Atlas of AI – Book reviewAtlas of AI, , Yale University Press (2021)
Introduction
The introduction of the book starts by questioning the analogy between human and artificial minds. The claim is that this analogy neglects the embodiment, relations, and environment in which humans function. AI is usually understood as disembodied intelligence. Instead, the book wants to show that AI is not artificial and disembodied, but rather very embodied and material, made from natural resources (Chapter 1), human labor (Chapter 2), and people's data (Chapter 3). It also includes decisions about the classification framing (Chapter 4) that can impact on people's identity, it aims to interpret human emotional states (Chapter 5), and is used as a tool for state power (Chapter 6). AI is also not autonomous, since it needs extensive data or rules from humans. In my view, this framing is interesting and worth being discussed and presented. The book structure follows this scheme, which provides the backbone for the narrative.
One thing I noticed in reading the introduction is that the author seems to equate AI with machine learning. It is mentioned that the term AI is mostly used in marketing rather than the computer science work, while machine learning is more commonly used in the technical literature. These statements seem to imply that machine learning is all there is in AI as a science and technology. This is not true. Entire scientific communities study and work areas of AI that are not based on machine learning techniques, such as search, optimization, planning, scheduling, knowledge representation, and others. Although some scientific conferences are focused on machine learning (such as NeurIPS), it is enough to go and check the list of topics of major AI conferences such as AAAI and IJCAI, or top journals such as the AI Journal or the Journal of AI Research, to see that machine learning is just a part of what is being studied and advanced by AI researchers. So, I would claim that AI researchers do use the term Artificial Intelligence in the scientific and technical literature.
Section snippets
Chapter 1 - Earth
Chapter 1 considers the AI's need for raw material, such as lithium, that is necessary to build the rechargeable batteries of AI-powered devices. The author travels to places where the “computational extraction”, that is the extraction of raw material to support computational techniques and devices, take place, such as Silver Peak in Nevada, where there is a massive underground lake of lithium, and describes the danger of environmental damage, miners' illnesses, and displaced communities
Chapter 2 - Labor
Chapter 2 discusses the human-technology relationship in the workplace. It starts with an analysis of the human-robotic hybrid work in warehouses and assembly lines, and positions the current practices within the history of standardization, simplification, optimization, and scaling that started much earlier than AI was in the picture, in environments like the car manufacturing and the meat market. It then describes the use of technology for workers' surveillance and tracking, and it also
Chapter 3 - Data
Chapter 3 describes the implications of the AI's need for large amounts of data. Over the years, data about people (voices, face pictures, text, etc) became the technical baseline to train and test AI techniques, often losing the connection to the people and the situations in which this data was collected. Now regulations require some form of consent, but in the early days of AI data was obtained with no attention to privacy. Later, with the advent of Internet, data of all kinds became readily
Chapter 4 - Classification
Data labeling can be done only after the possible classes (that is, the labels) have been defined. So a crucial decision in building AI models is to decide on such classes. Chapter 4 focuses on the assumptions and the implications of this decision. Any specific classification reflects a vision of the world, and assumes a form of universality, for example regarding genders and other people's features.
The chapter also links the bias issue to classification: the absence of some labels may be the
Chapter 5 - Affect
Chapter 5 focuses on a specific AI application: the recognition of human emotions from an analysis of the face of a person. This task is based on the assumption that there is a set of emotions that is universal and innate, and that they can be detected automatically by a machine. The chapter traces back this assumption to extensive pre-AI work on physiognomy and affect recognition, that received huge attention and funding but also controversial feedback on the validity of the assumptions
Chapter 6 - State
Chapter 6 claims that many AI practices have been derived from military priorities and methodologies. This is mostly due to the fact that military funding agencies supported AI research since its start, from computer vision to automatic translation to automated vehicles. So, a similar classification and surveillance framing has been applied also to civilian settings such as banks, airports, and streets. This allowed to move from law enforcement, which comes after a criminal incident happened,
General comments
I conclude my assessment with some general comments about the book.
In the various chapters, I really appreciated the innovative angle of exposing several often neglected actors and assets that greatly influence the methods and results of the AI lifecycle: the extraction of natural resources; the essential role of low-paid workers; the relationship between data, people, culture and context; the role of classification in defining a vision of the world; the controversies of affect recognition; and
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (10)
- et al.
Human + Machine: Reimagining Work in the Age of AI
(2018) Web site with description of the initiative and sessions