How to Do Things with Words

From J.L. Austin's speech acts to fuzzy logic, semiotic control theory, and ChatGPT

John Langshaw Austin’s How to Do Things with Words, based on lectures he gave in 1955 and published as a book in 1962 (two years after his death from lung cancer at the age of 48), challenged the positivist idea that all meaningful linguistic utterances are descriptive and can be assigned a propositional truth value. During World War II, Austin was the leader of a massive intelligence group known as “The Martians,” which was responsible for aggregating, processing, and analyzing vast amounts of diverse intelligence data in preparation for D-Day. (Thomas Nagel’s excellent London Review of Books article on Austin has a detailed discussion of Austin’s wartime activities and how they have influenced his philosophical project.) As part of his “ordinary language philosophy,” Austin proposed a theory of speech acts centered around the notion of a performative utterance—that is, an utterance whose role is not to describe the world but to produce an action on or in it. For such performative utterances, the notion of truth does not make sense; what makes sense is whether they succeed or fail in a given situation. Accordingly, Austin classified such performative speech acts as felicitous or infelicitous. In a way, infelicitousness is the ordinary language philosophy counterpart of logical inconsistency of a formal system in positivist linguistic philosophy. Yet, it is a much richer concept than logical inconsistency. If you make a promise and then fail to make good on it, then the performative act of making that promise was infelicitous. If you make a joke and it doesn’t land, that’s infelicitous too. This is related to the concept of “open texture” of language, a term first used by Friedrich Waismann (a card-carrying member of the Vienna Circle) to describe the inherent vagueness or indeterminacy of empirical statements made in a natural language. For example, if I say that I am in a bad mood and then someone asks me to list all the reasons for it, I would not be able to come up with a complete, exhaustive list.

Austin’s philosophy of speech acts recognizes the prescriptive and regulative role of language. That is, it views language as a control technology. Language is a means for organizing and structuring action and interaction; as a control technology, it is also a complexity-reducing device.¹ However, the open texture of language distinguishes it from control as programming, where one could (in principle) formally verify the correctness of a given program and, at the very least, expose any logical inconsistencies. Hilary Putnam’s Representation and Reality took the concept of open texture further, arguing that it presents a fundamental challenge to purely computationalist approaches to truth, intentionality, and semantics.

Now, while Austin and the members of his Oxford circle were influential in the realm of Anglo-American analytic philosophy at least for a while², it was never their intention to endow the idea of “doing things with words” with a computational character. Natural language was messy, embedded, and embodied; computation was clean, abstract, and disembodied. Thinking of felicitous or infelicitous speech acts in quantifiable terms would have never occurred to them, you just knew it when you saw it. Paradoxes of self-reference and quandaries of undecidability were bad enough for computational theories of cognition. How could you possibly compute with words?

As it happens, computing with words was exactly what Lotfi Zadeh wanted to do from the moment he introduced the idea of a linguistic variable and the theory of fuzzy sets in the 1960s. Zadeh was born in Baku, Azerbaijan, in 1921, moved to Iran at age 10, graduated from University of Tehran with a degree in Electrical Engineering in 1942, and was a faculty first at Columbia University and eventually at Berkeley after getting his MS degree at MIT in 1946 and PhD at Columbia in 1949. Zadeh’s profile in IEEE Spectrum mentions that, “as a child, [he] was surrounded by governesses and tutors, while as a young adult, he had a personal servant.” He was a pioneer in systems and control³ who invested a great deal of effort into promoting the conceptual unity of electrical engineering and computer science. One of his motivations for developing fuzzy logic was the disconnect he perceived between the high effectiveness of mathematical system theory in dealing with “mechanistic systems” and its low effectiveness in dealing with “humanistic systems.” In a 1961 paper called “From circuit theory to system theory,” he wrote:

There is a fairly wide gap between what might be regarded as ‘animate’ system theorists and ‘inanimate’ system theorists at the present time, and it is not at all certain that this gap will be narrowed, much less closed, in the near future. There are some who feel that this gap reflects the fundamental inadequacy of conventional mathematics—the mathematics of precisely-defined points, functions, sets, probability measures, etc—for coping with the analysis of biological systems, and that to deal effectively with such systems, which are generally orders of magnitude more complex than man-made systems, we need a radically different kind of mathematics, the mathematics of fuzzy or cloudy quantities which are not describable in terms of probability distributions. Indeed, the need for such mathematics is becoming increasingly apparent even in the realm of inanimate systems, for in most practical cases the a priori data as well as the criteria by which the performance of a man-made system are judged are far from being precisely specified or having accurately-known probability distributions.

