Introduction
In this paper, we will review Erik J. Larson’s book, The Myth of Artificial Intelligence: Why Computers Can’t Think the Way We Do.1 Larson’s observations on current AI algorithms and his discussion on differences between human and machine intelligence prompt us to think deeper about AI, human and machine intelligence, AI’s function in the techno-industrial system, and the conditions of its development. And we should think deeper on these issues to see beyond the haze of utopian or dystopian sensationalist claims about AI. According to Larson, the myth about AI is the belief that the field of AI is approaching towards a generally intelligent machine by the methods it utilizes as of now. According to this myth, no conceptual breakthrough is necessary; we only need to develop further the current algorithms and hardware to reach artificial general intelligence (AGI.) Moreover, the myth purports that the coming of AGI is inevitable. The myth of AI has another dimension related to this: Human mind is a more developed version of current machine learning algorithms, but it is not fundamentally different from them. Since there is no fundamental difference between the human mind and current AI, further developments in AI will eventually reach a human-like general intelligence. Larson says that this aspect of the myth denigrates the human mind, causes distrust about its abilities, and might constitute the biggest obstacle to reaching AGI.
We agree with Larson that today’s AI algorithms and the human mind are fundamentally different. But we reach this conclusion from a different perspective. Larson has a humanistic view of human intelligence. He sees the human mind as a general-purpose cognitive organ. This view regards human intelligence as something unique with a mysterious essence. Though Larson never explicitly says that the human mind has a mysterious essence that makes it unique, from his observations about the human mind and his discussion about the differences between human and machine intelligence, it is clear that his view of the human mind is humanistic. This view inevitably tends to regard human mind as unique and mysterious. In contrast, we believe that the human mind is shaped by evolution. It is adapted to solve reproductive problems our species faced during its evolutionary history. The human mind is fundamentally different from today’s AI (and will be fundamentally different from a future possible generally intelligent machine) because evolutionary pressures humans faced were unique to our species. Today’s AI algorithms are developing in fundamentally different conditions for solving different problems. That is why human intelligence is fundamentally different from machine intelligence, not because it has a mysterious capability that makes it a general-purpose cognitive organ.
We begin by discussing the differences between humanistic and evolutionary views of the human mind. We then summarize the obstacles to AGI discussed by Larson. This discussion allows us better understand where the field of AI stands today and the characteristics of today’s AI systems. We then discuss the aims of the techno-industrial system for developing AI. What could be the expectations of its developers from AI? This discussion takes us to think about the role of AI in the system. To what problems and fields it could be applied, and what could be the consequences of its use?
Humanistic vs. Evolutionary View of the Human Mind
According to Larson, the human brain is a general-purpose cognitive organ. It is not narrow and task-oriented. It can learn and perform diverse tasks such as playing chess, driving a car, speaking a language, etc. Larson says that “intelligence of the sort we all display daily is not an algorithm running in our heads, but calls on the entire cultural, historical, and social context within which we think and act in the world.” He thinks that the human mind has a general “learning” ability. The human brain does the same things when it learns to speak, drive a car, solve mathematical equations, and detect cheaters in social situations. Larson’s conception of the human mind, the humanistic view, is prevalent in social sciences. According to this, the human mind consists of general purpose or content-independent cognitive mechanisms. The human mind has an essence that makes it conscious, which gives it the ability of understanding and learning. As a domain-independent general-purpose information processing system, the human mind would learn everything from scratch by observing its environment. Thus, it would be equally easy for human minds to learn to walk, speak, and engage in abstract logic.
Larson contrasts these purported abilities of the human mind with the narrowness of today’s AI systems. AI systems are not general-purpose, and they are task-oriented. For example, an AI system that can play GO and beat the best human players cannot play chess. Larson claims that without a change in the perspective and a conceptual and theoretical breakthrough, today’s approaches to artificial intelligence wouldn’t be able to surmount the narrowness problem and reach human-like general intelligence. Larson’s critique of the AI field and his explanations about the shortcomings of current methods are illuminating in understanding where the field stands today. Some of his observations offer antidotes against the exaggerated claims of doom or boon about the inevitable coming of AGI and superintelligence. However, since his starting point is an erroneous conception of the human mind, he tends to underestimate the potential of the AI field.2 Despite emphasizing page after page the fundamental differences between human and machine intelligence, he falls into the trap of anthropomorphism. His only reference point for intelligence is the erroneous humanistic view of human intelligence. He cannot see that Darwinian selection could shape fundamentally different intelligence through different evolutionary pressures. To appreciate this fact, we should turn to evolutionary psychology and its findings about the human mind.
Larson divides the capabilities of intelligence into two domains: System X capabilities (narrow capabilities such as playing chess, making complex arithmetic computations, memorizing the whole Internet, and coughing up this memorized data), and System Y capabilities (broad capabilities such as understanding, creating novel insights and ideas, reaching synthesis using available information, etc.) Larson says that we can build algorithms exceptionally intelligent in System X capabilities. But we cannot do the same thing about System Y capabilities because we do not know how to simulate those capacities in code; we do not know how to reflect them in algorithms. And without System Y capabilities, we wouldn’t be able to design machines that have general intelligence.
According to Larson, we should design machines with System Y capabilities to create AGI. Larson says that we are not able to simulate those capacities in code. But this is not an accurate picture of the reality. In fact, the problem might be more fundamental than that. We not only cannot code those capabilities in algorithms, we don’t know what these capacities are and how our brains manifest those capabilities. Understanding, learning, creativity, etc., are concepts that are extremely broad. They are inadequate to describe what really happens in our brains when we “learn,” “understand,” or “create.” For example, “learning” gives the impression that the human mind does the same thing when it learns different things. As if “learning” is a well-defined cognitive operation that is the same for every talent: Learning to drive a car and speaking a language is the same. But this is not so. Neural mechanisms for learning a language and driving are distinct, and they are at different parts of the brain. In the case of language, we have specialized structures shaped by evolution to make us speak and understand the spoken language: Wernicke’s and Broca’s areas. Therefore, learning a language isn’t the consequence of general learning ability. It is the consequence of neural mechanisms that evolved to make us speak and understand what is spoken. Evolutionary psychology has demonstrated that the human mind is not a general-purpose cognitive organ but a bundle of specialized cognitive mechanisms that evolved to solve recurrent reproductive problems.
