Problematic and background

‘AI understands what is real from within a virtual world, whereas HI creates what is virtual in a real world’ (Yu-cheng, 2020a, p. 120). Marvin Minsky defines AI as “the science of designing machines capable of doing things that require intelligence when they are done by humans”. Co-pioneering with Minsky, John McCarthy elaborates on AI as “any intellectual activity can be described with sufficient precision to be simulated by computer science, electronics, and cognitive sciences” (Lexcellent, 2019, p. 5). For both of them, it seems that the term “intelligence” can be defined, realized, or imitated by and through certain designs of machines (Floyd and Bokulich, 2017; Husbands et al., 2008). Back in 1950 when Alan Turing considered the question “Can machines think?” He reframes the question with a game called the “imitation game” (Turing, 1950, p. 433). The original version of the imitation game approaches the question of thinking by not directly answering what is intelligence. Instead, intelligence has been defined as to what extent HI can be imitated by a machine “through TEXTS” (Turing, 1950, p. 434). Hereby, the question “Can machines think?” can be replaced by a new one: “What will happen when a machine takes the part of A (a woman) in this game?” Since then, intelligence becomes something material and can be formalized and “AI therefore aims to reproduce mental activities with the help of machines, in the field of understanding, perception, or decision” (Lexcellent, 2019, p. 6).

Imitating requires firstly building a certain “distance” between the imitator and the imitated. The task of creating distance can be achieved through formalization. The inventor of binary arithmetic, Gottfried Wilhelm Leibniz, believes that everything can be represented with 0 and 1, which influences today’s digital computing (Davis, 2018, p. 171; Strickland and Lewis, 2022). It is Leibniz, and Sir Isaac Newton as well, who formalizes the modern Calculus, and it is also this formalization that establishes a further distance between humans and the world (Robering, 2014). Representing through formalization hence becomes unavoidable when exploring how to make machines think like humans. In the beginning, the idea of formalization involves the appearance and maintenance of ‘distance’. When distance is created, connections or disconnections are established as well. Distance is a source of decidability, as Baudrillard argues, “By the abolition of distance, of the ‘pathos’ of distance, everything becomes undecidable” (Baudrillard, 2005, p. 75). In his introduction to The Intelligence of Evil, Chris Turner points out that Baudrillard’s understanding of the pro-duction of the world, that is, simulation, indicates that “a reality is being produced that is extreme itself, extreme in the absence of critical distance it grants us, in the all-enveloping nature of its short-circuited, real, asphyxiating immediacy (Baudrillard, 2005, p. 8).

Considering AI will be expected to become more human-like, it implies that a distance between humans and machines firstly has been created and then has to be overcome. In doing so it makes AI also a product of formalization in which a distance is established that is too large to be perceived, thanks to pursuing exhaustive logic. The logic of exhaustion aims to cancel distance after creating it. Baudrillard uses hyperreal or vertigo to describe a rudderless state arising from the absence of a sense of distance (Baudrillard, 1983, p. 147; Gane, 2003 (1991), pp. 102–103);Footnote 1 This vertigo or rudderless feeling requires an intervention of certainty to provide a direction for re-establishing a sense of distance (a sense of direction is also built on distance). “Certainty” is technical/technological, and it can only be provided with technics/technologies that can somehow exhaust something. It turns out that a re-established sense of distance becomes a product of vertigo and is built on a distance being expanded beyond human perception because of technological interventions. Hereby, we can have two kinds of distance: primary and secondary distances. Primary distance is a distance created through cognitive processes to observe the world, and it is a characteristic of HI, presented as the logic of illustration. Secondary distance is a re-established sense of distance, based on a logic inherent in tools (technology). The logic inherent in tools (technology) can be generalized as the logic of exhaustion whose aim is to exhaust everything. For which Heidegger has convincingly elaborated why technology is a mode of revealing and modern technology is not just to reveal something but also a mode of challenging and of “Enframing” (Heidegger, 1992, pp. 13–18).

Since certainty is both technical and technological and has to be provided with due technics/technologies, its impact upon HI is to transfer the way to know the world from the illustrative to exhaustive logic. Following Baudrillard, it can be said that certainty is only an illusion that HI needs to provide a direction. It offers a possibility of making things plausible. The plausibility of what can be seemed certain becomes an important operation of HI. The study of AI may lead us to know better how certainty functions as insurance to keep HI to “grasp optimally” the world, following Merleau-Ponty (Merleau-Ponty, 2005 (1945)). Those will be realized in AI also reflects HI. To what extent the logic of exhaustion may replace the logic of illustration all over daily life and human beings will be a crucial question to be answered. However, following the logic of exhaustion, AI points to a world without uncertainty, and an exhaustive-logic-based certainty has been gradually programmed into HI as its only mode to grasp the world (Clark, 2013). Contrary to AI, the cognitive process of HI bases itself not on exhaustion, but rather on illustration. Illustration is a way of expressing and extending uncertainty.

To some extent, both the logic of illustration and of exhaustion are forms of representation. Computing machines are developed from philosophical conventions of rationalism, and formal languages that computers employ are a type of exact or abstract representation. West and Travis elaborate on how the concept of representation is directly related to modern western philosophical traditions such as rationalism, dualism, formalism, and mechanism. These traditions began with Hobbes, Descartes, Leibniz, Locke, and others, and they have inspired the works of Babbage, Turing, von Neumann, and Wiener. These traditions are also the foundation of theories proposed by several pioneers of AI, such as McCarthy, Minsky, Simon, and Newell (West and Travis, 1991, p. 72).

In pursuing formalization, rationalism guarantees a path towards a formalized understanding of the world. Those which cannot be formalized are to some extent to be seen as irrational. For the past few years of frustrating attempts to realize features of HI by computer scientists, most of them continue to rely on exhaustive logic. This may be because HI cannot be easily described or because to computer scientists, reality and truth are deemed as concepts that can be formalized: “Many researchers in AI are convinced that even the world of everyday life—of metaphor and ambiguity, for example—can be programmed into formal rules that a machine can understand, if only we specify them in all their complexity” (also ref. Davis, 2018, p. 171; Wolfe, 1991, p. 1079). However, there is an increasing focus on how to know better HI works in order to develop AI (Chen and Nonaka, 2022; Walsh, 2018). For example, researchers start to think that how to make AI deal with reflexivity, or so-called ‘intuition’ (Ichijo, 2022), or to make AI more empathetic, or emotionally responsive (Barbey et al., 2014; Franzoni et al., 2019). All these indicate that on one hand, formalization is still a way to realize AI, and on the other hand, it is exactly this way that becomes an obstacle encountered by AI developers. The further problem will be: is it enough for developing a human-like AI if the complexity of HI and its relation to the world (knowing processes) can be fully formalized?

