13 Artificial life, artificial agents, virtual realities: technologies of autonomous agency

Colin Allen

13.1 Introduction

This chapter discusses ethical issues arising in connection with artificial life (non carbon-based), artificial agents (physically embodied robots and software agents or ‘bots’) and virtual reality. To different degrees, each of these technologies has a tinge of science fiction, but they are nevertheless all far enough advanced to raise current and pressing ethical issues. Each of them can be considered in the context of general ethical issues raised by philosophers of technology. These include worries about unintended harms, ranging from incremental erosion of human freedom and dignity as a result of over-dependence on machines, to the danger of a huge catastrophe by technological failure. They also include general worries about the difficulty of assessing risks and rewards of technological development, and about the potential of technology to concentrate political and economic power in the hands of those who have privileged access or technical knowledge, creating a de facto technocracy or fostering economic injustice.
The specific ethical issues on which this chapter is focused arise from the potential for autonomous agency that is afforded by computational agents and environments. The chapter begins with a brief survey of the technologies involved, subdividing both artificial agents and virtual reality to produce a list of five different kinds of technology. Next comes a discussion of the notion of autonomy as it applies to people and to information-processing artefacts. The ethical issues arising in the context of each of the five technologies are then described, followed by some concluding remarks.
13.1.1 Introduction to the technologies
We begin by clarifying the kinds of technology to be discussed under the headings of artificial life, artificial agents and virtual reality.
Artificial Agents may be embodied or virtual. Embodied artificial agents are designed to operate autonomously in physical environments ranging from the nanoscale to the macroscopic domains of interstellar space. Autonomous robots are the most obvious examples of embodied artificial agents, but cyborgs and other forms of technologically enhanced human beings may also qualify as embodied artificial agents. Virtual artificial agents, sometimes known as software ‘bots’, are designed to operate inside computationally generated environments. These entities may be embedded in standalone applications, such as the artificially intelligent agents that have long been included in computer games. Or they may operate in networked environments, including the Internet as a whole and specific portions of the Internet that make up the so-called ‘virtual worlds’ for multiuser games (e.g., World of Warcraft) and multiuser social environments (e.g., Second Life). Like ‘life’, the proper definition of the term ‘agency’ is a matter of controversy, and the concept of artificial agent stretches ordinary usage. In this chapter, ‘agent’ will be used to refer to any system that can use information adaptively to achieve some goals. The ethics of both embodied and virtual artificial agents are discussed below.
Virtual Realities (VRs) are computer-generated environments designed to exploit the sensory systems of human beings so as to produce a sense of presence in those environments. Machine-centred virtual realities are computational environments that are designed to make things easy for computers. For instance, the Internet Protocol (IP) address space is a machine-centred virtual reality which is structured to make it easy for computers to pass messages across the Internet. Machine-centred virtual realities make little or no attempt to replicate the look or feel of physical environments for their human users. Nevertheless, early users of computer networks such as the ARPAnet (predecessor to the Internet) were often thrilled by the sense of remote presence provided by a virtual terminal connection to another computer that could be located hundreds or thousands of kilometres away. Anthropomorphic virtual realities are computational environments which attempt to make it easy (or engaging) for humans to interact with machines and other humans by replicating the kinds of experiences humans have of their physical (and social) environments. They vary in the degree to which the full range of human experience is replicated, and in the extent to which the faithful simulation of a physical environment is the end in itself, as for example in flight simulators, or an instrument for effective presentation of other kinds of information, as imagined by the original cyberpunk author William Gibson in his début novel Neuromancer.
13.1.2 Autonomous agency
Each of the five technologies just described contains the potential for autonomous agency either of the technological artefacts themselves or of humans operating within the virtual environments provided by the technology. There are divergent views among philosophers about how to understand the notion of autonomous agency.
The Kantian tradition within moral philosophy holds that genuinely autonomous agency is possible only for fully rational persons, and this is often taken to be restricted, at least on this planet, to mature, adult human beings. The Kantian conception is tied up with ideas about the importance of conscious reflection on reasons for beliefs and actions, and on the capacity to be guided by consciously adopted normative principles in the face of temptation or tendency to act differently. Such a conception is also entangled with the thorny notion of free will.