Larson criticizes today’s AI systems as being narrow. He says that they are only good at solving specific tasks. For example, an algorithm designed to play Go cannot recognize images; an algorithm designed to recognize images cannot chat with humans. Even for seemingly similar tasks like playing different board games, an algorithm designed to play Go cannot play chess. Larson contrasts this narrowness of AI with the general character of the human mind. We can learn different tasks and perform those tasks reasonably well. The same person can play go, chess; recognize images; drive a car; speak languages, and understand spoken words. But, as we said above, evolutionary psychology has demonstrated that we are good at many tasks not because we have a general-purpose brain but because we have many special-purpose neural mechanisms that evolved to solve different recurrent evolutionary problems.
These mechanisms evolved to solve recurrent adaptive problems such as survival and growth (food acquisition and selection, finding a place to live, combating predators and other environmental dangers, combating disease), mating strategies (short-term and long-term mating strategies of men and women), problems of parenting, problems of aiding genetic relatives, etc. The human brain consists of layer-on-layer of these ad hoc mechanisms. According to this view,
Rather than a single general intelligence, humans possess multiple intelligences. Rather than a general ability to reason humans have many specialized abilities to reason, depending on the nature of the adaptive problems they were designed by selection to solve. Rather than general abilities to learn, imitate, calculate means-ends relationships, compute similarity, form concepts, remember things and compute representations, evolutionary psychology suggests that the human mind is filled with many problem-specific cognitive mechanisms, each designed to solve different adaptive problems.3
Some evidence shows that not even our logical reasoning is general or content-independent. Humans are better at reasoning and solving evolutionarily relevant problems, problems that strongly affect our reproductive success. One of the clearest demonstrations of this fact is the Wason Selection Task. In this task, when the question (conditional hypotheses of the form If P then Q is violated or not) is formulated as an abstract problem, less than 25% of college students complete the task correctly. When the question is formulated in a way that requires the participators to detect cheaters, their performance increases dramatically. This time more than 75% of the subjects completed the task correctly. See below two Wason selection tasks that have identical logical structures:4
Humans are vastly better at solving the same logical problem when it requires them to detect cheaters. When the problem is formulated as in “b” above, it is easy to see why we need to check what the 16-year-old is drinking and the age of the person drinking beer, and it is equally easy to see why it is not necessary to check what the 25-year-old is drinking and the age of the coke drinker. It is much harder to reach the same conclusion at “a” for “D” and “7” when the same problem is formulated abstractly.”5 People are better at solving the problem in “b” because detecting cheaters is an evolutionary relevant problem. People who cannot detect cheaters will forego their resources to cheaters and will be at a reproductive disadvantage. People who are good at detecting cheaters will outproduce those who are not good at it. We are good at solving social contract problems because we have specially evolved cheating-detecting algorithms.
The human brain isn’t content-independent and general purpose for good evolutionary reasons. Since adaptive problems (finding food, shelter; decision on fight or flee; mate selection; etc.) have extremely wide-range applications, learning all the instances and different forms of them one by one would be very inefficient. Humans make decisions on how to behave in evolutionary-relevant situations, and these behaviors have consequences on their reproductive success. However, individuals cannot assess the long-term reproductive success of their decisions. Only natural selection can do that. That means that individual “learning” cannot use the result of those decisions as criteria for successful behavior. As a result, Darwinian selection acts on the specific neural mechanisms that create those behaviors and selects those mechanisms that produce relative reproductive success. These are innate mechanisms that evolved to solve recurrent adaptive problems. These mechanisms lead us to certain behaviors in evolutionarily relevant occasions: preferences in mate selection, cheater detection, fight and flight responses, suspicion against strangers, disgust about certain smells and sights, compassion and aggression depending on the situation, etc. We have built-in motivational systems such as the dopamine system and emotions and feelings that drive us to certain behaviors. These are the things that make us autonomous.
These observations demonstrate that human intelligence isn’t general-purpose as today’s machine learning algorithms. The human brain is the product of an evolutionary process, and it was designed by natural selection to solve recurrent adaptive problems. The human brain is a bundle of these ad hoc mechanisms (algorithms) that are adapted to solve specific problems. So, the fact that today’s machine learning algorithms are content-specific and designed to perform on narrowly-defined tasks cannot be used as an argument to rule out the possibility that today’s machine learning algorithms could be scaled-up to fully autonomous artificial intelligence. Today’s AI systems are good at fulfilling specific tasks because they are designed that way. But this does not mean they are bound to remain separate and cannot communicate with each other. Different algorithms specialized in diverse tasks might communicate with each other and constitute the layers (ad hoc mechanisms) of a future general machine intelligence. Today’s narrow algorithms might be the rudimentary first components of a future fully autonomous machine intelligence. For example, more general-purpose algorithms like ChatGPT could call narrower algorithms that specialize in specific tasks (playing chess, image recognition, etc.) to carry out those tasks.
Larson never considers this possibility and criticizes Watson, an AI system developed by IBM to challenge the human contestants at Jeopardy!, precisely for using subsystems. He says that Watson uses a strategy based on the idea of a society of minds. It consists of numerous sub-modules responsible for specific tasks and reaches its final decision via the concerted action of these sub-modules. Larson discusses the techniques and algorithms Watson uses to show that it doesn’t have any real understanding. From his observations on Watson, one gets the idea that Larson regards as intelligent only those cognitive operations that we cannot explain explicitly. If someone were to explain in specific terms what constitutes exactly human understanding, the exact information processing mechanisms of the human mind, Larson would probably say that humans don’t have real understanding. He wants intelligence to be something mysterious, and when its constituent parts and the exact operations that produce it are explained, it becomes a pseudo-intelligence. According to Larson, the intelligence demonstrated by Watson is not genuine because so much human planning and care went into its design, because it uses lots of specific sub-modules to decide, because it fetches the answers from Wikipedia, etc, etc. For Larson, genuine intelligence is something we cannot explain.