It is in this sense the logic of illustration is introduced. Here, illustration is similar to William James’s ‘fringes of consciousness’ or ‘the vague awareness of humans in recognizing familiar faces in a crowd’ proposed by Hubert Dreyfus (1965, p. 21; Norman, 2017; William, 1890). The logic of illustration appropriated by HI makes its environment “observable and accountable”. It cannot be explained easily with some exhaustive technics/technologies. Illustration extends what cannot be perceived, understood, or included through maintaining distance, whereas exhaustion narrows down the world by cancelling distance. In other words, when AI knows its world through the logic of exhaustion, it runs itself within a closed system built and categorized with plenty of data. How AI makes the world observable and accountable may be different from HI’s. Ethnomethodology will provide some insights into this and will be elaborated on in the following sections.

The comparisons and contrasts between illustration and exhaustion in this paper aim to identify how HI and AI make things observable and accountable. In doing so, the concept of formalization needs not to be discarded because abstract thinking is also a major feature of the human mind. From the perspective of sociotechnical systems theory, when a system reduces the complexity of its environment, it does so by increasing the complexity of itself, by both which also increases the complexity of its environment (Luhmann, 1986 [1978], 1995 [1984]). This echoes Dreyfus’s argument that the human mind has a capacity to tolerate ambiguity, and the complexity of the environment cannot be exhausted due to its ambiguity in the human mind. These perspectives can be further explained through discussions of reflexivity and observations of related phenomena with ethnomethodology (Yu-cheng, 2020b).

From illustration to exhaustion: the emergence of homo technicus

Since the emergence of mankind as homo technicus, the relationship between man and tools—later, machines—has never been contradictory (José María Galván, 2003; José M. Galván and Luppicini, 2014; Stiegler, 1998).Footnote 2 In shaping and affecting each other, humanity and technology constitute a spiral through which the illustrative logic gives way to the logic of exhaustion. The outcome is ‘the mandate to recreate, in symbolic form, the totality of the environment in which a thinker operates—to develop a mental simulacrum of the thinker’s external world’ (West and Travis, 1991, p. 74). On the way to realize the “totality”, this “mental simulacrum of…”, the idea of exactness, completeness, certainty, and so on, have been emphasized. Exactness requires a certain degree of exhaustion and pursuing exactness also accompanies a further deepening of the exhaustive logic. So to speak, AI as an algorithm that is constituted of a definite number of codes can be considered an exhaustive system in that it exhausts itself within its own codes.

HI illustrates the world, whereas AI exhausts it (see Fig. 1). With the influence of formalization, scientists tend to equate the brain with the mind, as exemplified by saying ‘You are your brain’ (Swaab, 2014). In his discussion of minds and computers, Wolfe first distinguishes between the biological brain and the social mind: ‘The biological brain and the social mind work in radically different ways: one seeks information as complete and precise as possible; the other does not need hardwired and programmed instructions…because it can make sense out of ambiguity and context’ (Wolfe, 1991, p. 1078). This ‘[making] sense out of ambiguity and context’ is a key feature of the illustrative logic here. Although humans, similar to machines, require frames, scripts, or rules to think and act, they differ from machines in the ability to use them as “references”. Wolfe argues that because human memory is imperfect, these rules cannot be programmed in detail (Wolfe, 1991, p. 1080). ‘Incomplete’ refers to the brain as an ‘ongoing accomplishment’, following ethnomethodology (Garfinkel, 1992 [1967]). Due to the incompleteness of the human brain, HI cannot exhaust itself from within.

Fig. 1: Learning processes of AI and HI (Yu-cheng, 2020b, p. 122).
figure 1

HI illustrates the world, whereas AI exhausts it. In approaching the world (reality as well), illustration extends, creates, and imagines what cannot be perceived, understood, or included through maintaining distance, whereas exhaustion narrows down, summarizes, and simplifies it by identifying with it.

Full size image

The word illustrate shares an etymological root with illuminate. Light is a key to seeing visually; light illuminates an object, allowing it to appear in a visual sense. Even without any knowledge about an object, people can describe it in some manner, depending on what is illuminated by light. It means that the description does not need to be precise to be sufficient to convey information for other people to understand. As such, illustration encompasses and unfolds incompleteness, and it never aims at reaching any exactness, which is how HI knows the world.

Incompleteness is an operational feature that AI does not possess in the sense that AI has to be able to exhaust itself within its own codes. Otherwise, nothing is going to happen except error messages or crashes of operations. For that reason, coding experts have to always run ‘check’ after writing codes to make sure it is ‘complete’. Therefore, AI is both structurally (hardware) and operationally (software) complete, deterministic, and closed. To Wolfe, ‘…the everyday world provides the background or tacit knowledge that makes it possible to act in a contingent world, to act, as Dreyfus (1991) puts it, without a theory of how we act’ (Wolfe, 1991, p. 1082). For Weber and Mead, ‘What makes human intelligence different [from machine intelligence] is that, in human societies, people alter the rules they are expected to follow by incorporating information from the contexts and situations in which they find themselves together with others’ (Wolfe, 1991, p. 1091, the italics are mine). In other words, for HI, rules are not to be exhausted; rather, they are a point of reference with which HI illustrates the world, making it “observable and accountable”.

