The myth of artificial intelligence: Why computers can't think the way we do. Erik J. Larson. Cambridge, MA: Harvard University Press, 2021. 320 pp. $29.95 (hardcover). (ISBN 9780674983519)
本书批判了当前AI发展必然实现通用人工智能的流行观点，指出其基于对人类智能的误解，并探讨了AI神话的社会影响，适合关注AI本质与局限的读者。
“Is this the start of the great AI jobs bloodbath?” (The Daily Mail, 19 May 2023) “AI is ‘clear and present danger to education’” (The Times, 20 May 2023) “Scientists use AI to discover new antibiotic to treat deadly superbug” (The Guardian, 25 May 2023) “AI creators fear the extinction of humanity” (The I, 31 May 2023) This small sample of recent headlines in UK newspapers captures some contemporary hopes and fears around artificial intelligence (AI). It is unusual for “technology” to hit the front pages, but AI is very much doing so at the time of writing. This intense public interest may reflect that, unlike many of the technologies that we use, AI is based on a deeply resonant idea, something one might say of mythic dimensions. Erik Larson's The myth of artificial intelligence: Why computers can't think the way we do prompts explorations of what it might mean to argue that AI is a myth or approach it as a myth. For Larson, the so-called myth of AI is the claim that current developments will inevitably achieve general AI (something akin to real human intelligence) as opposed to narrow AI (AI that has been trained for a specific task, such as to play chess). Larson is not suggesting that it is a myth that general AI is possible in itself, or even that it is the wrong aspiration, just that we should not believe it is inevitable or even likely on current development paths because it is based largely on an inaccurate, or outright wrong, idea of what comprises human intelligence. Human intelligence has three dimensions according to Larson. Much of computing is deductive—for instance, rule-based—which has taken us only so far with intelligent systems. The current phase of AI is based on inductive thinking, that is, on learning from data. Yet for Larson, such inductive systems can still only deliver narrow AI. This limitation is because they do not reproduce the other aspect of human thinking based on abductive logic that uses context to understand which is the most probable of possible explanations. Currently, it is not known how to enable a computer to perform this kind of so-called thinking. Larson briefly mentions the way that the earliest version of the web 2.0 myth of people power was overtaken by the idea of big data and the way that huge investments in data-driven AI has placed immense power in hands of a few powerful companies possessing the funds and data to develop it. However, he does not dwell on this issue of unregulated power; indeed, it seems not to be a central concern of the book. For Larson, the myth that inductive AI can generate general AI is dangerous because the investment in the project is doomed to fail and further diverts from exploring the true scientific task of uncovering the mystery of human intelligence. This seems to downplay the achievements of data-driven AI while glossing over many problems that feeding data reflecting social biases into data-driven systems appears to be creating with current AI. Across multiple domains, AI is already delivering on exciting possibilities such as enabling new forms of scientific discovery or saving lives through filtering health data at scale. We seem to be far from reaching the limits of what this form of AI can achieve, so this early focus on the limits of this type of AI seems premature. Additionally, AI is mired in wider debates about bias, privacy impacts, effects on environmental sustainability, and challenges to human autonomy. These concerns are central to current debates. By not more fully examining them establishes a gap in the book. Fundamentally, Larson assumes that achieving general AI should indeed be the ultimate goal. He does not consider the possibility that computation cannot ever replicate human thinking (partly because it remains unclear how it works) or that trying to reproduce human intelligence is an undesirable outcome because of the societal dangers that it creates. Larson employs the term “myth” in the sense of a widely-held but ultimately false belief. Myths, however, can also be seen as traditional stories that carry deep cultural (or even universal) significance. Following this direction of thought, one could turn to Mayor's (2018) book Gods and Robots, to further unearth the depth of our fascination and fear with artificial life. As a historian of ancient science, Mayor seeks to demonstrate the importance of the myth of technology-made life in Greek mythology, philosophy, and literature to show how and why people were already imagining it long before it was technically feasible. For Ancient Greek myth appears to abound in stories of life that is made rather than born. Hephaestus created Talos, a bronze giant robot, to protect Crete from strangers. It patrolled the island's shores three times a day, and would throw boulders at intruders or roast them by squeezing them against its red-hot body. Talos obviously had some level of intelligence and autonomy to carry through its tasks, so it is no mere machine. In some versions of the legend, it is credited with emotions and