In contrast to the Kantian conception, there are deflationary accounts of autonomous agency which entail only that a reasonable selection is made among different behavioural options that take some account of the environmental challenges and contingencies but is not entirely determined by the environmental factors. Thus, for example, a rat faced with two food odours leading in opposite directions approaches one of them without that selection being strictly determined by the external cues. To say that the selection is reasonable is not to say that the agent is capable of self-consciously thinking about the reasons for choosing one way rather than another. Rather, it is to attribute the selection to the operation of a mechanism that has been deliberately designed or has evolved to make such decisions so as to satisfy some goals.
In this chapter, the phrase ‘autonomous agent’ is to be understood as something close to the latter ‘deflationary’ notion. If the Kantian insists, he or she can have the words. The important point is that our five emerging technologies represent something relatively new in the history of technology, namely artefacts which are not merely static, like bridges or buildings, and not mere vehicles for direct human agency, such as automobiles or can openers. These new technologies possess within them mechanisms that allow them to operate flexibly and without direct human supervision in a variety of conditions. On some views (e.g., Dennett 1987, Floridi and Sanders 2004) the agency of such artefacts is a matter of a stance taken by observers who are not mindful of the mechanistic principles by which they operate, and who may even be incapable of understanding them in such terms. On other views, the agency may be inherent to the structure of the mechanisms themselves (e.g., Wallach and Allen 2008). In other words, if an agent is a certain kind of complex mechanism, then it is not just a ‘stance’ but a proper description when agency is attributed to things having the right kind of complexity. For the purposes of this chapter it is not necessary to take a stand on this issue.
ALife and the technologies of embodied and virtual artificial agents obviously involve autonomous agency in the just-described sense. It may be less obvious how both forms of VR also involve this concept. Arguably, however, these technologies provide the environments where artificial autonomous agents have reached their most advanced forms. In machine-centred VR, millions of programs running on networked computers take ethically significant decisions without human oversight every millisecond of the day. Computers, for example, govern almost every credit card transaction on the planet, deciding whether to accept or decline charges without direct human supervision. Stock market trades are carried out by software using sophisticated analytical tools to determine the timing of sales and purchases. These programs are themselves artificial agents, but they are operating in a designed environment that allows them to make their decisions. A software trading system cannot operate unless the market provides pricing information directly to the software. Access to the data necessary for such agents to operate is part of what is afforded by the machine-centred VR. Thus, the ethics of constructing such environments cannot be separated from questions about the autonomy of the agents operating within it.
For anthropomorphic VR, the questions also include what kinds of entities should populate these environments. For instance, multi-user games in virtual environments gain some of their interest from the interactivity (and unpredictability) of the behaviour of other people operating within the same VR space. But people using VR also want artificial elements of these environments to behave in autonomous, interesting ways. Computer-generated characters inhabiting the space need to be as engaging as human-operated avatars. Furthermore, in anthropomorphic VR the autonomous agency afforded to people navigating the VR space can be technologically enhanced but it can also be artificially limited. The opening up of new possibilities for action and the closing down of other possibilities may have important consequences for the benefits and harms to people, whether directly internal to the virtual realms or indirectly through the effects on how people transfer what they have learned in the VR context to the real world. Thus, autonomous agency is also an important issue within anthropomorphic VR environments.

13.2 Ethics of virtual autonomy

We now turn to the particular ways in which issues of autonomy and agency play out in our five selected technologies.
13.2.1 ALife
During the 1960s, the exponential growth of speed and storage capacity in integrated circuits was recognized, giving us what is now known as Moore's Law (Moore 1965). The 1970s saw the introduction of ‘genetic algorithms’ (Holland 1975). Together, these developments suggested that the enormous power of evolution would soon be harnessed for computational purposes. In the 1980s, early proponents of ALife imagined that success was just around the corner. However, the task of engineering virtual computational environments in which ALife can flourish has proven much more difficult than its early advocates assumed. Partly this is due to the fact that ALife researchers started out with what we now know in hindsight to be overly simplistic ideas about the relationship between genotypes and phenotypes. For example, if an ALife researcher wanted to evolve artificial creatures whose artificial neural networks were capable of some task, they typically assumed it was appropriate to have one ‘gene’ for each connection in the network to determine the strength (‘weight’) of that connection. The fact that there aren't enough genes to code for every connection in the brain had been understood from at least the 1950s, even by computer scientists. But the one gene-one weight approach to ALife was nonetheless adopted as a reasonable simplifying assumption to allow modelling to begin.