An opposite viewpoint has been expressed by some scientists. In Understanding Beliefs, Nilsson examines how HI and AI learn about the world. He argues that they are barely indistinguishable; HI describes the world through various “beliefs”, and AI through “models”; beliefs and models are mediums for understanding and used to represent the world, yet the real world is something that cannot be directly touched or understood (Nilsson, 2014, p. 116). For him, beliefs are not different from models. However, what he may overlook is the question of how beliefs and models emerge. Models have to be material; hence they can be exhaustive. Models need to be derived from and can be applied to objects, facts, and so on. They have to be exhaustive if they want to represent their objects appropriately. The opposite is true for beliefs, which do not need to be material; in other words, beliefs do not need to be based on reality or real existence. In this sense, beliefs extend the world, and they are also illustrative.

The argument will be that equating beliefs to models as Nilsson suggests is exactly what can be observed today: a transition from the logic of illustration to the logic of exhaustion, from incompleteness to completeness, from uncertainty to certainty, and from ambiguity to exactness. HI has been considered a model that can be simulated. The illustrative logic has been replaced gradually with the logic of exhaustion, and the latter has dominated how we know ourselves and the world. Therefore, Homo technicus, to some extent, can be reconsidered as a species that makes everything as exhaustive as possible, including itself.

Following Bernard Stiegler’s discussion of homo technicus, which means mankind by nature is technical, Liu explores the difference between HI and AI by extending Stiegler’s distinction of de-fault and default (Yu-cheng, 2020a). Due to the lack of any instinct, as a fault of Epimetheus in the Greek myth, HI learns by de-fault -the removal of the flaw- when encountering their environments. In doing so, their environments become a product, a default, of human beings’ inability to exhaust them. In other words, in de-faulting, a default has been generated in the sense that it cannot be exhausted. However, AI’s default does not emerge from the evolutive processes of de-faulting. The default of AI steers its own operations –its programmers code it and make it work according to certain rules, and aims at eliminating those that are seen by human beings as a product of de-fault, the environment. For HI, default is produced in the action of de-fault, whereas for AI, HI’s de-fault becomes its environment which has to be dealt with, with AI’s default—its programs, lines of code, and so on. Again, de-fault, characterized by HI, extends the world, creating possibilities, and makes sure noises from the environment can be treated properly. In this sense, it corresponds to the logic of illustration. On the other hand, AI’s default is designed by engineers, it exhausts not only its operations but also its environment, within definite numbers of lines of code. Back to the earlier discussion of distance, for HI, de-fault denotes a distance between human beings and their environments, just like illustration indicates a distance between the illustrator and the illustrated, whereas for AI, the distance has been cancelled since its environment can only be represented by AI with AI’s default. There is no space for interpretation between AI and its default. In a word, how its environment has been presented refers to the way it is exhausted. In this case, it is AI presenting a world of nothing but a result of the exhaustive logic.

In this regard, HI’s de-faulting aims not at exhausting its environment. Its understanding of the world emerges from the processes of knowing, accepting, modifying, and unfolding the flaw, the fault, that is, the lack of any instinct. HI’s default emerges from de-faulting. In this way, de-fault is exploratory. It corresponds to the logic of illustration, yet the latter emphasizes more a change of distance. AI’s default is given by something that is not intrinsic to itself. In knowing the world, for AI, HI’s de-fault, resulting in the inexhaustible, becomes AI’s environment that has to be exhausted. HI’s de-fault illustrates and hence enriches its environment, whereas AI’s default attempts at exhausting everything, making its environment ‘count-able’, controllable and predictable. Since HI’s de-faulting explores and extends its environment, producing something new, it brings about a possibility of reflexive operation. In terms of Garfinkel, HI’s making social settings observable and accountable, that is, its de-faulting as well as illustrating, requires an operation of “reflexivity” (Garfinkel, 1965). However, computing machines do not operate with reflexivity. In accordance with Niklas Luhmann’s social systems theory, such programs are incapable of observing their own observations (Luhmann, 1982, 1986, 1995 [1984], 1997, 2006 [1991]). AI’s default, framed by the logic of exhaustion, transforms HI’s de-fault, that is, those that have not been exhausted, into something exhaustible, becoming a part of its default. AI’s understanding depends on to what extent that de-fault can be included in its default. It is different from HI in that the understanding based on the transferring from de-fault to default is not equivalent to HI’s de-fault. For example, when converting color into binary code, the color disappears in the code. Furthermore, the color codes cannot be used to express those colors, as well as those feelings, moods, and meanings attached to them, as HI’s de-fault does.

The concept of reflexivity has been described in Ethnomethodology as “members’ accounts, of every sort, in all their logical modes, with all of their uses, and for every method for their assembly are constituent features of the settings they make observable” (Garfinkel, 1992, p. 8). Every observation has included a previous observation that has been made observable by social members, just like every de-fault has also been based on its previously produced default. In so doing, every effort HI makes to describe reality alters how reality is presented to it. That’s how and why Garfinkel refuses to replace indexical expressions with objective expressions (Garfinkel, 1992, pp. 4–6). In this sense, it resonates with Wolfe’s discussion on the relationship between the biological brain, social mind, and memory (i.e., the imprecise nature of memory). For HI, the past leaves behind traces that may be unclear and imprecise, and we can only rely on these traces to search for the past. For AI, the so-called past is recorded and preserved exactly and precisely; these records are no traces of the past but the past itself (O'Keane, 2021). With ethnomethodology, when HI uncovers traces of the past indexically to understand the ‘what, how and why’ of an event or phenomenon, it finds that these traces are not so clear and sometimes cannot be reminisced exactly. Then, HI may activate various mechanisms to fill in these blanks and piece together an image of the event. These mechanisms include those that allow us to interpret and even to create certain traces, as well as to determine how these traces are connected to each other. The operation of reflexivity intertwines with the use of indexical expressions, and they are interdependent on each other. Both reflexivity and indexical expressions are lacking in AI in that for the latter everything is stored without any loss, and nothing can be forgotten, hence leaving no room for interpretation.

