- an artificial device that can sense its environment and purposefully act on or in that environment;
- an embodied artificial intelligence; or
- a machine that can autonomously carry out useful work (2012: 8)
The most important aspect of Winfield’s definitions is that, despite differing morphologies, all robots have bodies. AI, however, is software and, therefore, need not be embodied, though it requires computing hardware to run on. Advanced robots employ AI for functions including perception, planning actions, and learning, but a robot body does not necessarily entail AI, nor does an AI system necessarily entail a robot body.
To distinguish actually-existing AI from its speculative future incarnations, it is helpful to employ the following three categories: narrow AI, artificial general intelligence (AGI), and artificial superintelligence (ASI). Actually-existing AI is narrow: ‘the vast majority of current AI approaches … are primarily designed to address narrow tasks’ (Johnson et al. 2016: 4246). Most AI research, all commercial applications of AI, and the AI that consumers use daily, are such task-based tools. They are functionally more akin to microscopes than the anthropomorphic and politically active droid $\mathrm { L } 3 { - } 3 7$ in Solo. These systems have none or very little ability to do anything beyond their particular domain of functionality. An AI system that recognizes faces in photographs is not going to be able to process recordings of speech, play Go, or compose emails, and it is definitely not going to be able to speak Farsi. We will discuss dozens of existing narrow AI systems over the course of this book.
On the basis of generality, narrow AI is contrasted with artificial general intelligence (AGI), which refers to AI with ‘the capacity for efficient cross-domain optimization’ or ‘the ability to transfer learning from one domain to other domains’ (Muehlhauser 2013). AGI refers to an AI with the capacity to engage and behave intelligently in a wide variety of contexts and to apply knowledge learned in one context to novel situations, meaning it would be ‘capable of reasoning across many intellectual domains’ (Baum 2018a: 3). As of 2019, AGI remains a speculative technology, although serious research is now being conducted on it in both public and private institutions. We discuss AGI in Chapter 3.
Artificial superintelligence (ASI) is yet more speculative. While an ASI ‘is likely to have general intelligence’ (Baum 2018a: 3), it specifically refers to an AI ‘that greatly outperform[s] the best current human minds across many very general cognitive domains’ (Bostrom 2014: 63). ASI is a science fiction staple, but serious discussion of it also occurs in academic circles where it is often seen as swiftly following the creation of AGI (see e.g. Bostrom 2014; Torres 2018; Baum 2018a; 2018b). Most commonly, the scenario imagined is that an AGI gains the ability to self-modify and evolves into a god-like ASI with unpredictable powers. The consequences of such an event are impossible to predict with certainty, but the mere possibility of it occurring compels thinkers and institutions – including Nick Bostrom and the Future of Humanity Institute at Oxford, Seth D. Baum and the Global Catastrophic Risk Institute, as well as Eliezer Yudkowsky and the Machine Intelligence Research Institute (MIRI) – to argue that we must seriously research the possibility now.
10
inhuman power