This stance was partly based on the old dogma from biology that genes code for proteins in a one-to-one fashion. However, the genome mapping projects that were started in the 1990s and culminated with the publication of the first complete human genome in 2003 taught biologists that the development of organisms is far more complicated than this dogma suggests. The construction of cells and living organisms requires many resources not just from outside the DNA but from outside the cell and outside the organism itself. There is much machinery between transcription of a piece of DNA and the appearance of a specific protein in a cell. Furthermore, the so-called regulatory sequences of DNA need not code for proteins at all, and even those pieces of DNA that are used for the production of proteins do not directly specify the proteins expressed, but are subject to processes such as post-transcription editing by RNA. The result is that there is no one–one correspondence between genes and products, let alone between genes and phenotypic traits. The interactions are fearsomely complex, and it is part of the reason why humans and flatworms can be so different anatomically and behaviourally despite the fact that they have roughly the same number of genes, with many of those genes even being shared. So far this discussion has focused on the evolution of virtual organisms for computationally defined environments, but progress is also being made on evolvable hardware. Much of this work reinforces the lessons about the complexity of physical interactions and the surprising ways in which physical components can interact, even sometimes confounding the expectations of engineers (Thompson 1996, Lohn and Hornsby 2006).
Despite the utility of evolutionary algorithms for finding efficient solutions to difficult hardware and software design problems, these approaches yield systems that can be hard to analyse and understand. This lack of transparency, due to the complexity of the interactions involved, may be a source of worry about effective risk analysis for those who wish to deploy artificially evolved hardware or software. Nevertheless, current ALife technology is far from sustaining adaptive, self-regulatory and self-reproducing lineages of autonomous agents. The technological limitations noted here for both virtual and physically embodied ALife should, therefore, temper concerns about ALife somehow escaping the clutches of scientists, and running amok. Worries about the perils and ethical consequences of life in silico are destined to remain science fiction into the foreseeable future. Many commentators believe, however, that a more plausible path towards artificial life forms consists in hybridization of biological life forms with machines: cyborgs.
13.2.2 Embodied artificial agents
Embodied artificial agents are designed for operation in physical environments. Cyborgs are one main genus of embodied artificial agents, the other main genus being robots, which are completely inorganic in nature.
Cyborg technology covers a very wide spectrum of hybrid systems. It may be difficult to give a sharp definition that captures just what interests us about cyborgs in this context. A common dictionary definition, for example, identifies a cyborg as a ‘person whose abilities are extended beyond normal human limitations by mechanical elements built into the body’. In this case, anyone with dental implants might count as a cyborg, if those artificial teeth are considerably stronger or more cavity resistant than their natural counterparts. Prosthetic devices of all kinds are capable of enhancing human abilities, and the 2008 case of whether double amputee Oscar Pistorius could compete in the Olympic Games with his ‘Cheetah’ blade legs highlighted the extent to which artificial limbs might provide an ‘unfair’ advantage because their material characteristics provide more efficient energy storage and return than natural legs. In the end, Pistorius failed to beat the qualifying time for entry to the 400 m event, although his time certainly put him well beyond average human running speed. Perhaps, however, the dictionary definition excludes Oscar Pistorius from the category of cyborgs on the grounds that his Cheetah blades were strapped on, not ‘built into’ the body. Still, it is not too much of stretch to imagine such prostheses soon being surgically implanted via an artificial knee or hip joint. Other borderline cases of cyborg technology include robotic exoskeletons that are being developed for military use (e.g., by the Utah-based robotics company Sarcos) and hybrid robots that use actual biological neurons grafted onto a multi-electrode array (MEA) to control simulated and embodied systems (Potter, Wagenaar and DeMarse 2006; as widely reported by news media, Potter trained 25,000 rat neurons on an MEA to control a flight simulator).