In summary, the emergence of homo technicus paves the way for the logic of exhaustion, which means not only AI and machines are a result of it, but also human beings start to observe and think of their environment with the logic. With Stiegler’s distinction of default and de-fault, the former has become the logic of exhaustion, assigned to machines, whereas the latter belongs to human beings, yet is gradually replaced with the former. The difference between them is the idea of distance. The illustrative logic appropriates and maintains a distance between the illustrator and the illustrated. The aim of the illustrative logic is not to equate the illustrator to the illustrated. In doing so, it keeps open possible interpretations. That is how it extends the world. Following ethnomethodology, the logic of illustration can be elaborated with the ideas of ‘indexical expression’ and ‘reflexivity’. Both of them also require a distance to work. However, the logic of exhaustion is to eliminate the distance between the exhauster and the exhausted, to equate the knower and the known. Hence it lacks room for ambiguity, inexactness, uncertainty, interpretations, and so on. Therefore, it can be argued that AI cannot observe its own observations, and does not operate reflexively. The question altered from the above discussion will be: is it possible to create an AI, if it runs still according to the exhaustive logic, that can observe its logic and establish distance with what it exhausts? In other words, can AI be reflexive? Inspirations from ethnomethodology would be beneficial to probe into it.

Problems with natural languages and conversation analysis

Mastery of natural languages and AI’s exhaustive logic

Scientists have hoped to create an AI that can communicate with humans using natural language since the Dartmouth Summer Research Project on Artificial Intelligence in 1956. The question will be: what is natural language? And what features does it have? According to ethnomethodology, natural language is indexical, reflexive, and contextual. Natural language is not located in the realm of the exhaustive logic. A preliminary answer has been suggested in Dreyfus’s discussion on the three information-processing abilities specific to humans: fringe consciousness, essence/accident discrimination, and ambiguity tolerance. As he has argued, “…in perception we need never appeal to any explicit traits. We often recognize an object without recognizing it as one of a type or a member of a class” (Dreyfus, 1965, p. 40). He concludes, “The fact that we need not conceptualize or thematize the traits common to several instances of the same pattern in order to recognize that pattern, distinguishes human recognition from machine recognition which only occurs on the explicit conceptual level of class membership” (Dreyfus, 1965, p. 42). It could be said that on one hand for HI to reach such explicitness is not possible, and on the other hand, due to the lack of explicitness, what HI achieves is an ability of “context-dependent ambiguity reduction”: “In recognizing certain complex patterns, as in narrowing down the meaning of words or sentences, the context plays a determining role. …The context not only brings out the essential feature, but is reciprocally determined by them” (Dreyfus, 1965, p. 42). For AI, ambiguity needs to be exhausted, while for HI it needs simply to be illustrated in that ‘context-dependent’ refers to a tolerance of something that is not explicit enough to be discerned. The further question will be: Can context be exhausted?

In his discussion of Wittgenstein’s study of natural language and his idea of ‘family resemblances’, Dreyfus points out that “formalizing family resemblance in terms of exactly similar traits would eliminate the openness to new cases which is the most striking feature of this form of recognition” (Dreyfus, 1965, p. 43). Here the ‘openness to new cases’ implicates also a kind of enriching, extending, and unfolding the world as the illustrative logic suggests. When recognizing with the exhaustive logic, whether AI or HI’s reducing ambiguity and complexity of its environment—that is, knowing the world—depends on a kind of ‘overconfident rationality’. The well-known ‘The McDowell–Dreyfus Debate’ may describe part of it: “Is human experience permeated with conceptual rationality, or does experience mark the limits of reason? Is all intelligibility rational, or is there a form of intelligibility at work in our skillful bodily rapport with the world that falls outside our intellectual capacities?” (Schear, 2013) While contrasting perceptual consciousness with conceptual rationality, Dreyfus and McDowell approach the world differently. The former seeks to “identify and describe forms of ‘absorbed coping’ that do not come within the scope of conceptual rationality”, whereas the latter denotes a process of conceptualization when HI get to know its environment (Schear, 2013, p. 2). As mentioned above, rationalism guarantees a way to formalization, looking forward to a world without ambiguity, expected to be realized in machines and AI algorithms. The process of formalization requires and results in a separation of body and mind, which is also a theme suggested by René Descartes having a huge impact on rationalism, and also bringing disasters to the world (Schiphorst, 2009; Shilling, 2005; Todes, 2001). It is this ‘overconfident rationality’ that encapsulates the logic of exhaustion that considers natural language in a dissectible way.

There are two implications or premises of this ‘overconfident rationality’. One is to consider the environment (the world) as having-been-exhausted. The exhaustive logic can work only if its object can be seen as exhausted, just like machines and AI algorithms do. The other is that in this already-exhausted environment, since everything can be exhausted, there exists a form of cognition that can exhaust everything, yet we have not discovered where it is and how it works. Both AI and the recently emerging Master Algorithm are products of this viewpoint.Footnote 3 Both of the two implications cannot explain well how the same social settings, same words, or same sentences have different meanings or interpretations to different people if let says everything has been or can be exhausted. In a word, following ethnomethodology, the logic of exhaustion requires no reflexivity. It cannot and does not need to observe itself in that a distance does not exist to make an observation. The fact that new cases will be included or excluded depends on the logic’s structural and operational closure since exhaustion indicates a certain closure not just in its form but also in its content.

What natural language does is to make sure it has enough complexity to deal with those that cannot be exhausted. In this sense, natural language is illustrative, reflexive, and open to its environment. This also means that natural language is different from machine language in that it can tolerate ambiguity as HI does. In line with Dreyfus’s argument, HI has not only ambiguity tolerance but also reflexive ambiguity tolerance. By contrast, AI algorithms have only reflective ambiguity solutions. In terms of Garfinkel, social members’ accounting practices are reflexive in that every account of social settings includes their previous accounts of them. Hence, to make social settings “observable and accountable” is not just considered an ongoing accomplishment, but also indicates a continuous change exists through social members’ accounting practices. In other words, this reflexive ambiguity tolerance allows social members to observe themselves, and the observations of their accounting practices also constitute their observations. Considering AI, although it can process ambiguity (albeit in deterministic forms), it cannot observe its operations (observations). Processing ambiguity for AI is on one hand accomplished through exhausting itself (within its codes), and on the other hand, its decision made relies on a complete running of its codes, rather than on tolerance of something incomplete, or it cannot exhaust.