Although the machinery of fuzzy set theory and fuzzy logic was framed in mathematical terms and could be conceived as just another nonstandard logic, it had a polarizing effect that was amplified by the culture wars and the science wars of the time. Rudolf Kalman, another pioneer of systems and control and also someone who was not exactly known for a diplomatic disposition, wrote the following in 1972:

I would like to comment briefly on Professor Zadeh’s presentation. His proposals could be severely, ferociously, even brutally criticized from a technical point of view. This would be out of place here. But a blunt question remains: Is Professor Zadeh presenting important ideas or is he indulging in wishful thinking? No doubt Professor Zadeh’s enthusiasm for fuzziness has been reinforced by the prevailing climate in the U.S.—one of unprecedented permissiveness. ‘Fuzzification’ is a kind of scientific permissiveness; it tends to result in socially appealing slogans unaccompanied by the discipline of hard scientific work and patient observation.

The mathematician William Kahan was similarly unsparing:

‘Fuzzy theory is wrong, wrong, and pernicious.’ says William Kahan, a professor of computer sciences and mathematics at Cal whose Evans Hall office is a few doors from Zadeh’s. ‘I cannot think of any problem that could not be solved better by ordinary logic.’ What Zadeh is saying is the same sort of things ‘Technology got us into this mess and now it can’t get us out.’ Well, technology did not get us into this mess. Greed and weakness and ambivalence got us into this mess. What we need is more logical thinking, not less. The danger of fuzzy theory is that it will encourage the sort of imprecise thinking that has brought us so much trouble.’

It is not my intention here to debate the merits of fuzzy logic. The interesting part is the genealogy of ideas that eventually led Zadeh to proclaim in a 1996 paper that “fuzzy logic = computing with words” and then, in another paper published in 1999, to argue that, in contrast to computing with numbers that represent measurements, computing with words (or, more broadly, with propositions in natural language) is about manipulation of perceptions. Unlike idealized numerical measurements, perceptions are often vague, flexible, fallible, context-dependent—in other words, they perfectly exemplify the open texture of experience even when (or especially when) described in natural language. In that same paper, Zadeh says:

We cannot build robots which can move with the agility of animals or humans; we cannot automate driving in heavy traffic; we cannot translate from one language to another at the level of a human interpreter; we cannot create programs which can summarize nontrivial stories; our ability to model the behavior of economic systems leaves much to be desired; and we cannot build machines that can compete with children in the performance of a wide variety of physical and cognitive tasks.

What is the explanation for the disparity between the successes and failures? What can be done to advance the frontiers of science and technology beyond where they are today, especially in the realms of machine intelligence and automation of decision processes? In my view, the failures are conspicuous in those areas in which the objects of manipulation are, in the main, perceptions rather than measurements. Thus, what we need are ways of dealing with perceptions, in addition to the many tools which we have for dealing with measurements. In essence, it is this need that motivated the development of the methodology of computing with words (CW)—a methodology in which words play the role of labels of perceptions.

The list of failures, as Zadeh presented them in 1999, is a mixture of things we still cannot do well (that includes robots, large-scale economic modeling, or artificial machines that can perform on par with children in certain physical or cognitive tasks) and of things that today’s large language models can perform with ease (translation from one natural language to another or summarizing a story). On the one hand, it would certainly be apt to describe what LLMs are doing as a species of computing with words. On the other hand, the range of capabilities posited by Zadeh as the goal of computing with words is much wider than what LLMs can do. What can account for this?