Setting definitional matters aside, the kinds of cyborgs that have the most interest for the purposes of this chapter are those which offer cognitive enhancements to humans through implanted computational devices. For many citizens of technologically advanced societies, cell phones, PDAs and pocket Wi-Fi devices have become indispensable external aids to our fallible memories and limited knowledge bases. Prototypes for cell phone implants have already been developed, but while these are physically embedded within the human body, they so far involve no direct links to the human nervous system. External readings of brain waves via EEG technology are being tested for the control of speech synthesizers, wheelchairs and avatars in virtual environments. The prospect of even tighter integration of information technology with human minds via direct implantation of devices to the human nervous system seems to offer enormous possibilities for cognitive enhancement. Neural-cognitive implants are being investigated mostly with medical applications in mind, but some for the sheer thrill of being at a technological frontier. Thus, brain–computer interfaces are being investigated for speech synthesis and brain-implanted electrodes have been used to control prosthetic limbs in monkeys and humans (Lebedev and Nicolelis 2006). At the more speculative and futuristic end of the spectrum, Kevin Warwick, professor of cybernetics at Reading University, has experimented with a variety of implants (Warwick 2004).
These technological developments raise some very general ethical issues that accompany the use of any technology. For example, the adoption of advanced technologies frequently widens the gap between haves and have-nots. Much has been made of the ‘digital divide’ between those who have access to computers and the Internet, and those who do not (see, e.g., Servon 2002). The ‘cyborg divide’ could be just as significant if these enhancements provide major advantages to those who adopt them. Specific enhancements also raise particular ethical issues. For instance, there are already hearing enhancers that can increase hearing sensitivity twentyfold, which means that you may no longer safely assume that a distant person cannot hear your private conversation. Human-cyborg technologies afford greatly enhanced agency in the real world through an increase in cognitive and physical capacities. If a cyborg human can do the physical work of ten men or the intellectual work of a team of experts, this has potentially profound consequences for increasing the freedom and autonomy of those who are enhanced, and decreasing the freedom and autonomy of those who are not.
While cyborgs are still mostly confined to the research laboratory, robots have long been a presence in the human environment. Industrial robots are very widely used, but these special-purpose machines are typically bolted to the factory floor where they do one job, and one job only. In non-industrial applications, however, autonomous or semi-autonomous robots that are free to roam and may have multiple capabilities are becoming mass-market consumer items as well as occupying more specialized niches. One morally significant area where the widespread use of robots is currently envisaged is in the burgeoning field of elder-care (Anderson and Anderson 2008). Technologically advanced countries are facing a large bulge in their elderly populations. Japanese society is particularly faced with the challenge of insufficient health care workers to take care of the aging population, and the introduction of robots in this context has been made an official government policy. So-called ‘carebots’ may be responsible for dispensing medicines, making sure they are taken, encouraging exercise and providing basic companionship (Floridi 2008a).
The field of robotics is also a major focus for military planners, for example the US Army's Future Combat Systems program. At the same time there is increasing penetration of semi-autonomous robots into homes, whether as programmable robotic toys (e.g., Sony's now-discontinued AIBO robotic dog, or current offerings such as Ugobe's Pleo robotic baby dinosaur) or as household appliances (e.g., iRobot's Roomba robotic vacuum cleaner). Roboticist and iRobot founder Rodney Brooks argues that there is presently an inverse relationship between price and autonomy (Brooks 2007). The expense of military robots means that there is an incentive to place them under human supervision to protect the investment. But home robots must be cheap to be commercially successful, and their owners do not want to have to supervise them continually. Neither does the average home contain the infrastructure that would be required for comprehensive monitoring and control. Thus the expensive robots that have been deployed for space exploration or military applications have thus far been almost entirely tele-operated, while the cheap robots intended for the mass market are relatively uncontrolled. Although Brooks may, for the time being, be right about the relationship about being cheap and out of control, the logic of military deployment nevertheless drives towards giving robots greater autonomy (Wallach and Allen 2008). For instance, the ‘unmanned’ drones that have been used extensively by the US military for missions in Afghanistan, Pakistan and Iraq, are in fact tele-operated from a base in Nevada by crews of four or more highly trained individuals. Increasing the autonomy of the drones would allow more of them to be flown with the same number of operators. The advantages for military superiority of having greater numbers of fighting units provides an incentive for increasing machine autonomy.