As far as natural language processing and comprehension technologies are concerned, the premise and presumption of machine translation are that it must first exhaust all aspects and possibilities of languages and their rules. However, the use of natural language—or what Garfinkel refers to as the mastery of natural language—is not based on the exhaustive logic. Learning and using natural languages are related to Wittgenstein’s philosophical concept of a language game, which not only varies by time and space but also by cultures and social contexts, in terms of what he called ‘pointing’ as a basic cognitive process (Wittgenstein, 1953, 2001). In other words, natural language is developed and appropriated based on accepting and including ambiguity, instead of rejecting and excluding it. There are plenty of cues or indexicals hidden in the use of it. Dreyfus, citing Bar-Hillel’s research on language translations, argues that “in order to use a computer to interpret these cues, we would have to formulate syntactic and semantic criteria in terms of strict rules; and our use of language, while precise, is not strictly rule-like” (Dreyfus, 1965, p. 33). Since every social member is also a master of natural language, it would be beneficial to understand what it means by ‘mastering natural language’ if AI is expected to be as competent as HI.

Natural language used by social members is both indexical and reflexive. It is interesting to point out that Garfinkel’s discussion of natural language aims to understand how sociological reasoning accomplishes its work, and how social members who are mastery of natural language can reach these accomplishments with the indexical properties of natural language, which include the idea of reflexivity. In persons doing sociology, natural language provides circumstances, topics, and resources of their inquiries, and it is this (use of) natural language that also furnishes the technology of inquiries and their practical sociological reasoning with “its circumstances, its topics, and its resources” (Garfinkel and Sacks, 2005 [1986], p. 157). Garfinkel describes this reflexivity as indexical properties of natural language. In other words, reflexivity, either following Garfinkel or Luhmann, should be considered a source of ambiguity, uncertainty, and indexicality. Reflexivity indicates a distance that is somehow created and maintained. Therefore, an operation without reflexivity would be an operation of eliminating distance, and in so doing, the logic of exhaustion would be possible as well.

The aim of doing sociology, for Garfinkel, just like doing science and developing AI as well, is to “accomplish a thoroughgoing distinction between objective and indexical expressions with which to make possible the substitution of objective for indexical expressions” (Garfinkel and Sacks, 2005 [1986], p. 158). The indexical properties of natural language, which presented itself as being full of ambiguity, make sure that a kind of remedial practice of practical sociological reasoning becomes “unavoidable and irremediable” (Garfinkel and Sacks, 2005). Respectively speaking, the indexical properties in the use of natural language follows the illustrative logic, whereas the substitution of them with objective expressions corresponds to the logic of exhaustion. The question for Garfinkel is “What is about natural language that makes these phenomena observable-reportable, i.e., accountable phenomena?” (Garfinkel and Sacks, 2005 [1986], p. 160) Social members’ mastery of natural language indicates that there are methods for them to produce and recognize formal structures of everyday activities through their “doing formulating” (Garfinkel and Sacks, 2005 [1986], p. 164).

While discussing the connections between how social members’ “doing formulating” and mastering natural languages, Garfinkel points out that ‘…the very resources of natural language assure that doing formulating is itself for members a routine source of complaints, faults, troubles, and recommended remedies, essentially’ (Garfinkel and Sacks, 2005 [1986], p. 170). In other words, for social members, “doing formulating” is a process characterized by complaints, faults, troubles, and recommended remedies for these complaints, faults, and troubles. They are the indexical expressions that have to be remedied for making social settings accountable. Therefore, mastering natural languages is closely related to indexicality. Inversely, indexicality requires social members to master natural languages if they have to communicate with each other. Mastery, in this sense, does not mean that social members achieve a complete comprehension or formalization of those indexical properties of natural language (this is what algorithms do). Rather, mastery indicates an ability to approach and deal with those indexical expressions.

By contrast, the language used by AI algorithms on one hand is given from the outside, and on the other hand, it does not include indexical expressions such as complaints, faults, troubles, or remedies. In a word, it is developed with objective expressions, 0 and 1, and so on. To what extent can we say that AI is a master of its language? To answer it, firstly, AI has to be capable of observing its observations. More specifically speaking, AI is the language itself, which means that there is no distance between AI and its operations, and there is no reflexivity in its operations. AI will not, need not, and cannot complain about what its language brings to it. For example, an algorithm does not complain about whether binary code is easy to use and does not make mistakes while using it. This means that the logic and usage of binary code do not change. This approach differs from social members’ use of natural languages. New meanings may be assigned to words and sentences at any time; new feelings and emotions, as well as their interpretations, may vary across different times and social settings. Although indexes are used in algorithms, they are simplified or reduced to rules; they also differ from indexicals in natural languages in that they require fixity, operability, and controllability.

AI cannot observe itself in that a distance cannot be established from within between AI and its observations. The distance is created by its programmers outside of it.Footnote 4 In pursuing the logic of exhaustion, a distance between the exhauster and the exhausted, the formalizer and the formalized, and the default and the defaulted, has been or has to be cancelled to describe a perfect correspondence, as possible as it could be, between two sides. Those which cannot be exhausted, formalized, or defaulted are to be seen as noises, outsiders, or troubles, that is, indexicals, that have to be eliminated until both sides are equal to each other. However, the logic of illustration requires no exact correspondence between two sides. Illustration deals with indexicals as a resource for transmitting information. Every illustration not only tackles indexicals but also produces or unfolds other indexicals that will or will not enter into the next illustrations. This is the operation of reflexivity.

The operation of AI is reflective, not reflexive. Reflective actions indicate no distance between a subject and its action,Footnote 5 whereas reflexivity means that the subject can establish a distance with itself, and observe its observations. Whether the subject is aware of this action of establishing distance, the outcome has always to be a product of it. Although it seems that AI can monitor its operations and debug itself, its every decision is still a result of the logic of exhaustion, which means it exhausts itself within its lines of code. In addition, it is the distance that makes “making sense of …” possible. It means the production of meanings emerges from a distance established and maintained somehow and somewhat.