I would like to offer an explanation which is partly historical and partly conceptual. While Zadeh’s fuzzy decision theory originally drew sharp criticism in the West⁴, it caught on in Japan and in the Soviet Union. In Japan, it was mostly adopted in consumer electronics and home appliance industry⁵; in the USSR, though, it fell on the fertile soil of a uniquely Soviet approach to cybernetics. Starting in the 1960s, there was a great deal of dialogue between Soviet cyberneticists (who, unlike their Western counterparts, readily embraced both the analog and the digital methods) and psychologists, who were interested in the art and science of human problem solving. This interest brought together the mathematician Dmitrii Pospelov and the psychologist Veniamin Pushkin, who were convinced that Herbert Simon and Alan Newell’s formalization of complex problem-solving as pattern recognition and local search (what Pospelov and Pushkin termed the “maze approach”) did not account for human creativity and insight the way their own “model approach” did. According to Pospelov and Pushkin, what Simon and Newell described was only the final stage of a more complex process involving the construction of an appropriate linguistic description of the problem that would allow the problem-solving agent to reason coherently from the starting conditions to the desired goals. In other words, before you can even begin making local moves in the maze, you have to first build the maze and figure out the rules of the game, and that’s where creativity and insight would come in. They presented these ideas in their 1972 book Reasoning and Automata, where they made the connection to the problem of control of large-scale complex systems.

The first thing to note was that the notion of large-scale complex system was ill-defined. Crediting Mikhail Bongard, another prominent Soviet cyberneticist, Pospelov and Pushkin argued that we should apply this label to those systems that are not amenable to strict formalization. In fact, such formalization may not even be desirable in the first place. In his 1986 monograph Situational Control: Theory and Practice, Pospelov listed the salient features of such nontraditional (or ill-defined) control systems:

1. Uniqueness.

2. Lack of a formalizable purpose.

3. Lack of optimality.

4. Dynamic or evolutionary nature.

5. Incompleteness of description.

6. Free will.

Taken together, all of these factors call for a new approach to control, so-called situational control. This terminology reflects the notion of situation as a summary of all relevant information pertaining to the current configuration,structure, and function of the control system. The process of controlling such a system amounts to selecting one control action from a finite repertoire of available actions that would cause a change from one situation to another. In a certain sense, the notion of situation is related to the notion of state: Both are meant to capture the current information available to the decision-maker for the purposes of prediction and control, but the key assumption underlying the construction of the state space is that of completeness, namely, that all relevant system attributes are amenable to observation, categorization, and abstraction. Unlike the state, the concept of situation reflects all the above peculiarities of working with large-scale complex systems and is inevitably influenced by subjective inputs, biases, and preferences of system operators and users.

This brought to the foreground the role of natural language as the only available means of describing and controlling such systems. They recognized that natural language could be the future of computer programming decades before Andrej Karpathy’s viral tweet about English being “the hottest new programming language.” From the very beginning, situational control theory recognized the open texture of language as both a challenge and an opportunity for designing radically new approaches to control. The approach they took was similar in spirit to Zadeh’s construction of fuzzy logics that starts with standard propositional logic and then “fuzzifies” it by relaxing the Boolean set membership criterion to a membership function taking values in the interval [0,1]. In the same vein, situational control theorists took as their starting point the standard notion of a formal system from logic. A formal system is composed of axioms, transformation rules, and a semantic interpretation. As such, it is a closed system—all semantically valid statements can be derived from the axioms by mechanical application of the transformation rules. Situational control theorists proposed to “open up” formal systems by making all their ingredients amenable to modification based on external inputs (e.g., by hooking them up to a learning system). This, they argued, transformed the elements making up the formal system into signs as they are studied in semiotics—that is, entities that can stand in relation to themselves, external objects, and external subjects via their respective syntax, semantics, and pragmatics imposed in a given context by convention. So, the term “semiotic control” and “semiotic system modeling” was also applied.

