Pierre-Simon Laplace in the preface to Essai philosophique sur les probabilités, 1814)
This was not merely a fantasy, but reflected the paradigm and ideal of classical physics in the 19th century. The successes of physics were so spectacular that many scientists succumbed to the temptation to reduce all natural phenomena to Newtonian mechanics. The Laplacian demon was a myth, but it nevertheless had a major influence on intellectual history. Positivism, materialism, and behaviourism were all inspired by the idea that everything can be reduced to mechanics. It took the paradigm shifts of physics in the 20th century to recognise the error.
The paradigm of the internet economy is comparable to that of classical physics. If, in Laplace’s quotation, one replaces atoms with consumers and position and forces with data, it describes quite well the worldview and vision of some digital corporations and AI start-ups. What the perpetual motion machine was to classical physicists seems to be strong AI to some computer scientists today. If one accepts that the Turing test is not truly convincing, it is not apparent how a theory of strong AI could be verifiable or falsifiable. One could compare this to speculation about the existence of parallel universes, which in principle cannot be empirically verified. In both cases, it is speculative thinking—with the difference that leading AI researchers repeatedly make public predictions about when the goal of strong AI will be achieved. With regard to the limits of natural science, Hans Küng pointed out that “in the last questions there is freedom” (Küng, Lesch, Was die Welt im Innersten zusammenhält). What existed before the Big Bang cannot be answered with the tools of physics. Freedom cannot be diagnosed on the basis of neuronal excitation patterns. The difference between humans and the future possibilities of artificial intelligence is one of these last questions. People are free to understand themselves as neuronal machines and accordingly regard strong AI as merely a matter of time. But they are also free to believe that mind cannot, in principle, be reduced to mathematics. Here, the confrontation between natural science and religion in the question of the existence of God seems to repeat itself—only now it is about the human being.
Problematic in this context is that the marketing departments of digital corporations and the media reinforce AI storytelling, thereby providing yet another example of the phenomenon of Narrative Economics (Robert Shiller). As an example, consider the solution of the protein-folding problem with the help of artificial intelligence. In a competition hosted by the University of Maryland (Critical Assessment of Protein Structure Prediction Contest – CASP), the aim was to predict protein structures from given amino acid sequences (Spektrum, 2021). Google subsidiary DeepMind developed an AI called AlphaFold, which was able to solve the extremely difficult protein-folding problem. The Spektrum report states: “An artificial intelligence outperforms its human competitors.” And further: “The machine ‘understands’ physics in some way alien to us, which enables it to calculate how the atoms of the protein molecule arrange themselves.” The name AlphaFold naturally recalls another world-famous AI, AlphaGo, which was able to beat the world champion in Go. The quote comes from a scientist (the biologist John Moult) and seems to point to something magical. It is language that misleads here. It is not an AI that outperforms its human competitors, but other humans who developed the method and the algorithm: John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W. Senior, Koray Kavukcuoglu, Pushmeet Kohli & Demis Hassabis. It is not the case that the machine understands physics—or anything at all. Here too, it is only mathematical statistics and learning from data, even if humans cannot retrace the details of the solution path.
This points to one of the main theses of this analysis. It is less about the inferiority of humans compared with artificial intelligence, and more about the superiority of individual humans with the help of AI. This fact is obscured by language (AI as subject) and storytelling techniques.
In view of the still-accelerating process of digitalisation and the obvious advances in machine learning, the question of the difference between humans and artificial intelligence concerns all of us: the individual citizen as well as society, the natural sciences and the humanities, manufacturers as well as users of AI technology. Finding answers is difficult for several reasons. First, many analyses qualify themselves with a “not yet.” What may be a difference today could already be obsolete with the next technological leap. Furthermore, the respective definitions of the human being and artificial intelligence are anything but clear and stable. The question itself is therefore already a moving target. The greatest challenge, however, is the lack of clear and comprehensible language with which all participants in the discourse could communicate across disciplinary boundaries. While many laypeople marvel at the results of applying artificial intelligence and uncritically adopt the language of the technology industry, AI experts are only in exceptional cases able to explain their field to everyone and, for example, adopt the perspective of the humanities. A comparison with the history of physics shows that this is not a truly new phenomenon. Radically different, however, is the fact that this is not a predominantly scientific discourse within informed specialist circles (such as the debate on quantum theory), but a broad public discourse. In particular, the public actively participates in the experiment (as data providers and users).