imagination. Hephaestus also designed Golden Maidens that bustled about as servants, understanding his needs. Other artificial life included the “living statues” created by the master craftsman, Daedalus. For Mayor, these myths are imagining genuine artificial life as products of technology—as life through craft—and are not based on magic or the will of a god. Mayor considers them as ancient thought experiments. The development of these ideas may reflect recognition of the power of metallurgy and the ability of craftsmen to copy natural forms in a striking way. But in philosophy and literature, such myths were also material to explore deeper moral concerns about human freedom and choice. For example, Mayor notes that such automata are usually deployed by powerful tyrants, echoing our own fears that robots will promote surveillance and control. Such stories, Mayor also illustrates, appear in the mythologies of other cultures. These ancient stories seem to be evidence of a persistent human interest in the imagination of artificial life as a vehicle to pose fundamental existential dilemmas about the human condition. Interestingly, in her closing chapter, Mayor emphasizes that these imaginings in ancient times produced emotions of “awe, dread and hope” (Mayor, 2018, p. 213). We seem to feel some of the same sentiments around current AI. The project to produce AI that mimics human intelligence, endorsed by Larson, may be something that science will aim to achieve, but it should not be ignored that it is emotionally provoking and will prompt profound questioning about what it means to be human. Each small moment of convenience—be it answering a question, turning on a light or playing a song—requires a vast planetary network, fueled by the extraction of non-renewable materials, labor and data. (Crawford and Joler, 2018) The myth of AI here is that it is simply based on clever algorithms. Abstracted from its social and historical context, the myth masks the material impacts and powerful vested interests that lie behind it. These insights are more fully articulated and extended in Crawford's 2021 book, the Atlas of AI. Like Larson, Crawford locates AI in a historical context, but a much longer, deeper one in the history of technologies and power. One of the major extensions to the previous work is that she analyses the nature of bias in AI, particularly the myth that AI can be neutral. For instance, she illustrates how a number of attempts to classify people in the context of AI have been damagingly reductive, such as trying to identify profound and personal aspects of identity from photographs. While these classifications purport to be neutral or objective, they are actually simplistic and normative and often have deeply sexist and racist assumptions built into them. The systems trained on these classifications then reproduce their assumptions and in turn, can cause untold harm. Crawford's point is that the underlying idea of classifying populations is inherently problematic and usually linked to forms of coloniality. It is more than the data being used to train algorithms is biased; rather, it is a much deeper problem of how attempting to classify people “objectively” can never succeed. Making decisions about how to train systems cannot therefore be within conceptions used in computer science. These problematic assumptions are a more fundamental problem with Larson's general AI project than not fully understanding human intelligence. The pattern of AI being used as a tool of surveillance and control over work is further illuminated by Crawford (2021). These uses thereby situate more power in the hands of employers and institutions. This pattern of impact of technologies like AI on work are explored further by Munn (2022) in Automation is a myth. Munn's claim is that the pervasive story of impending automation advanced in the media today, and that has been promised for about 100 years, is a myth. But it is a very powerful and dangerous myth because it presents the spread of technologies like AI as an apolitical story of technical progress masking its links to capitalism, coloniality, and patriarchy. The myth of automation, for Munn, is made up of three myths. First is the “myth of automated autonomy” (Munn, 2022, p. 9). This is the claim that any work can be fully automated. Munn argues that this is not possible. Whether automated autonomy is a utopian dream (of the freedom from work) or a dystopian nightmare (of loss of human control and a role) complete automation is always a fantasy. In reality, because work is often very complex, there are always gaps in what can be automated that have to be filled by human labor. Such a gap in the potential to automate is probably closer to what we feel than that there is a gap in our understanding of human intelligence as Larson claims. Computers can never really do everything humans do in multiple ways. These gaps, moreover, are generally filled by work that is often menial or unpleasant, such as the task of combing through content that is flagged for moderation because of toxic topics. It is usually precarious work and rendered invisible by being crowd-sourced and typically located in the global south. Munn does chart how the discourse about autonomy for AI is shifting thinking towards