Ethics of Artificial Intelligence and Robotics
Artificial intelligence (AI) and robotics are digital technologies that will have significant impact on the development of humanity in the near future. They have raised fundamental questions about what we should do with these systems, what the systems themselves should do, what risks they involve, and how we can control these.
After the Introduction to the field (§1), the main themes (§2) of this article are: Ethical issues that arise with AI systems as objects, i.e., tools made and used by humans. This includes issues of privacy (§2.1) and manipulation (§2.2), opacity (§2.3) and bias (§2.4), human-robot interaction (§2.5), employment (§2.6), and the effects of autonomy (§2.7). Then AI systems as subjects, i.e., ethics for the AI systems themselves in machine ethics (§2.8) and artificial moral agency (§2.9). Finally, the problem of a possible future AI superintelligence leading to a “singularity” (§2.10). We close with a remark on the vision of AI (§3).
For each section within these themes, we provide a general explanation of the ethical issues, outline existing positions and arguments, then analyse how these play out with current technologies and finally, what policy consequences may be drawn.
1.1 Background of the Field
The ethics of AI and robotics is often focused on “concerns” of various sorts, which is a typical response to new technologies. Many such concerns turn out to be rather quaint (trains are too fast for souls); some are predictably wrong when they suggest that the technology will fundamentally change humans (telephones will destroy personal communication, writing will destroy memory, video cassettes will make going out redundant); some are broadly correct but moderately relevant (digital technology will destroy industries that make photographic film, cassette tapes, or vinyl records); but some are broadly correct and deeply relevant (cars will kill children and fundamentally change the landscape). The task of an article such as this is to analyse the issues and to deflate the non-issues.
Some technologies, like nuclear power, cars, or plastics, have caused ethical and political discussion and significant policy efforts to control the trajectory these technologies, usually only once some damage is done. In addition to such “ethical concerns”, new technologies challenge current norms and conceptual systems, which is of particular interest to philosophy. Finally, once we have understood a technology in its context, we need to shape our societal response, including regulation and law. All these features also exist in the case of new AI and Robotics technologies—plus the more fundamental fear that they may end the era of human control on Earth.
The ethics of AI and robotics has seen significant press coverage in recent years, which supports related research, but also may end up undermining it: the press often talks as if the issues under discussion were just predictions of what future technology will bring, and as though we already know what would be most ethical and how to achieve that. Press coverage thus focuses on risk, security (Brundage et al. 2018, in the Other Internet Resources section below, hereafter [OIR]), and prediction of impact (e.g., on the job market). The result is a discussion of essentially technical problems that focus on how to achieve a desired outcome. Current discussions in policy and industry are also motivated by image and public relations, where the label “ethical” is really not much more than the new “green”, perhaps used for “ethics washing”. For a problem to qualify as a problem for AI ethics would require that we do not readily know what the right thing to do is. In this sense, job loss, theft, or killing with AI is not a problem in ethics, but whether these are permissible under certain circumstances is a problem. This article focuses on the genuine problems of ethics where we do not readily know what the answers are.
A last caveat: The ethics of AI and robotics is a very young field within applied ethics, with significant dynamics, but few well-established issues and no authoritative overviews—though there is a promising outline (European Group on Ethics in Science and New Technologies 2018) and there are beginnings on societal impact (Floridi et al. 2018; Taddeo and Floridi 2018; S. Taylor et al. 2018; Walsh 2018; Bryson 2019; Gibert 2019; Whittlestone et al. 2019), and policy recommendations (AI HLEG 2019 [OIR]; IEEE 2019). So this article cannot merely reproduce what the community has achieved thus far, but must propose an ordering where little order exists.
1.2 AI & Robotics
The notion of “artificial intelligence” (AI) is understood broadly as any kind of artificial computational system that shows intelligent behaviour, i.e., complex behaviour that is conducive to reaching goals. In particular, we do not wish to restrict “intelligence” to what would require intelligence if done by humans, as Minsky had suggested (1985). This means we incorporate a range of machines, including those in “technical AI”, that show only limited abilities in learning or reasoning but excel at the automation of particular tasks, as well as machines in “general AI” that aim to create a generally intelligent agent.