Human brains evolved under specific evolutionary pressures that were unique to our species. Machine intelligence also evolves subjected to Darwinian selection. Large organizations (corporations and states) develop AI algorithms. These organizations are in an unconscious competition with each other. Organizations that have the most conducive characteristics to their propagation and survival are the ones that survive and propagate themselves. These organizations use AI to solve practical problems they encounter. They try to increase their efficiency and gain an advantage in their competition against other organizations. AI systems that bestow benefits to these organizations will be selected (through an unconscious Darwinian selection) during this process. So, evolutionary pressures operate on machine intelligence as well. Machine intelligence also tries to solve adaptive problems. However, these adaptive problems would be very different from the problems human minds evolved to solve. As a result, machine intelligence probably would be quite distinct from human intelligence. But this doesn’t mean that machine intelligence cannot reach full autonomy (and this is basically what people generally mean when they talk about general intelligence). After all, human intelligence is a product of material factors. It doesn’t have a unique essence (an essence beyond matter) that makes it fundamentally different from other intelligences.
However, the AI field should surmount formidable obstacles to reach that level. Larson’s discussion of these obstacles gives us a good understanding of where the field stands now and the hurdles ahead to reach that level. The road goes to that level (if it ever arrives it) is long and arduous.
Obstacles to Artificial General Intelligence
I. Problems of Deduction
Until the 1990s, the AI field was dominated by the deductive approach. According to this, programmers would spoon-feed the machines virtually all the possible deductive inferences. They would construct the thinking ability of the machines from scratch. For example, the Dartmouth workshop proposed this:
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. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.6
Of course, these researchers would understand that it wasn’t so easy “to describe every aspect of learning or any other feature of intelligence,” and “using language, forming abstractions and concepts, solving kinds of problems now reserved for humans” were not the kinds of things that could be programmed in one summer. The number of deductive inferences that needed to be determined and coded into machines was impossibly high. In fact, as Larson demonstrates, even if it were possible to spoon-feed all the possible deductive inferences to a machine, this wouldn’t be enough for general intelligence because deductive reasoning has inherent shortcomings.
A well known example of a deductive reasoning is below:
All men are mortal.
Socrates is a man.
Therefore Socrates is mortal.
The first two statements are premises, and the last one is the conclusion that we infer from these premises. As Larson says, “deduction supplies a template for ‘perfect’ and precise thinking for humans and machines, and primarily for this reason it has been investigated in mathematics and the sciences, and used successfully in several important applications in the field of AI.” But Larson adds the following caveats about deductive reasoning: it is relatively easy to determine whether deductive arguments are valid or not using formal rules. The argument is valid if the conclusion is true whenever the premises are true. But this doesn’t say anything about the soundness of the argument. For example, consider the following valid argument:
If it’s raining, then pigs will fly.
It’s raining.
Therefore, pigs will fly.
In this argument, the premises are false. So, deductive rules themselves don’t say much about the truth of those arguments. We need to rely on empiric knowledge to know the truth value of premises, and this depends on commonsense knowledge about the world. Deductive reasoning can be coded into machines, but without commonsense knowledge, machines would make silly errors.7 On the other hand, deductive reasoning, on its own, wouldn’t solve the problem of relevancy. Larson gives the following example:
All males who take birth control pills regularly do not get pregnant.
A man takes his wife’s birth control pills regularly.
Therefore, the man does not get pregnant.
That is a sound and valid deductive argument, but it is totally irrelevant because men don’t get pregnant anyway. Larson says that a robot might have vast data of arguments like these, but it wouldn’t understand anything; it wouldn’t know what’s relevant and silly. To surmount the relevancy problem, knowledge of causation is necessary. Machines should know if some event actually brings about a result or makes something happen. However, detecting causality is daunting because real-life events are complex; they have many possible causes. As Larson points out, deductive reasoning cannot establish causality.
As a result of these inherent weaknesses of the deductive approach, the AI field entered into one of its so-called “winter” periods, a period where enthusiasm, hope, and funding for the AI field wanes. However, the coming of the internet and the development of more powerful processors started a new AI summer during the 90s. This time, the field would use another type of logical reasoning: induction.
II. Problems of Induction
Induction is to reach general hypotheses from particular observations. A famous example is that after seeing lots of white swans, someone might conclude that all swans are white.8 Or if we observe a traffic jam in a particular neighborhood after each football match, we would conclude that after this match, we will experience a traffic jam in that neighborhood also. Inductive reasoning gives us the ability to predict. We make predictions based on our previous observations and experiences. However, conclusions reached using induction cannot be proved or guaranteed; they are always provisional. As Larson observes, enumeration (the number of samples we observed) is very important in inductive reasoning. As the number of observations increases, we assume that the conclusion that we reach is closer to certainty.9 An example of inductive reasoning is below:
N swans observed have been white [where N is some large number].
Therefore, all swans are white.
Inductive reasoning has a probabilistic dimension to it as well. A well known example is the following one:
Seventy-three percent of randomly sampled voters are for Candidate X.
Therefore, Candidate X will get about seventy-three percent of the votes.
In the example above, we use an observation based on a sample from a broader population to reach a conclusion about this population. Most of the time, the conclusion reached in this fashion will be valid, but it cannot be guaranteed a hundred percent. Some unexpected factors could cause it to be wrong.
Modern AI (machine learning algorithms) relies on this kind of statistical analysis and inductive reasoning. They detect patterns in a given sample and try to locate where in that pattern a new example would fit. They are simulations of real-life events; they try to approximate them. In fact, they don’t try to simulate the events themselves, only the outputs of these events. Using past samples (spam emails, for example), they try to classify a new occurrence (whether a given email is spam or not). The same logic is applied to different problems such as image and voice recognition, content classification and suggestion, go or chess-playing, etc. Since real-life events are very complex, their machine-learning simulations can only be approximations. Since their knowledge is limited to the sample to which they are exposed, they cannot anticipate unlikely occurrences.