recognizing the need to involve humans and for it to be more “human centered” (Munn, 2022, p. 21). But there are tensions because automation was always about efficiency and productivity, which do not reflect the more holistic or rounded needs from work of humans. The second myth is the “myth of autonomy everywhere” (Munn, 2022, p. 47). This myth claims that the wave of automation is happening inevitably and everywhere. Yet, the reality of the patchiness of automation is ignored, and moreover, the social and material contexts making automation possible as well as its actual impact in specific contexts are denied. The third myth is the “myth of automating everyone” (Munn, 2022, p. 81). This myth points to the unequal effects of automation on different groups. Because they already have fewer resources, some black workers, for example, may have less ability to ride the shockwave of automation. After all, automation happens in already racialized workplaces. Similarly, automation has a high impact on already precarious work in which women are over-represented. Automation effectively reproduces existing inequalities in labor markets rather than enhancing all work. Critically, the automation myth is dangerous because it masks the reproduction of capitalism, colonialism, and patriarchy behind a story of progress. Thus, it is not just a false belief but a powerful ideology that breaks down resistance to human values. This argument is an incisive and partly convincing to help cut through the hype around AI. Perhaps it is also an oft repeated myth that automation, including AI, can only impoverish work. This impoverishment may indeed be a tendency and it is one that we need to be constantly reminded of because the myth of automation has an ideological role in masking exploitation. At the same time, this account does appear somewhat one dimensional. The effects of automation, including AI, are likely to be rather drawn out and complex, with wholly new jobs and types of jobs being created as others decline, and some work groups gaining job enrichment and freedom from mundane tasks, others losing control, status or identity (Willcocks, 2020). The myths of AI and of automation are versions of a rather familiar myth of technological determinism and solutionism. This recurring myth is that technological innovation in itself drives social change (ignoring the social construction of technology), that it is progressive and inevitable, and promotes the idea that by itself technologies can solve complex social problems. Recognizing and challenging examples of this myth is an important dimension for everyone working in the information sciences and indeed beyond. Critical information literacy must rest on an appreciation of this myth. Raising understanding of the infrastructures that make the apparent “magic” of AI possible is central to what information professionals need to promote. Tools like ChatGPT do not simply pop up as free resources that one might choose to use or not; instead, it is important to recognize that they are created by humans, often with profit in mind. A full understanding of these facts could be described as algorithmic literacy. Algorithmic literacy in this context may have multiple components. It implies explaining how to use Chat GPT effectively, recognizing the need to formulate queries in the best form, and to recognize its information limits, such as that it is inaccurate and does not give its sources, was fed with data only up to September 2021. It is not irrelevant to acknowledge Larson's point that current AI reproduces only parts of what makes up human intelligence. Algorithmic literacy also implies prompting users to reflect on its impact on their experience: is it making learning processes too easy? How does it make them feel? This reflection would be to acknowledge the questions for the human condition that Moon traces back to ancient myths about artificial life. And also, fundamentally, drawing on the critiques of AI myths from Crawford and Joler (2018) and Munn (2022), we should be prompting users to consider whether they want to use AI at all when they realize the environmental impact (Ludvigsen, 2022); the ghost labor of Kenyan workers used to detoxify content (Perrigo, 2023); the way it was trained on copyrighted material without permission; the biases created within the system by the data it was fed with; the way our current use of it is training it further (our use of systems is data labor; Li et al., 2023); the privacy risks of sharing our data with it. Ultimately, these prompts point to the unregulated power of so-called “big tech” companies to control our world. Critical algorithmic and AI literacy must involve understanding and responding to such power. The myth of artificial intelligence emphasizes how the story of AI has become a central myth of the contemporary world. AI is fascinating because it is a complex, ambivalent story rooted in the existential quest of human self-understanding, even self-conceptualization. This book supplies a solid starting point for accessible and incisive insights into this myth, particularly from the perspective of computer science. Ultimately, it offers an important commentary on the myth of AI that is based on interrogating the nature of human intelligence.