In terms of Stiegler, when using tools to de-fault, human beings gradually externalize themselves and then see the world with this externalized body and mind. However, there is no such process of externalization exists in AI’s knowing its world. It can be said that human beings emerging as human species is an achievement of this externalization, that is, each of us can recognize ourselves as human species. Tools, machines and AI do not emerge as species to themselves. Rather, they emerge as “something of human beings”. Externalization refers to a distance established between human beings as species (technical) and human beings belonging to nature (natural). It is this externalization that procreates various technics of externalization such as languages, words, symbols, politics, cultures, and religions (Stiegler, 1998, p. 193). These technics of externalization reflexively provide the distance with meanings. AI is unable to establish distance with itself and is therefore unable to operate on the level of reflexivity.

In summary, in view of natural language, HI knows and approaches the world without the need to exhaust everything. With regard to machine languages and AI as well, plus their foundation, big data, it is often heard that with big data, we can see those unseen, or make invisible visible. Pursuing big data as a solution to better AI has also been pursuing the logic of exhaustion, since “big data” and its seemly beneficial in helping machines and AI to understand the world are the most important product of the logic. For exhaustion and big data as well, whatever is invisible can be made visible. It is not just the aim of science to exhaust everything whether visible or not, but also an ambitious drive to make humans “objective”. What can be transcended, in terms of transhumanist posthumanism, seems to be those that can be exhausted such as organs, senses, feelings, emotions, and last but not the least, the mind (Clark, 2013; Fukuyama, 2002; Lilley, 2013; More and Vita-More, 2013). Those unseen do not mean that they do not exist, and it is not necessary to make those unseen seen that can further prove they do exist. Regarding how critical realism explains reality, those unseen do have power or influence on the world (Elder-Vass, 2005, 2010; Vandenberghe, 2014). With the exhaustive logic, what AI and algorithms do is to transform, though violently, those unseen into those seeable to themselves, to put them into their learning frames, and to make decisions (Mann, 2017; Srinivasan, 2017).

The ‘et cetera’ principle as illustrative property in EM and conversation analysis

As an offshoot of ethnomethodology, the theory and methods of conversation analysis (CA), developed mainly by Harvey Sacks, explore how social members see, understand, and maintain visible and invisible orders through conversations (Sacks, 1992, 1995, 2006). Sacks claim that there are some ‘formal features’ that can be found in the analysis of ordinary conversations. For example, ‘People talk one at a time’, ‘Speaker change recurs’, ‘Sequences that are two utterances long and are adjacently placed may be “paired” activities’, ‘Activities can be required to occur at “appropriate” places’, ‘Certain activities are “chained”’ (Silverman, 1998, p. 103). These features constitute how social members make social settings, including conversations, accountable. It reminds us that conversations have also been used by Alan Turin to know if humans can be misled or simulated by machines. For now, CA happens to provide a comprehensive understanding of the distinction between the exhaustive and illustrative logic.

In a paper authored by Garfinkel and Sacks, On Formal Structures of Practical Actions, they explore how a formal structure exists and works with natural language mastered by social members. It is not hard to answer: “…conversationalists, in the course of a conversation, and as a recognized feature of that conversation, formulating their conversation” (Garfinkel and Sacks, 2005 [1986], p. 167). Social members “do formulating” as a master of natural language. By formulating means, as one of the subtitles in that paper suggests, “naming, identifying, defining, describing, explaining, etc., a conversation i.e., formulating a conversation, as a feature of that conversation” (Garfinkel and Sacks, 2005 [1986]). Viewing from CA, seeking any formal structure, just like any attempt at formalizing something, unavoidably needs to provide a possibility of exhaustion. In other words, in terms of Garfinkel, it needs to first draw a distinction between indexical and objective expressions and then substitute the latter for the former. In doing so, anything indexical becomes a problem, a trouble, or a de-fault, that has to be eliminated (Garfinkel and Sacks, 2005 [1986], p. 166). It is CA that provides exactly a counter point of view somehow can elaborate why and how considering the two logics are conducive to thinking of AI and HI. Accordingly, the question can be reversed as below: can formal structures, if can be found, of practical actions, including conversations, be applied to AI, and can it make AI think and act like humans?

The attempt at looking for formal structures, as well as for formalizing everything, presumes the logic of exhaustion, while the result of it simply demonstrates that those formal structures are a kind of machinery “in accountably rational discourse” (Garfinkel and Sacks, 2005 [1986], p. 173). Doing formulating indicates a “seen but unnoticed” formal structure does exist. It is rational in that “members are particularly knowledgeable of, sensitive to, and skillful with this work; with doing it, assuring it, remedying it, and the like” (Garfinkel and Sacks, 2005 [1986], p. 170). Since the properties of indexical expressions are troubles for conversations, their elimination depends on various attempts at satisfying the distinction of objective and indexical expressions in actual occasions for all practical purposes, and “providing objective expressions as substitutes for indexicals” through doing formulating (Garfinkel and Sacks, 2005 [1986]). That is to say, doing formulating not only refers to a formal structure, visible or not, but also tells an invisible story about how the logic of exhaustion works for it.

The fact that social members are required to do formulating points out exactly that the attempt at exhausting everything is not just a mission that cannot be completed but also has to be an endless and ongoing accomplishment. In ethnomethodology, Garfinkel proposes the “et cetera” principle to assist in understanding how exhausting everything is not realistic in social members’ accounting practices:

…the situations in which a rule is and isn’t applicable can never be finitely enumerated… In effect, every rule has an implicit et cetera at the end of its list of triggering conditions, and the agent must judge each time whether the current situation is one in which the rule may be applied, or, possibly, broken (Garfinkel, 1992, pp. 73–75).

Accordingly, the et cetera principle explains that when applying rules to something, it is not possible to exhaust it with those rules. Exceptions do always exist when applying them. It is either a matter of time or of space. The function of et cetera is to make sure a possibility of extension is not excluded. In other words, the principle ensures its object can be unfolded. In this sense, the logic of illustration presents itself as the et cetera principle when remedying the attempt of exhausting everything, in this case with rules. However, for AI and algorithms, “there are always exceptions”, by which means those cannot be exhausted, or those cannot be formalized, becomes a need-to-be-solved phenomenon. More importantly, it is not allowed for AI to use the et cetera principle to tackle these exceptions. For HI, the use of the principle means that exceptions are not a product of rules, but become part of them. It is one thing to seek a formal structure through studying social members’ doing formulating, it is quite another thing to apply this found formal structure to designing AI that can make social settings accountable, not just account-able, as social members do. In studying human–machine interactions, Hirst has pointed out that applying CA rules to designing human–machine interactions may be unrealistic. In his review of Button’s argument, CA rules have been appropriated by human conversers as “resources” in conversations, not strictly as rules, hence, exceptions as part of rules have been used also as resources (Hirst, 1991, p. 222). That is why HI explains exceptions, not simply rules out them according to the exhaustive logic. By contrast, for AI, exceptions are a product of rules and have to be eliminated, not explained as HI does, if rules are designed to be followed and cannot be changed by AI itself.