Machine learning algorithms have recently become so successful because Internet has provided lots of data to them. With more data, these algorithms become better at classifying new examples and predicting outcomes, but more data improves their performance to a certain extent. Sooner or later, models get saturated and cannot be improved by adding more data. More data doesn’t solve the problem of long-tail, infrequent events that could have immense consequences in real life. That is the reason, despite achieving spectacular results at tasks that have limited scope (such as playing chess, go, content classification, and image or sound recognition), machine learning algorithms couldn’t yet master the tasks that require them to grapple with all the details and unexpected events of real-life events such as fully automated driving.
In sum, machine learning algorithms demonstrate every weakness of inductive reasoning. As Larson says, in dynamic real-life environments, there is constant change. Change occurs in both predictable and unpredictable ways. That makes decisions based on induction problematic because there will always be things that will surprise the assumptions made based on past experiences. Inductive reasoning might work well in predictable environments with limited rules, such as board games (go, chess, etc.). But machines that rely on induction are not very good at navigating dynamic and complex environments. As Larson says, the main shortfall of current approaches is due to this fundamental problem of induction. Induction, inevitably, requires us to believe that “instances of which we have had no experience resemble those of which we have had experience.”10 In other words, during induction, we have to extend our experiences based on observed examples to unseen examples, and there is no guarantee that they will hold. Larson points out that more powerful learning approaches (data, observation, or equipment based) that rely on inductive reasoning cannot overcome these shortcomings. According to Larson, the AI field needs a new conceptual approach, a fundamentally new method to create artificial general intelligence. Adding new data and computing power to current methodologies (inductive and deductive-based) will not solve the inherent problems of these approaches.
III. Problems About the Common-Sense
Another big obstacle to AGI is the common-sense or world knowledge problem. Minds should have many default assumptions about the external world to navigate in real life. We cannot perform even the most basic tasks without these assumptions. These assumptions are about our environment: the objects in those environments, their relation with each other, whether they are solid, breakable, wet, etc. We know what solidness, wetness, fragileness, and flexibleness mean. We know that solid objects cannot pass through other solid objects. Our minds have many assumptions (the frame) of these kinds to achieve everyday tasks. These assumptions are innate, or we have the capacity to learn them quickly through our experiences. We’re not even aware of their existence most of the time, and we cannot imagine how widespread they are.
Larson says that common-sense is a rich understanding of the world and mainly requires two things: everyday knowledge about the world and inferential capacity to make use of this knowledge. Common-sense knowledge about the world is extremely wide. It includes, among other things, assumptions about the objects in our environment: knowledge about the categories of these objects like man-made tools vs. naturally occurring objects, living vs. non-living things, categories of animals, the difference between animals and plants; possible functions of the man-made tools; cause and effect relations; uncertainty; other agents’ beliefs, desire and intentions; etc. Larson says that a field in computer science (knowledge representation and reasoning) tried to face this challenge to no avail. According to Larson, current approaches to AI cannot surmount this problem. The field needs a conceptual breakthrough, not better ways of doing the same things.
Larson tells us about a project conducted by DARPA that tried to spoon-feed computational systems with everyday knowledge like Living humans have heads, Sprinklers shoot out waters, Water makes things wet, etc. Larson himself was involved in this project and explains why it failed. Larson gives two reasons for this failure. First, most of what we know about the world is implicit. We bring our knowledge into consciousness and make it explicit only when circumstances require it. So, it was not possible to exhaustively spoon-fed our world knowledge into machines. Besides, this implicit knowledge about the world is enormous. Spoon-feeding a computer with common-sense knowledge would take a lifetime, if not more.11 Moreover, even the exhaustive list of all the common-sense data points wouldn’t be enough. In addition to that vast database, machine intelligence should be able to use these data in relation to each other. It should know the connections between the data points and the relations between them.
In sum, it is not possible to teach computers all the things about the real world. We cannot teach them all the things we know about the world because this knowledge is hopelessly broad, and even we ourselves don’t know the extent of our real-world knowledge. Most of this knowledge is implicit and unconscious. On the other hand, computers should put all this knowledge in a structure by which they can evaluate and use them in relation to each other. They should know how each fact in that database is related to another. That is a daunting task because these relations are not static and can arise spontaneously. It means that two items in a real-world knowledge database that look initially unrelated might become related in a real-world event. Therefore, machines should have inference abilities to make sense of this vast knowledge to act sensibly in real-life situations.
Larson’s Magic Solution: Abduction
According to Larson, the field should figure out how to express as computer code abduction to surmount these obstacles. Abduction is a type of logical inference distinct from the well-known two logical inferences, induction and deduction. Larson claims that abduction is the capacity that gives humans the ability to guess, hypothesize, or intuit. These abilities go beyond mere calculation. Whereas calculation connects known dots (applying the rules of algebra, for example), intuition or guesswork explains what these dots mean. It is about reaching an original conclusion by using the available data.
Larson explains with the schema below why abduction, induction, and deduction are distinct logical inferences and cannot be converted into each other.
As can be seen in the table above, abduction is, in fact, a logical fallacy in propositional logic. That is why, says Larson, deduction or induction cannot be converted into abduction. And if intelligent inference requires abduction, we cannot get there through deduction or induction. AI scientists should find a way to code abduction into machines.
Larson defines abduction as reaching conclusions based on observations. However, in abduction, conclusions don’t follow observations mechanically. During abduction, we conclude through hypothesizing. Hypothesizing is very close to guesswork, but it isn’t pure random guessing, either. “We guess, out of a background of effectively infinite possibilities, which hypotheses seem likely or plausible.” Larson gives the below example:
The surprising fact, C, is observed.
But if A were true, C would be a matter of course.
Hence, there is reason to suspect that A is true.
