Changing the world with AI

Artificial intelligence (AI) is a technical term referring to artifacts used to detect contexts or to effect actions in response to detected contexts. Our capacity to build such artifacts has been increasing, and with it the impact they have on our society. This article first documents the social and economic changes brought about by our use of AI, particularly but not exclusively focusing on the decade since the 2007 advent of smartphones, which contribute substantially to “big data” and therefore the efficacy of machine learning. It then projects from this political, economic, and personal challenges confronting humanity in

the near future, including policy recommendations. Overall,

AI is not as unusual a technology as expected, but this very

lack of expected form may have exposed us to a significantly increased urgency concerning familiar challenges. In particular, the identity and autonomy of both individuals and nations is challenged by the increased accessibility of knowledge.

01 Towards a New Enlightenment? A Transcendent Decade

1. Introduction

The past decade, and particularly the past few years, have been transformative for artificial intelligence (AI) not so much in terms of what we can do with this technology as what we are doing with it. Some place the advent of this era to 2007, with the introduction of smartphones. As I detail below, at its most essential, intelligence is just intelligence, whether artifact or ani- mal. It is a form of computation, and as such a transformation of information. The cornucopia of deeply personal information that resulted from the willful tethering of a huge portion of society to the Internet has allowed us to pass immense explicit and implicit knowledge from human culture via human brains into digital form. Here we can not only use it to operate with human-like competence, but also produce further knowledge and behavior by means of machine-based computation.

For decades—even prior to the inception of the term—AI has aroused both fear and ex- citement as humanity contemplates creating machines in our image. This expectation that intelligent artifacts should by necessity be human-like artifacts blinded most of us to the important fact that we have been achieving AI for some time. While the breakthroughs in surpassing human ability at human pursuits, such as chess (Hsu, 2002), Go (Silver et al., 2016), and translation (Wu et al., 2016), make headlines, AI has been a standard part of the indus- trial repertoire since at least the 1980s. Then production-rule or “expert” systems became a standard technology for checking circuit boards and detecting credit card fraud (Liao, 2005). Similarly, machine-learning (ML) strategies like genetic algorithms have long been used for intractable computational problems, such as scheduling and neural networks, not only to model and understand human learning, but also for basic industrial control and monitoring (Widrow et al., 1994). In the 1990s, probabilistic and Bayesian methods revolutionized ML and opened the door to some of the most pervasive AI technologies now available: search through massive troves of data (Bishop, 1995). This search capacity included the ability to do semantic analysis of raw text, astonishingly enabling Web users to find the documents they seek out of trillions of Web pages just by typing only a few words (Lowe, 2001; Bullinaria and Levy, 2007).

Today, AI breakthroughs in surpassing human ability in certain activities make headlines, but AI has been a standard part of the industrial repertoire since at least the 1980s

This capacity to use AI for discovery has been extended not only by the massive increase of digital data and available computing power, but also by innovations in AI and ML algorithms. We are now searching photographs, videos, and audio (Barrett et al., 2016; Wu et al., 2016). We can translate, transcribe, read lips, read emotions (including lying), forge signatures and other handwriting, and forge video (Hancock et al., 2007; Eyben et al., 2013; Assael et al., 2016; Haines et al., 2016; Reed et al., 2016; Vincent, 2016; Schuller et al., 2016; Sartori et al., 2016; Thies et al., 2016; Deng et al., 2017; Chung and Zisserman, 2017). Critically, we can forge real-time audio/video during live transmissions, allowing us to choose the words millions “witness,” particularly for celebrities such as politicians for whom there is already a great deal of data for composing accurate models (Thies et al., 2016; Suwajanakorn et al., 2017). At the time of this writing, there is increasing evidence that the outcomes of the 2016 US presidential election and UK referendum on EU membership were both altered by the

02 The Past Decade and Future of AI’s Impact on Society by Joanna J. Bryson

AI detection and targeting of “swing voters” via their public social-media use (Cadwalladr, 2017a,b; ICO, 2018), not to mention the AI-augmented tools used in cyberhacking (Brundage et al., 2018). AI is here now, available to and benefiting us all. But its consequences for our social order are not only not understood, but have until recently barely even yet been the subject of study (Bryson, 2015). Yet now too, with advances in robotics, AI is entering our physical spaces in the form of autonomous vehicles, weapons, drones, and domestic devices, including “smart speakers” (really, microphones) and even games consoles (Jia et al., 2016). We are be- coming surrounded by—even embedded in—pervasive automated perception, analysis, and increasingly action.

What have been and will be the impacts of pervasive synthetic intelligence? How can society regulate the way technology alters our lives? In this article, I begin by presenting a clean, clear set of definitions for the relevant terminology. I then review concerns and suggested remediations with respect to technology. Finally, I make expert though unproven recommendations concerning the value of individual human lives, as individual human capacities come increasingly under threat of redundancy to automation.

2. Definitions

The following definitions are not universally used, but derive from a well-established AI text Winston (1984), as well as from the study of biological intelligence (Barrows, 2000, attributed to Romanes, 1883). They are selected for clarity of communication at least local to this chapter, about the existing and potential impacts of intelligence, particularly in machines. Intelligence is the capacity to do the right thing at the right time, in a context where doing nothing (making no change in behavior) would be worse. Intelligence then requires:

— the capacity to perceive contexts for action; — the capacity to act;

— the capacity to associate contexts to actions.

By this definition, plants are intelligent (Trewavas, 2005). So is a thermostat (McCarthy, 1983; Touretzky, 1988). They can perceive and respond to context: for example, plants to the direc- tion of light, thermostats to temperature. We further discriminate a system as being cognitive if it is able to modify its intelligence, something plants and at least mechanical thermostats cannot do. Cognitive systems are able to learn new contexts, actions, and/or associations between these. This comes closer to the conventional definition of “intelligent.”