To work effectively with humans, robots need to engage human interest and maintain it. The field of human–robot interaction (HRI) investigates visual, linguistic and other cues that support engaging interactions. Among the characteristics that are systematically investigated by HRI scientists are the emotionally significant facial expressions and non-verbal properties of speech, such as rhythm or prosody, that play a significant role in human social interaction. These ‘minimal cognition’ cues are a focus of behaviour-based robotics (Breazeal 2002), and have proven surprisingly effective in giving people the sense that they are dealing with intelligent agents. However, some critics fear that the addition of such traits to robots is fundamentally deceptive, relying on the strong tendency of humans to anthropomorphize objects by projecting human-like characteristics onto things that don't have them. This may prove especially problematic in the context of elder-care, with people who are relatively starved of human companionship (Turkle 2005, Turkle et al. 2006).
All of these applications of robotics – whether home, military or healthcare – involve their own ethical issues. This is especially clear for military applications where there is considerable concern about ensuring that combat robots follow acceptable rules of war, as encoded by the Geneva conventions for instance (Arkin 2007), and for healthcare applications where the issue of patients’ rights (for instance, to refuse medication) are of concern (Anderson and Anderson 2007). Arkin (2004) has also suggested that the likely military, service and sexual applications of robots may serve to revive callous attitudes towards life and liberty that characterized earlier stages of human history, thus undermining moral progress in domains such as the formal abolition of slavery. Arkin neatly captures these concerns with his title phrase, ‘Bombs, Bonding, and Bondage’.
More generically, the presence of autonomous robots in real-world environments, where it may not always be possible to constrain their actions to well-defined contexts (robots do roam), raises questions about whether these machines will need to have on-board ethical decision-making capacities (Moor 2006, Turilli 2007, Wallach and Allen 2008). With the current state of robotics and artificial intelligence, the suggestion that artificial moral agents are possible, let alone necessary, may seem far-fetched. However, assessments of the autonomy and morality of artificial agents may hinge less on ‘deep’ metaphysical facts about moral agency, and more on the fact that people will adopt different stances towards artificial agency based on whether they understand (or care to understand) the underlying mechanisms (Floridi and Sanders 2004, Grodzinsky et al. 2008). The need is illustrated by the fact that artificial agents operating in virtual environments are already autonomously making decisions with potentially significant ethical consequences, even though they are blind to those consequences. The computer that denies your credit card purchase makes no prediction about whether this will ruin your day or your life, and gathers no information that might help it make such a determination.
Robotics thus forces us to think hard about whether and how autonomous agency might be located or replicated in the computer algorithms that control our most advanced technologies. The challenge of building machines that can be regarded as artificial moral agents raises not only questions about the ethics of doing so, but also raises questions about the nature of ethics itself, such as whether ethical rules and principles are the kinds of things that can effectively guide behaviour in real-time decision-making (Wallach and Allen 2008).
13.2.3 Virtual artificial agents
The presence of bots in contexts where people are seeking social interaction has been a persistent concern to ethicists and technologists. These concerns predate the rich, virtual, multi-user environments exemplified by Second Life. They were raised in the context of the early text-based ‘MUDs’ (multi-user dungeons) and ‘MOOs’ (MUDS object oriented) that were constructed on early computer networks using only textual communication, and these concerns were raised even earlier than that in the context of artificial intelligence. A frequent concern is that people are easily tricked into forming social attachments to entities that are incapable of reciprocating them. This worry arose in the early days of artificial intelligence research in connection with ELIZA (Weizenbaum 1965), a program that simulated (or, perhaps, parodied) the question-asking technique of Rogerian psychotherapy, and that is credited as being the first ‘chatterbot’ or ‘chatbot’ – a piece of software which attempts to sustain a conversation with a human interlocutor. Carl Sagan enthusiastically suggested that in the future there would be computer psychotherapists available in street corner booths everywhere. Weizenbaum emphatically rejected this vision of the future and renounced his work on ELIZA, maintaining that computers would always lack the important human qualities of compassion and wisdom (Weizenbaum 1976).
The fact that people tend to overestimate the intelligence and complexity of computer programs is now known to AI researchers as the ‘ELIZA effect’. It is a manifestation of the previously mentioned human tendency to anthropomorphize things. However, this tendency may be exacerbated by the much richer graphics-based virtual environments that have become common in the forty years since ELIZA's inception. Furthermore, in mixed virtual environments where some avatars are human-operated and others operated by autonomous software, it may be especially difficult to distinguish which are which. This is partly because the range of human actions in a virtual world is itself limited by the constraints on expression and action that are imposed by the rules of the virtual environment, a topic we shall return to in the following sections.