According to this, abduction is concluding about the things that rarely occur. Surprise events that fall at the long tail of a distribution. These surprise events are a problem for machine intelligence only capable of induction. But a human mind, when confronted with such occurrences, can make adequate hypotheses about these occurrences and guess their causes. For example, if Julie doesn’t work on Saturdays, a person who sees her at work on a Saturday would assume all sorts of things that would explain why she is at work: she might be working extra hours, or she might be called in to cover someone who is sick that day. But a machine intelligence that is only capable of induction would face a long-tail problem. Larson says that in our daily lives, we constantly use abduction. Even seeing a flower and deciding to which particular species it belongs is an abduction: We hypothesize about the possible classification of that flower. Besides, we constantly update those conjectures according to the information we receive from the outside world. We keep those conjectures that are corroborated by the facts and discard or revise those that are not. He calls this the defeasible nature of abduction. In short, our understanding of the world is a combination of abductive conjectures on the one hand and testing of these conjectures with inductive and deductive inferences on the other hand. Deduction, induction, and abduction work in conjunction in the human mind:
Once an intelligent agent (person or machine) generates a conjecture [...] downstream inference like deduction and induction make clear the implications of the conjecture (deduction) and provide a means of testing it against experience (induction). The different logics fit together: “Deduction proves that something must be; Induction shows that something actually is operative; Abduction merely suggests that something may be.” Yet it’s the may be –the abduction– that sparks thinking in real environments.12
Larson sees and presents abduction as a magic formula. If AI scientists could figure out a way to code into machines abduction, they would, at last, create an artificial general intelligence. For Larson, abduction is an all-encompassing capability that bestows on the human mind all the things it can do with its great flexibility and generality. We evaluate possible scenarios, create conjectures, see and name the objects we encounter, decide what to do when we face an odd situation, and demonstrate common-sense and real-world knowledge thanks to abduction. We explained above that Larson sees the human mind as a general-purpose cognitive organ. All the abilities of the human mind come from a general and all-encompassing capability. And Larson finds in abduction this all-encompassing capability, sort of a secret ingredient that, when added to induction and deduction, completes the formula and creates a general intelligence. Larson uses abduction in his book (the secret ingredient to “general intelligence,” a general-purpose magic capability) as a blanket word that denotes mechanisms in our brains that we know nothing about their exact neurological functioning. What Larson expresses with abduction is, in fact, the product of a myriad of evolved cognitive mechanisms (which only together make us autonomous and flexible and give us common sense and word knowledge). Abduction is far from explaining what these mechanisms are. By attributing all the crucial abilities of the human mind to one type of inference (abduction), he does the opposite of what he tries to do with his book: he underestimates the problems of creating an artificial general intelligence that would have human-like qualities. If we could code abduction, we would create artificial general intelligence. But the matter is much more complicated than that. As we explained above, evolutionary psychology demonstrates that the human mind does all the things it does not because it has a general-purpose, all-encompassing ability that could be applied to every problem the human mind faces but because it has a myriad of ad hoc evolved problem-solving mechanisms that were shaped by evolution. Automating only one remaining elusive inference wouldn’t be enough to solve the general intelligence problem. AI field should identify and code into machines enough ad hoc problem-solving mechanisms that would give the machines the same flexibility and generality humans have.
Does the AI Field Really Aim to Create a Human-Like General Intelligence?
Larson is convinced that the AI field wants to create an artificial human mind. But how do we know that that is the real aim of the field? It is true that in mainstream polemics, movies, and scary or jubilant superintelligence narratives of some experts, the final goal is portrayed as creating human-like general machine intelligence. That may be so for the individuals that work on AI. And some of the leading figures might claim that they want to create human-like machine intelligence. However, these people don't work in a vacuum, and their projects are funded by large organizations (states and corporations). These organizations are driven by unconscious mechanical factors, not by the desires and wishes of individuals who work for them. They try to solve practical problems they face and increase their efficiency. They develop AI technology for this reason. Circumstances drive them to harness more data and analyze it more accurately. They "want" to track, understand, and manipulate the masses. They "want" better forecasts about extreme weather events to prepare themselves better to reduce the damage. They "want" to solve enormously complex problems human intelligence cannot solve, like the protein folding problem. They "want" to calculate how plasma behaves in a fusion reactor to control it better in a magnetic field. They "want" to eliminate humans from the economy because humans have all sorts of “problems” such as fatigue, lack of motivation, need to sleep and eat, and many other psychological and physiological traits that make them less ideal for these organizations than smooth-functioning machines. These large organizations do not "care" if the machines they create are conscious, have feelings, are creative or appreciate art, really understand their human interlocutors, etc. Since they focus on practical problems they face and the possible benefits of AI technology, they won't expend resources on developing human-like artificial minds. So, machine intelligence will evolve according to the needs of these organizations. What the individual AI researchers or entrepreneurs say about their aims doesn't matter much in practice. What they do in practice and what the large organizations that develop these artificial minds need matter the most.
According to Larson, the only real test to assess whether a machine is intelligent or not is the Turing Test. Larson falls into the trap of anthropomorphizing here. He will only regard those machines as genuinely intelligent if they demonstrate human-like abilities such as understanding, creativity, and intuition. But these are concepts that don’t explain what they denote. We say that we understand something and feel the consequences of “understanding,” but we don’t know what happens in our brains when we understand. We don’t know the precise mechanisms in our brains that make us understand. The same is true for learning, creativity, and intuition. They are ambiguous concepts and convenient for Larson precisely for this reason. Whatever the actual performance of an AI system, he dismisses it as not really intelligent by explaining how it does what it is programmed to do, as if he is unveiling the trick behind an illusion. For Larson, genuine intelligence is something unexplainable and mysterious. Larson would only see those intelligence as genuinely intelligent we cannot explain their operations. If there is an explanation for intelligence, then it is not genuine intelligence; it is only a trick that mimics intelligence. But the human mind has “real” intelligence. Since it has “real” intelligence, it should perhaps have an essence? Something beyond matter which makes it really intelligent? Larson’s thinking seems to go that way. He says that thinking machines will have two inherent problems: originality and initiative. They won’t be able to find their own problems. They will always do the things they are programmed to do. This discussion is about autonomy as much as about intelligence. Humans and other animals have their own purposes and move autonomously to fulfill them. These purposes are the products of evolution, and the ultimate aim of these animals is to perpetuate their genes. During their evolutionary history, they acquired cognitive mechanisms that direct them to do the things that would help them survive and reproduce. Humans, for example, don’t always act consciously by calculating rationally the consequences of their behavior. We reach most of our decisions unconsciously. We make many decisions in our daily lives, and most of them are automatic. So, we do not always have originality and initiative as well. As a good humanist, Larson seems to think that our minds have a mystical trait that makes us intentional free agents: As if we create our intentions and purposes without any influence, and something inside us makes us act freely, following only our willpower. That is not the case. We also do, in a way, the things which we are programmed to do. Our programmer is the evolution, and our programs are the ad hoc neural mechanisms that evolved to solve recurrent adaptive problems. Our minds don’t have any mysterious, unique essence that makes us original, that bestows us free will. What Larson calls the intentionality or originality of humans are, in fact, the manifestations of their autonomy. And machines could also develop full autonomy through Darwinian selection (this time, it would be an artificial selection). Perhaps, today’s narrow AI systems could become ad hoc problem-solving mechanisms of the future, more general and fully autonomous machines.