The function of natural language in semiotic control theory was to provide the control engineer with a means for expressing concepts, relations, and functional roles that are relevant for the control problem at hand. So, just like Zadeh’s fuzzy logic that started with the recognition of inherent vagueness of natural language for describing perception but ended up with yet another formal system, semiotic control theorists developed a formal Situational Control Language that bore very little resemblance to the open texture of natural language. They were remarkably open-minded and broad in their adoption of various formalisms—in addition to Zadeh’s fuzzy logic, Pospelov’s book discusses variants of temporal logic, causal logic, pseudo-physical logic, John Stuart Mill’s inductive logic, and Indian Navya-Nyāya logic. It has philosophical sections on concept formation, Piagetian psychology, and the fundamental incompleteness and absurdity of knowledge.⁶ Ideas from cybernetics and pattern recognition, such as the perceptron or clustering, appear as well. The historical overview at the end of the book traces the development of ideas that led to situational control theory and lists some practical successes in Soviet industry (shipping port logistics, control systems in aviation, oil refinery management, medical diagnostics, etc.). Nevertheless, just like fuzzy logic, situational control is now mainly a historical curiosity because, just like fuzzy logic, it did not exactly give up the idea of fully formalizable system design. For all their talk of ill-defined boundaries and vagueness of natural language, both fuzzy logic and situational control envisioned systems as symbolic and propositional all the way through. Thus, they ended up in the same blind alley as GOFAI, eventually outstripped by connectionist methods like deep neural nets and, now, LLMs and other Transformer-type architectures.

If we now return to Zadeh’s list of canonical tasks that rely on computing with words, we can see why some of the things he lists are now easily within the capabilities of state-of-the-art LLMs (such as machine translation or text summarization), while some are still beyond their reach. As I wrote in one of my posts about Daoism and AI, replacing fully propositional symbolic systems with massive connectionist architectures that rely on words (or tokens) as an information interface with their environment had momentous consequences. Some tasks can indeed be handled by systems that take tokens as inputs, generate tokens as outputs, but look nothing like “theorem provers” internally. Lotfi Zadeh’s insistence on referring to perceptions as amenable to “fuzzy” tokenization was directionally correct, but limiting tokens to words in a natural language was not. In the abstract of his 1999 paper, he wrote that

computing with words (CW) is inspired by the remarkable human capability to perform a wide variety of physical and mental tasks without any measurements and any computations. Familiar examples of such tasks are parking a car, driving in heavy traffic, playing golf, riding a bicycle, understanding speech, and summarizing a story.

However, most of the above examples are not instances of computing with words. Rather, they are examples of computing with natural signs, some of which may be words in a natural language, but many are not. Moreover, they do not admit an easily quantifiable metric of success like perplexity or cross-entropy or expected reward. The considerations of open texture and felicitous acts loom large, as they do in any context involving biological organisms or any other embedded and embodied system. Biological evolution has endowed us with incredible capacity for doing things with natural signs. Language is just a small part of it, and there is now a great deal of work using Transformer-like architectures on suitably tokenized nonlinguistic data (vision, motion, etc.). In fact, according to the “throwing hypothesis” put forward by the neuroscientist William Calvin, our ability to “compute” with language could have emerged by recruiting existing neuronal architectures that had evolved for better control of motion sequences, including timing, planning, and the use of forward predictive models with token-like information interfaces. If that’s the case, then “computing with words” is just a special case of “computing with signs,” the architectures for which were already present in hominid brains.

1. Mark Wilson develops this theme further in Physics Avoidance, when he talks about the reduction of syntactic complexity afforded by the use of language to announce the “shifting of investigative moods” in complex problem-solving.

2. Some interesting intellectual genealogies can be traced here. For example, Gilbert Ryle, a member of Austin’s circle, was Daniel Dennett’s advisor.

3. With Charles Desoer, he coauthored one of the first textbooks on mathematical theory of linear systems.

4. With some exceptions, such as Richard Bellman (the inventor of dynamic programming), who had coauthored several papers with Zadeh on fuzzy decision-making and on pattern classification.

5. Zojirushi rice cookers are amazing. Is it because of their neuro fuzzy technology?

6. Remember, we are talking here about a book published in the Soviet Union, although roughly a year into the beginning of perestroika.