Although autonomous agents in virtual worlds are sometimes difficult for other users to distinguish from real human beings operating in those worlds, they are just as frequently all too easy to identify. And while there are many systems operating in virtual environments without direct human oversight, all of them are ‘ethically blind’ (Wallach and Allen 2008) in the sense that they lack specific information that would be relevant to ethical decision-making, they have no means of assessing the ethically relevant effects of their decisions, and they lack other capacities such as empathy that are important to human morality.
Work is also being done on virtual equivalents of cyborgs – on-screen avatars that combine tele-operation by humans with software enhancement. For instance, an on-screen avatar in a virtual reality context may derive its behaviour and appearance from the actions and facial expressions of a person connected to bodily motion sensors and cameras capable of providing enough information to render a three-dimensional model of the actor in the virtual setting. But such a tele-operated avatar can also be filtered through an intermediate layer of software and enhanced in specific ways, for instance by blending the appearance of the remote operator with the face of a famous person, or by enhancing and sustaining facial expressions that are known to influence the responses of other people to the avatar. The potential for ‘persuasion’ by non-verbal means is only just beginning to be investigated, but the potential for ethical abuse in such situations is already clear (Bailenson et al. 2008).
13.2.4 Machine-centred VR
We have already touched upon machine-centred VR in discussing the bots that operate there. While the autonomy of such agents is relatively limited at present, nevertheless the construction of machine-centred VR gives artificial agents an advantage compared to humans attempting to operate in the same environments. The simple example of online auctions illustrates this. Because the format in which information is presented favours machines, they are much more capable than humans of precisely timing a final bid to win an auction at the last second.
Of course, it is not all bad, and people reap benefits from having their software systems operate in environments that have been shaped so as to make it easy for the software rather than the people. Nevertheless, the design decisions involved in shaping such environments often have unintended effects, just as the shaping of our physical environments to accommodate motorized vehicles has had unintended effects on our ability to walk or use other modes of transportation, which in turn has both positive and negative consequences for general health and well-being. Similarly, the original use of the very limited set of characters belonging to the ASCII code made it easy for machines to process documents originally written in English, and for computers (and their English-speaking users) to communicate over networks. But the limitations of ASCII made it difficult for speakers of languages other than English to represent their documents in machine-processable formats and to communicate via email, instant messaging, and other information and communication technologies (ICTs). The adoption of Unicode, with its much richer character sets, has gone a long way towards rectifying these problems, of course, but not without the expense entailed in updating the software and revising archived materials so as to provide forward compatibility. Furthermore, ICTs embody other culture-specific assumptions about communication, such as the relative importance of words or text over non-verbal aspects of communication – for example, the facial and bodily gestures which in some cultures are important for showing proper respect to others. Thus it is important to recognize that technologies which appear to be value-neutral from one cultural perspective may in fact raise ethical issues in a different cultural context (Ess 2009).
The ethical issues raised by machine-centred VR environments are thus not just about the artificial agents operating within them, but concern the limitations placed by these environments upon people who attempt to operate within them. Current metaphors for information retrieval are frequently library-based metaphors, such as browsing and searching. Science fiction writers have long dreamed of technologies that would transform digitally encoded machine-centred VRs into information environments, epitomized by the brain–computer interface that the primary character of William Gibson's Neuromancer uses to ‘enter’, perceive and move around in cyberspace as if it were physical space. The field of information visualization is beginning to provide tools that support rich visual presentation of abstract information, and has been especially successful in providing visual representations of information networks. The much richer anthropomorphic VR access to the machine-centred underpinnings of the Internet envisaged by science fiction writers such as Gibson awaits the invention of appropriate physical, spatial and visual metaphors for the underlying information.