Larson discusses Stuart Russell’s objections about the Turing Test. In Human Compatible13, Russell points out that Turing Test is a human-centered test, and it is debatable whether the criteria it involves (understanding human language and engaging in an open-ended conversation) is suitable to assess machine intelligence. According to Russell, “the Turing test is not useful for AI because it is an informal and highly contingent definition: it depends on the enormously complicated and largely unknown characteristic of the human mind, which derives from both biology and culture. There is no way to ‘unpack’ the definition and work back from it to create machines that will provably past the test. Instead, AI has focused on rational behavior, and thus a machine is intelligent to the extent that what it does is likely to achieve what it wants, given what it has perceived.”14 Russell emphasizes here that the field of AI focuses on practical problems. And bestowing human-like emotions, drives, urges, and interests to machines might not be among those problems. Take consciousness, for example. As long as AI systems accomplish the tasks they are designed to accomplish, why would the organizations that develop AI "bother" whether these machines are conscious or not? As long as large language models like ChatGPT attract the attention of humans, lure them into speaking and sharing information with them, gather their thoughts and turn them into analyzable data points, inculcate into people the “correct” values, and manipulate their behavior why would their developers bother that if they really understand what the humans are saying? What matters is the competence of AI systems, not their consciousness.
Larson objects to these criticisms and says that they amount to dodging the problem. He says that since the beginning, creating a self-conscious machine was the ultimate aim of the field. And now, after realizing that the consciousness problem is much more difficult than initially assumed, some people declare that it wasn’t their goal anyway. As we said above, Larson thinks that the AI field aims to create a conscious machine. However, apart from reminding us that from Hollywood movies to sensationalist claims and predictions of Ray Kurzweils and Elon Musks, machines that turn suddenly conscious have been the whole focus of mainstream debate about AI, Larson doesn’t explain why consciousness and the whole plethora of human intelligent abilities should be the goals of the field in practice. Mainstream focus on these sensationalist claims doesn’t mean that the AI field really aims to create a human-like conscious machine. This shallow sound and fury about killer anthropomorphic robots or super-intelligent god-like machines that would kill all humans accidentally (or intentionally) or usher a new heaven on earth is mostly entertainment and distraction. It is a propaganda tool to divert attention from the immediate and more “mundane” threats of existing machine learning algorithms.
Function of AI in the Techno-Industrial System
Rather than focusing on this sensationalist propaganda and entertainment about AI, it would be better to look into the matter from a historical perspective to understand the role of AI in the techno-industrial system. Since the agricultural revolution, human societies have become more complex in an accelerated fashion. Their population, area, and the amount of energy and material they used increased constantly. Their activities required ever more data (statistics on production, consumption, and taxes; archives on ownership status and contracts; laws organizing the relationships among people.) This vast data was necessary for the new sedentary, complex societies that emerged with the agricultural revolution. After a certain threshold, this information overload became too much for the mere human brain power to handle. The writing was invented to keep records on production, taxes, ownership statuses, laws, etc. As societies got more complex, writing systems and data collecting, recording, and manipulating technologies developed further: from simpler number systems to Arabic numerals that use value systems on positions and the number zero. Computer technology represents a new phase in this development. It is a direct consequence of information inflation. Data storage and processing technologies have evolved from manual human brain power calculation to mechanic calculators of the 16th century to the big frame computers of the early 20th century to the supercomputers of the 21st century. Data collection technologies have evolved from manual data collection to face-recognition algorithms.
Computer technology has become a necessity, just as writing became a necessity when human societies reached a certain level of complexity. Each step further in the development of data storage and processing technologies has rendered human capabilities (mere brain power in calculating or memorization) in these areas more obsolete. Sometime during the middle of the 20th century, the information load became so huge that it became impossible for humans to manipulate this vast data. Humans continued to program the computers that carried out data storage and manipulation as computers began to replace human calculators.15 Nowadays, algorithms that processors use to manipulate data have become so complex that the capabilities of human programmers are now becoming inadequate to program computers. With the accumulation of vast amounts of digital data and the developments in processor technology, we now witness machines that program themselves (machine learning or artificial intelligence).
In sum, computer technology and AI is the answer the techno-industrial system gives to the ever more complexification of human societies and the resulting information glut. The system used and continues to use human brain power to conduct its activities and solve problems it faces. But human brain power is becoming more and more inadequate for the cognitive tasks the current level of complexity entails. As we explained above, this was the situation for a long time for mere human brain power. Throughout history, various tools have been used to complement human brain power, from writing to abacuses to mechanical calculators. After the industrial revolution, the information glut has increased so enormously that human brain power isn’t sufficient anymore for the cognitive tasks of the system, even with all the appendages that help it. The techno-industrial system needs something more powerful, and AI is a potential answer. The system tries to develop a new cognitive organ for its needs.