Take the long short-term memory (LSTM) algorithm of recurrent neural network (RNN) for example, it has been applied effectively to language modelling, speech-to-text transcription, machine translation, and so on, since it emerges in 1997 (Sherstinsky, 2020). According to IBM’s learning materials, RNN is defined as follows: “A recurrent neural network (RNN) is a type of artificial neural network which uses sequential data or time series data. …Like feedforward and convolutional neural networks (CNNs), recurrent neural networks utilize training data to learn. They are distinguished by their “memory” as they take information from prior inputs to influence the current input and output. While traditional deep neural networks assume that inputs and outputs are independent of each other, “the output of recurrent neural networks depends on the prior elements within the sequence (the italics are mine).”Footnote 6 There are several key terms worthy of mentioning: sequential, time series, memory, and long short-term. LSTM appears to be solving problems encountered by RNN in that it only processes its inferences from every previous variant. For example, in the sentence ‘I am a boy’, ‘I’ determines or predicts ‘am’, or ‘a’ ‘boy’, instead of vice versa. The relation between ‘I’ and “boy’ is not and will not be considered, that is, ‘I’ cannot be used to predict ‘boy’, ‘boy’ “I’. The problems have been stated as exploding gradients and vanishing gradients.Footnote 7

It is not just time that plays an important role in predicting results, but also distance places a crucial limit on RNN. The ‘long’ in LSTM indicates a solution for both time and distance problems through a key function, memory. The solution, on one hand, extends its sequential scope and on the other hand makes LSTM can consider not just variants next to each other, it can also process those that are not. In this sense, it can be argued that LSTM can better deal with so-called ‘context’ in communications. Despite its attempt at solving these two most important elements in conversations, depending on processing sequential data has become its ceiling in advancing itself. To some extent, LSTM can find a more suitable model or structure from an extra memory mechanism to predict the results. It is still a model or structure that confines itself in the exhaustive logic. As the Vanilla LSTM system developed by Sherstinsky indicates, its requirements have included terms such as ‘explicitly and unambiguously mapping…’, ‘account for all components of the system’, ‘the most inclusive form of the LSTM system’, ‘leaving nothing to imagination or guessing’, and ‘everything should be explained and mad explicit’ (Sherstinsky, 2020, p. 2). In other words, LSTM’s better performance in language processing still comes from its more powerful exhaustive ability. That may be a reason for some computing scientists to suggest another model called ‘hierarchical neural attention encoder’ to better address, if not impossible, these problems.Footnote 8 The term ‘attention’ corresponds to the et cetera principle, not just in the vertical but also in the horizontal sense. The direction might be right according to EM and CA. However, while it focuses on time or sequential issues, it overlooks another one, that is, the issue of reflexivity.

In applying the et cetera principle, HI can learn from exceptions and makes them a resource for its accounting practices. AI algorithms cannot learn from them and only treat them simply as a product of rules that have to be excluded. Therefore, it requires AI of dealing with exceptions in its exhaustive system. The result of it is further and more complete formalization with the logic of exhaustion. Following this, when applying CA rules, a product of formalization, to human–machine conversations, it assumes that human conversations can be exhausted to a certain extent, that is, a formal structure that can be found to ensure mutual understanding. In this sense, CA rules are to be viewed as being created or represented in human–machine conversations. However, quite contrary to it, human conversers do not literally use CA rules in conversations, which means human conversers do not really know what rules they use in conversations. Accordingly, any attempt at applying CA rules to simulating conversations between humans and machines will encounter a problem that in the first place does not have to exist that way between human conversers. That CA rules are induced from human conversations does not mean that they can exhaust them (Hirst, 1991, p. 223). The inspiration of CA rules refers not to rules themselves, but to the ideas of relevance and context, with which indexical expressions make accounting practices reflexive. As Hirst suggests:

It’s one thing to use a grammar of CA rules to perform post hoc recognition of the high-level structure of a conversation, or to generate an example of (both sides of) a conversation. It’s quite another to use such rules to actually participate in a conversation without making a mess of things (1991, p. 223).

Therefore, the et cetera principle provides another way to elaborate on how social settings, including conversations, are made observable and accountable, and it presents a great entry to distinguish HI from AI, the indexical from objective expressions, and the illustrative from exhaustive logics. Social members’ accounting practices are reflexive and indexical, and they are not like to be exhausted as AI algorithms do, with current mathematical models. It is also inspiring that inquiries about how social settings and conversations are made observable, reportable, and accountable, should contribute more to the study of human-machine interactions and AI as well. As indicated in a paper that discusses what will be the result of applying literally algorithmic methods to everyday activities, the authors conclude that every command simultaneously brings about something that has to be further defined or clarified, that is, something that has to be exhausted; otherwise, the action will not proceed, in that case, the two people described in that research will encounter problems in simply moving forward (Ziewitz, 2017).

Conversations and the act of making them accountable are embedded in various contexts and clues which are provided by cultures and will be appropriated for all practical purposes. Take the use of emoji for an easy example, emoji is some kind of visual presentation that can deliver information in conversations. The coding and decoding of emojis are according to various contexts and clues that appear during conversations. For each converser has encountered a situation of “double contingency”, in terms of Luhmann, each converser has to decide alone what an emoji conveys. It describes a situation that both sides of conversers cannot know how exactly each of them makes it accountable. CA rules provide a possibility to see some sort of order exists in conversations, which makes human conversers talk like humans. It is not just words, contents, or sentences that make social members’ conversations accountable. Rather, it is pauses, silences, and indicator terms that provide accountability to human conversers.