As in complexity science, we must succeed in viewing the system as a whole. Behind every AI there is a developer, behind every model an analyst, behind every theory a paradigm, behind every investment a business model, behind every AI story a narrator, behind every (human) decision a value framework. What makes the matter truly complex is that it is not only about something external to humans (such as Jupiter’s moons), but about the human being itself. People’s self-image in dealing with artificial intelligence becomes a reality of its own and must therefore be included in the view of the system as a whole. If, as a result of ongoing interaction with intelligent machines, people begin to perceive themselves as machines, this will change their dreams and restrict their scope for action. A changed self-image results in a changed image of the human being. A changed image of the human being leads to different decisions regarding the meaningfulness and limits of applying AI—both in companies and in personal contexts. That is why a clear separation between scientific discourse, product marketing, and science fiction is so important. The way artificial intelligence is spoken and written about creates a reality of its own and, in the worst case, fuels a process of dehumanisation in thinking and action. Strong AI would thus become a self-fulfilling prophecy—not in the sense that AI approaches the level of humans, but that humans become machines.
The differences between humans and artificial intelligence, identified on the basis of introspection accessible to every person, mathematics, and computer technology, can provide orientation.
It is not claimed that the list of differences discussed here is complete in any way—or even could be. What matters, however, is to bring the differences between humans and artificial intelligence into a shared language. “For whatever people do, recognise, experience, or know becomes meaningful only to the extent that it can be spoken about” (Arendt, Vita Activa).
References
Brockmann, D. (2021). Can’t see the wood for the trees.
Die Welt. (1996). https://www.welt.de/print-welt/article652666/Computer-schlaegt-Kasparow.html.
Hemel, U. (2020). Critique of Digital Reason. Herder.
Hemel, U. (November 2021). From the deficit model of the human being to digital humanity. What distinguishes humans from artificial intelligence? From Philosophie InDebate: https://philosophie-indebate.de/3964/indepth-longread-vom-defizitmodell-des-menschen-zur-digitalen-humanitaet-was-unterscheidet-menschen-von-kuenstlicher-intelligenz/ accessed
Lem, S. (no date). Solaris.
Microsoft. (2016). Retrieved from https://blogs.microsoft.com/ai/historic-achievement-microsoft-researchers-reach-human-parity-conversational-speech-recognition/
New York Times. (2011). Retrieved from https://www.nytimes.com/2011/02/17/science/17jeopardy-watson.html
Shane Legg, M. H. (2007). p. https://dl.acm.org/doi/10.5555/1565455.1565458. Retrieved from https://dl.acm.org/doi/10.5555/1565455.1565458
Spektrum. (2020). https://www.spektrum.de/news/deepmind-will-problem-der-proteinfaltung-geloest-haben/1802324.
Statista. (no date). Retrieved from https://www.statista.com/statistics/802690/worldwide-connected-devices-by-access-technology/
Statista. (2020). Retrieved from https://de.statista.com/statistik/daten/studie/309656/umfrage/prognose-zur-anzahl-der-smartphone-nutzer-weltweit/
TEDBlog. (2013). Retrieved from https://blog.ted.com/how-did-supercomputer-watson-beat-jeopardy-champion-ken-jennings-experts-discuss/
Turing, A. (1936). On Computable Numbers, with an Application to the Entscheidungsproblem.
Wikipedia. (2021). Retrieved from https://de.wikipedia.org/wiki/Intelligenz