We said above that through a process of Darwinian selection fully autonomous machines could develop one day, machines that have artificial general intelligence. This selection process is not unique to intelligent machines. All the cultural artifacts of humans (the products of human civilization not inherited by genes) are subject to Darwinian selection. These artifacts evolve through constant changes, additions, and upgrades. A selection process that is similar to natural selection drives this process. Those artifacts that bring advantages to the functioning of human organizations are retained and diffused. They are retained and diffused because they may increase the energy efficiency of a given process; they may make it possible to exploit a new energy source and create more destructive ways of applying violence; they may increase the speed of communication and transportation or bestow on human societies more effective methods of data collection, storage, and analysis. These cultural phenomena are parts of larger units, human organizations (companies, states, nations, etc.). Darwinian selection operates on these organizations as well. In fact, selection pressures that apply to cultural artifacts first act on these organizations and affect these artifacts through them. Cultural artifacts are selected according to the “reproductive” benefits they bestow on these organizations. Human organizations are integrated systems that have their own metabolisms and dynamics. Their ultimate aim is to perpetuate themselves. That is not something they pursue consciously; it just happens. Those organizations that have and develop the most conducive characteristics to perpetuate themselves are the ones that survive and continue to exist. Those that are not so good at these things are eliminated or absorbed by the more successful ones. So, there is an unconscious, mechanic competition among human organizations. They ought to improve their processes by finding more effective ways of doing things and enlarging their operations. They ought to absorb more material and energy and process them in their metabolisms more effectively.
Artificial intelligence entails enormous possibilities in that regard. We mentioned the problems related to the complexity above. Organizations that could solve these problems would reap great benefits. Another crucial area AI is applied is the control of human behavior, the perennial problem of complex human societies since the agricultural revolution. Unfortunately, the techno-industrial system has made considerable progress in this area in recent years due to AI. The system has gained enormous capacities for surveillance and manipulation. It can now track ever more aspects of people's lives and dazzle them with massive intensity by hacking their dopamine system. It bombards people with a constant barrage of ridiculous and meaningless sounds and images. Many people have lost their ability to concentrate; their attention spans continue to decrease even more. Each new generation is affected even more intensely because each newer generation is subjected to this regime from an ever-early age, and the technologies of electronic lobotomy are progressing with each passing day. This internal siege of the masses makes it easier to tolerate the miserable existence modern life entails. On another front, AI bestows enormous advantages for an external siege through tracking technologies such as face recognition systems and data storage and analysis tools that lay bare desires, fears, relations, and daily practices of people. In Industrial Society and Its Future16, Ted Kaczynski saw the control of human behavior as one of the biggest challenges of the technological system. Despite subjecting them to ever more indignities and forcing them to live ever more artificial lives in ever more unnatural environments since the publication of that work, the technological system has controlled human behavior to a large extent. AI has played an extensive role in this.
Apart from controlling human behavior, the techno-industrial system faces other challenges in its war against Nature. And AI could become one of its most destructive weapons in this war. First of all, AI technology could be applied in diverse sectors. It could work in nearly every domain. Today, it is already in use in such diverse fields as medical image analysis and drug design, algorithmic trading in finance and credit risk assessment, customer sentiment analysis, content generation with language models, in robotics (autonomous weapon systems, drones, automobiles, humanoid robots, manufacturing robots, etc.), the control and prediction of manufacturing processes, forecasting of demand, inventory management, the management of power grids, in tracking and assessing the student performance, in creating customized education programs, in monitoring crops, in precision agriculture, in weather forecasting, etc. In the future, AI could be used in personalized healthcare based on individuals’ genetic and health data, robotic surgery systems might become more sophisticated, and AI systems could improve drones, ships, construction machines, and self-driving cars even more. The system could even try to automate entire cities with their transportation, logistics, waste management, and construction activities managed by an AI as an integrated autonomous system. The complete automation of production could also be possible. AI can create and manage manufacturing systems from the design phase to the production phase. That would entail the integration of the design phase in software and the manufacturing phase in plants. Software programs used in the design of products would be in direct communication with the machines used in manufacturing, and the production facilities would be constructed according to the needs of this design. Imagine these as self-contained machine hives sucking enormous quantities of material and energy and transforming them into mass products. AI would be used in solving complex problems such as advanced materials development. If it could give the system the ability to construct materials atom by atom, the system would try to circumvent the energy and material scarcity it could face by attempting to manufacture them directly. Language models like ChatGPT could evolve to understand better context, emotions, and nuances in human communication. Thus, they could better lure humans to communicate with them and learn their desires, fears, and inclinations more deeply to inculcate better the values and worldviews conducive to the system’s more effective functioning.
The system will utilize these diverse uses of AI to face the challenges related to climate change and biodiversity loss. The system is trying to update its energy infrastructure to adapt to climate change, a consequence of the system’s functioning (burning of fossil fuels). That entails the electrification of energy use and production, the construction of so-called renewable wind and solar power plants, and the renovation of the electricity grids to accommodate this change. Electrification of energy production and consumption entails much more complex power grids that should accommodate the intermittency of solar and wind, with the addition of a massive new consumer of electricity, electric cars. It would need AI’s data analysis and forecasting powers to design and run these complex energy networks. However, these more traditional methods might not be sufficient to mitigate the effects of climate change, and the system could try to take into its own hands the governing of the atmosphere.17 In such a case, it would use AI to create more accurate and detailed climate models. It would try to predict the consequences of its meddling using AI if it tries to intervene directly in the processes of the atmosphere. The system’s massive use of energy and materials and its extractive and transformative activities create all sorts of problems for the ecosystems of the Earth. Its activities are encroaching or even going beyond the planetary boundaries. The techno-industrial system is still dependent on ecosystemic functions to sustain itself. So it will attempt to ration the consumption levels of its members and try to monitor and control the ecosystems to better take advantage of them. The former would entail the near-total tracking of the daily lives of its members to ensure that they won’t go beyond their allocated footprint. The face recognition and image recognition algorithms would track what people buy; AI systems would calculate the material and energy consumption implicit in those purchases. Smart cars would relay to an AI system how many kilometers their owner traversed, etc. For the latter, the techno-industrial system would use AI-powered satellite data and remote sensing technologies to monitor the wild ecosystems. Drones would buzz over the wilderness and collect all sorts of data. The system would check the condition of the species (their population, health, etc.) The accumulation of undergrowth in the forests and the age of the individual trees would be monitored and assessed. With these kinds of data, the techno-industrial system would subject the wilderness to rational control and exploitation to ensure its “sustainable services.” It is easy to see what this would entail: the attempt at total rational control and subjugation of wild Nature.