Take “word vector” technology for another example, it has been applied to developing AI with the ability to know what words, sentences, and paragraphs mean, and to generate responses with them. A word vector is a mathematically represented meaning of a word by calculating how often words show up next to each other. It is the frequencies of words that decide what should be meaningful to the algorithms. Surrounding words are simply numbers that inform computing machines what should be considered as a response to them. The function of context is reduced to the calculation of word frequencies. In the realms of NLP and NLU, it is either to train algorithms to raise their ability to calculate through deep learning or to increase more and more parameters and training data to obtain better performances (for example, Open AI releases models from GPT-2, GPT-3, to GPT4) (Brown et al., 2020). Either way assumes that with paramount parameters and data, better performance of understanding natural language can be achieved and even closed to humans’ mastering of it. It can be said that either way depends still on the logic of exhaustion. Therefore, pursuing always-considered-not-enough parameters and data is the only way to simulate HI or make AI think and act more like humans. If the et cetera principle can provide something to the logic of exhaustion, it will be that substituting it for the logic of illustration brings more difficulties to its realization, just like the dissatisfaction of replacing indexical with objective expressions, as Garfinkel argues. More importantly, the substitution or replacement is guiding the human species toward a future of posthuman in the sense of transhumanism. Everything that cannot be exhausted, formalized, digitized, datafied, or objectified, will be considered as not a default.

Conclusions

This paper aims to explore the connection between AI and transhumanist posthumanism. Following ethnomethodology, the author distinguishes the logic of exhaustion from the logic of illustration. Both the two logics are ways to approach and understand the world, however in an opposite direction. The logic of illustration, appropriated by the human mind, creates distance between humans and the world, whose aim is to make things understandable, with no regard to how things can be truly represented or not. With the development of science and technology, especially in pursuing exactness, steadiness, and predictability, the logic of exhaustion has attempted at eliminating the distance, resulting in a situation in which both humans and machines are seeking to become each other. Though post-modernists reject grand theories, what comes after prevails the logic of exhaustion, as another kind of grand thinking that everything can be and should be exhausted or formalized to know the world better. When refocusing on the logic of illustration, and considering HI an ongoing accomplishment of reflexive accounting practices, it may be beneficial to some cutting developments of AI, such as empathy, friendliness, and unbiasedness.

Since ethnomethodology studies how social members make social settings “observable and accountable”, it is also contributable to understanding what it means and how is it possible that AI can make things “observable and accountable”. With the distinction, HI illustrates the world, while AI exhausts it. HI makes sense out of ambiguity, whereas AI reduced ambiguity to certainty by exhausting those that cannot be exhausted. In the section “From illustration to exhaustion: the emergence of homo technicus”, homo technicus has been defined as “a species that makes everything as exhaustive as possible”. With a discussion of Stiegler’s playing the two words, default and de-fault, a difference between HI and AI can be further pointed out. For HI, de-fault generates default, whereas for AI, its default, given from the outside, based on the logic of exhaustion, eliminates de-fault, that is, HI. In other words, de-fault creates distance between humans and their world, while default is to equate AI to its world or to make the world presented through AI (as algorithms, codes, or anything formalized).

In terms of Garfinkel, HI’s de-faulting, illustrating, and making social settings observable and accountable require an operation of “reflexivity”. The operation of reflexivity intertwines with plenty of indexical expressions. Both reflexivity and indexical expressions are lacking in AI in that for the latter everything is stored without any loss, and nothing can be forgotten, hence leaving no room for interpretation. In a word, the emergence of homo technicus paves the way for the logic of exhaustion, which means not only AI and machines are a result of it, but also human beings start to observe and think of their environment with the logic. Following ethnomethodology, the logic of exhaustion requires no reflexivity.

With Regard to social members’ mastery of natural language, it makes sure it has enough complexity to deal with those that cannot be exhausted. Natural language is illustrative, reflexive, and open to its environment. It also means that natural language is different from machine language in that it can tolerate ambiguity as HI does. Here comes back to the distinction of indexical and objective expressions, social members’ mastery of natural language can be observed in their “doing formulating”. Doing formulating indicates a formal structure that can be found through social members’ remedying indexical expressions. Mastery does not mean that social members achieve a complete comprehension or formalization of those indexical properties of natural language which is what algorithms do. Rather, mastery indicates an ability to approach and deal with those indexical expressions. Doing formulating not only refers to a formal structure, visible or not, but also tells an invisible story about how the logic of exhaustion works for it.

The discussion of the “et cetera” principle proposed by Garfinkel provides assistance in understanding how exhausting everything is not realistic in social members’ accounting practices. The principle explains that applying rules to something is impossible to exhaust it with those rules. For HI, the use of the principle means that exceptions are not a product of rules, but become part of them. It is one thing to seek a formal structure through studying social members’ doing formulating, it is quite another thing to apply this found formal structure to designing AI that can make social settings accountable, not just account-able, as social members do. Therefore, the et cetera principle provides another way to elaborate on how social settings, including conversations, are made observable and accountable, and it presents a great entry to distinguish HI from AI, the indexical from objective expressions, and the illustrative from exhaustive logics. In this section, two examples, the use of emoji and the “word vector” technology, are used to demonstrate how the trend is still toward the logic of exhaustion.

The way worked by HI, presented as the logic of illustration, is to enrich our understanding of the world, rather than to decide which parts of the world are “correct”. The latter is the logic of exhaustion. Regarding a possible transformation from human to post-human, the logic of exhaustion plays a crucial role in it. Not only will the biological brain gradually adapt to this exhaustive-logic-based technological intervention, which will evaluate everything with it, but a corresponding social mind will also be developed to support this logic. The ultimate goal of this logic is to exhaust everything, including the biological brain, with either perfect simulation or a brand-new designed algorithm. Even if we cannot truly understand the relationship between the biological brain and the social mind, we can still perform simulations in a manner that views these two as equivalent and treat these simulations as an alternative reality. The true reality—if one exists—may no longer carry any importance, as depicted in the movie Ready Player One; as the creator of the OASIS, James Halliday, states, ‘reality… is real’.