13.2.5 Anthropomorphic VR
Immersive, computer-generated environments attempt to present a ‘realistic’ world of experience to users. As with the previous section on machine-centred VR, our concern in this section is not with the agents populating these virtual realms, but with the ethical issues underlying the features of the environments themselves and the uses to which they are put (Brey 1999b). Anthropomorphic VR has been deployed for many purposes, including games, cybersex, tele-conferencing, pilot training, soldier training, elementary school education, and it is under investigation for use in many more applications, for instance police lineups. These applications make different demands on the amount of realism involved. Flight simulators, for instance, aim to be realistic in almost all respects. Educational applications for anthropomorphic VR, do not always aim for full realism, since pedagogical objectives may lead their designers to limit both the range of actions that individuals can take within the virtual world, and to limit the range and complexity of responses to user actions. Consider, for example, all the ways that a school science project or field trip can go awry. A VR environment that is intended to substitute for a field trip may be deliberately designed so that it is less likely to fail (Barab et al. 2007). One might ask whether such a deliberate departure from realism teaches children the wrong lesson about how easy it is to do good scientific research, by oversimplifying the task of data collection and setting up in them the expectation that the world is more predictable than it actually is. But this seems like a relatively minor worry when the alternative might be no field trip at all because of the difficulty and expense involved.
VR-based games typically depart from realism in numerous ways. Games may, for instance, enable users to pretend that they have magical powers that can transcend the limitations of the physical world. It is hard to see how exercising the capacity to act out such fantasies would be harmful or unethical, except insofar as the activities are so intrinsically rewarding that they lead to addictive behaviour. A more worrisome aspect of anthropomorphic VR concerns what it affords to individuals by way of freedom to act out violent or coercive fantasies. When the person is operating as a single user within an isolated VR application, the ethical issues raised by actions within that environment are necessarily indirect. Their indirectness may not make them any less pressing, however, as concerns are widespread about how the ability to rehearse violent, misanthropic or misogynistic actions in video games may desensitize users to those issues in the real world, or even provide a training effect that makes game players more likely to engage in similar actions outside of the VR context. However, there is also a longstanding response that violent games actually divert antisocial propensities into a context where they are harmless. The scientific research on this topic remains controversial (but see Ess 2009 for a balanced treatment).
When the VR context involves networked computers and multiple users, the ethical issues take on a different character because there is the possibility of direct harm to other users. A case of ‘cyberbullying’ of a teenager by an adult in the United States using the MySpace social networking site allegedly led to the teenager's suicide, but resulted only in conviction in 2008 on misdemeanour charges of illegal computer access; a more serious charge of conspiracy was dismissed. Social networking in VR environments also allows users to simulate acts of violence or coercion that would be unethical or socially unacceptable outside the VR context. For instance, a case of virtual rape has been reported in Second Life and investigated by Belgian police. Also, real opportunities for fraud and for theft of valuable property are made possible with Second Life. The question of whether VR relationships can be adulterous has also been raised (Thomas 2004, Stuart 2008; see also Ess 2009), and regardless of the philosophical discussion of this point, there have been cases where online relationships have resulted in real-world divorce.
In affording the opportunity to transcend physical or social barriers, anthropomorphic VR provides technological enhancement of personal autonomy. Such opportunities are not ethically neutral, however. They may affect other individuals directly, or have indirect effects by training or reinforcing habits of action that may spill over into the real world.

13.3 Conclusion

In this fast-paced tour of the ethics of artificial life, artificial agents and virtual realities it has been suggested that the common theme uniting these technologies concerns the ethics of autonomous agency. For artificial life, robots and software bots, there are ethical questions concerning the agency inherent in these systems themselves, whether they can be made to respect ethical boundaries, and the ways in which, as these systems become more sophisticated, they may lead humanity down a path which reduces human freedom or autonomy. For virtual reality, the ethical questions concern the ways in which they may provide new opportunities for human agency, or they may restrict those opportunities, and whether they may habituate or desensitize people to violence and other acts of oppression, or perhaps train people to act in antisocial or unethical ways.
Neither the benefits nor risks of these technologies are easily predictable. The histories of both techno-utopianism and techno-pessimism are full of embarrassing overstatements. The best we can do is articulate long-term goals for these technologies while carefully monitoring the current developments to see how well those goals are being accomplished, to identify unanticipated problems before the technologies become so entrenched that they are difficult to change, and to make the designers of these technologies more cognizant of the extent to which their computer systems embody their values (Nissenbaum 2001), many of which are culturally specific (Ess 2009). It is with this approach in mind that the chapter has eschewed futuristic questions about whether our artefacts will ever themselves be appropriately condemned for their moral failings (Dennett 1996). Such speculations are entertaining and make for great discussions in undergraduate philosophy courses, but the more pressing ethical issues confronting us in the technologies of autonomous agency are those outlined above.
Acknowledgements
I thank Luciano Floridi and Wendell Wallach for their comments.