AI is a manifestation of the system’s fuite en avant. It attacks the problems its existence creates from many angles using AI. However, the continuous development of AI and its application in diverse fields will only increase the system’s complexity and related problems. The system could push the biosphere’s limits using this technology even more, but this would create even more problems in the future. The gradual but relatively fast development of AI, following the cultural development pattern (an artificial Darwinian selection) we explained above, could one day produce a fully autonomous machine intelligence. This intelligence could have enormous capacities to control the physical world due to the advancements in robotics technology. The requirements of the large organizations that develop this technology will shape the characteristics of this intelligence. So, all the current polemics about the alignment problem are futile endeavors. The alignment problem expresses the possibility that future AIs may have goals that do not align with their human creators. This concept, as it is configured today, only deals with the problem of whether specific AI systems will follow the orders of humans and whether they will understand the humans’ intentions correctly or not. However, since AI technology will evolve following a blind Darwinian process, it will be impossible to plan and direct the future developments of AI; it will be impossible to determine the characteristics of future AI systems. They will evolve together with the large organizations that develop them. We cannot control either of them: in the narrower sense, the evolution of particular AI systems, and in the broader sense, the human organizations that develop these AI systems. Besides, the alignment problem would be irrelevant from our perspective, even though it were to be solvable. And we mean this both in the narrow sense that this problem is defined today and in the broader sense of the complete control of the long-term development of AI with the human organizations that develop it. Even in an implausible scenario that humans control perfectly the future evolution of AI, the world these human developers would create will be a hellish nightmare of a “perfect” brave new world. We don’t see a meaningful difference between what these machines and their human developers would do with the capabilities this technology entails. Both alternatives would be equally disgusting.
We mentioned the possibility that the future evolution of AI could create fully autonomous machines. We should think about this autonomy in two layers. At the lower level, there is the autonomy of certain technologies like fully autonomous drones, automobiles, or some software programs. At a higher level, we can talk about the complete autonomy of large organizations (companies, states, etc.) With the advent of fully autonomous machines, large-organizations could become independent from humans. These organizations have their own internal dynamics and metabolisms, and humans cannot control their long-term evolution. In this sense, they are already autonomous. However, as of now, they need humans as actuators. The vital force that moves them is humans. Since these organizations are dependent on the existence of humans, they have a biological aspect. Thus, they could only exist if the ecosystems remain within certain limits, boundaries that allow humans to live. If the future development of AI makes them fully independent from humans, this will make them independent of biological processes as well. In such a case, the techno-industrial system could continue to exist even if the Earth’s ecosystems are thoroughly devastated by chemical or nuclear pollution. Even if all the complex life on Earth vanishes, the techno-industrial system could continue to exist if it could develop AGI before collapsing. That would be the worst consequence of AI for wild Nature.
Larson never mentions such possibilities. His main concern is that the myth of AI relegates the human mind to a lesser position. For this reason, he says, we don’t trust any more individual creativity and the power of geniuses. He fears that this belief aggravates the existing obstacles in front of AGI. His concern is not about AI itself but the myth about AI. Despite repeatedly exalting the uniqueness of the human mind and adoring it, he wants to promote the advent of artificial general intelligence. It is such a paradoxical stance. He doesn’t seem to realize that AGI would inevitably make the human mind less valuable. If there is intelligence that does almost everything better than humans, what importance would the human mind have? If he really values the uniqueness of the human mind, why does he try to promote AGI? If he is not capable enough to see this glaring contradiction, there is only one explanation: he wants to leave himself to the safe bosom of conformism. In today’s world, the veritable dogma is technology. In the mainstream, one can only critique technology to help its development. A total, fundamental critique of technology is not permissible. The only permissible critique of technology could be a restricted critique about possible dangers, or one can critique some approaches to technology to indicate their shortcomings, as Larson does in this book. These are always done with a blatant display of good intentions. Like every good expert who wants to be accepted by the intellectual circle he belongs to and society in general, Larson has to signal the good intentions of his critique; he has to show that he doesn’t criticize the technology and techno-scientific development itself. He has to demonstrate that he only wants to contribute to AI’s development by pointing out the weaknesses of current approaches.
Karaçam
karapinusnigra@gmail.com
1 Erik J. Larson, The Myth of Artificial Intelligence: Why Computers Can’t Think the Way We Do, Belknap Press, 2021.
2 Sometimes, paradoxically, he underestimates the challenges artificial general intelligence would entail. We will see an example of this below.
3 David M. Buss, Evolutionary Psychology: The New Science of the Mind, Routledge, 2019
4 Jerome H. Barkow, Leda Cosmides, John Tooby; The Adapted Mind: Evolutionary Psychology and Generation of Culture, Oxford University Press, 1995, p. 182.
5 People perform differently in “a” and “b” not because the problem in “b” is more familiar than “a.” Cosmides and Tooby formulate the question at “a” in a more familiar form, but again in a way not about detecting cheaters. People perform significantly worse in this case also. See ibid.
6 See the Wikipedia article on the Dartmouth workshop: https://en.wikipedia.org/wiki/Dartmouth_workshop
7 For humans, it is easy to determine the truth value of the premises of this example. Because we have common-sense knowledge. And much of this common-sense knowledge is innate. Machines don’t have this common-sense knowledge, and to bestow them with this knowledge is daunting. We will return to this point below.
8 Until seeing a black swan, of course.
9 Of course, it depends on the total number of the population.
10 Larson quotes David Hume here.
11 For example, an exhaustive list of common sense knowledge should include this: pouring a liquid into a glass container with no cracks and only one opening will fill it up. It looks so simple and obvious to us, but machines have no clue about such simple facts.
12 Larson, ibid, Chapter 12.
13 Stuart Russell, Human Compatible: Artificial Intelligence and the Problem of Control, Viking, 2019
14 Ibid, p. 41.
15 There were literal human calculators up until the 60s who were responsible for making arithmetic calculations.
16 Theodore John Kaczynski, “Industrial Society and Its Future,” in Technological Slavery: Volume One, Fitch & Madison Publishers, 2019.
17 See the following article for the system’s possible reactions to climate change: The Possible Reactions of the Techno-Industrial System to Climate Change.