AI somehow gets closer to our skin than other technologies—thus the field of “philosophy of AI”. Perhaps this is because the project of AI is to create machines that have a feature central to how we humans see ourselves, namely as feeling, thinking, intelligent beings. The main purposes of an artificially intelligent agent probably involve sensing, modelling, planning and action, but current AI applications also include perception, text analysis, natural language processing (NLP), logical reasoning, game-playing, decision support systems, data analytics, predictive analytics, as well as autonomous vehicles and other forms of robotics (P. Stone et al. 2016). AI may involve any number of computational techniques to achieve these aims, be that classical symbol-manipulating AI, inspired by natural cognition, or machine learning via neural networks (Goodfellow, Bengio, and Courville 2016; Silver et al. 2018).
Historically, it is worth noting that the term “AI” was used as above ca. 1950–1975, then came into disrepute during the “AI winter”, ca. 1975–1995, and narrowed. As a result, areas such as “machine learning”, “natural language processing” and “data science” were often not labelled as “AI”. Since ca. 2010, the use has broadened again, and at times almost all of computer science and even high-tech is lumped under “AI”. Now it is a name to be proud of, a booming industry with massive capital investment (Shoham et al. 2018), and on the edge of hype again. As Erik Brynjolfsson noted, it may allow us to
virtually eliminate global poverty, massively reduce disease and provide better education to almost everyone on the planet. (quoted in Anderson, Rainie, and Luchsinger 2018)
While AI can be entirely software, robots are physical machines that move. Robots are subject to physical impact, typically through “sensors”, and they exert physical force onto the world, typically through “actuators”, like a gripper or a turning wheel. Accordingly, autonomous cars or planes are robots, and only a minuscule portion of robots is “humanoid” (human-shaped), like in the movies. Some robots use AI, and some do not: Typical industrial robots blindly follow completely defined scripts with minimal sensory input and no learning or reasoning (around 500,000 such new industrial robots are installed each year (IFR 2019 [OIR])). It is probably fair to say that while robotics systems cause more concerns in the general public, AI systems are more likely to have a greater impact on humanity. Also, AI or robotics systems for a narrow set of tasks are less likely to cause new issues than systems that are more flexible and autonomous.
Robotics and AI can thus be seen as covering two overlapping sets of systems: systems that are only AI, systems that are only robotics, and systems that are both. We are interested in all three; the scope of this article is thus not only the intersection, but the union, of both sets.
1.3 A Note on Policy
Policy is only one of the concerns of this article. There is significant public discussion about AI ethics, and there are frequent pronouncements from politicians that the matter requires new policy, which is easier said than done: Actual technology policy is difficult to plan and enforce. It can take many forms, from incentives and funding, infrastructure, taxation, or good-will statements, to regulation by various actors, and the law. Policy for AI will possibly come into conflict with other aims of technology policy or general policy. Governments, parliaments, associations, and industry circles in industrialised countries have produced reports and white papers in recent years, and some have generated good-will slogans (“trusted/responsible/humane/human-centred/good/beneficial AI”), but is that what is needed? For a survey, see Jobin, Ienca, and Vayena (2019) and V. Müller’s list of PT-AI Policy Documents and Institutions.
For people who work in ethics and policy, there might be a tendency to overestimate the impact and threats from a new technology, and to underestimate how far current regulation can reach (e.g., for product liability). On the other hand, there is a tendency for businesses, the military, and some public administrations to “just talk” and do some “ethics washing” in order to preserve a good public image and continue as before. Actually implementing legally binding regulation would challenge existing business models and practices. Actual policy is not just an implementation of ethical theory, but subject to societal power structures—and the agents that do have the power will push against anything that restricts them. There is thus a significant risk that regulation will remain toothless in the face of economical and political power.
Though very little actual policy has been produced, there are some notable beginnings: The latest EU policy document suggests “trustworthy AI” should be lawful, ethical, and technically robust, and then spells this out as seven requirements: human oversight, technical robustness, privacy and data governance, transparency, fairness, well-being, and accountability (AI HLEG 2019 [OIR]). Much European research now runs under the slogan of “responsible research and innovation” (RRI), and “technology assessment” has been a standard field since the advent of nuclear power. Professional ethics is also a standard field in information technology, and this includes issues that are relevant in this article. Perhaps a “code of ethics” for AI engineers, analogous to the codes of ethics for medical doctors, is an option here (Véliz 2019). What data science itself should do is addressed in (L. Taylor and Purtova 2019). We also expect that much policy will eventually cover specific uses or technologies of AI and robotics, rather than the field as a whole. A useful summary of an ethical framework for AI is given in (European Group on Ethics in Science and New Technologies 2018: 13ff). On general AI policy, see Calo (2018) as well as Crawford and Calo (2016); Stahl, Timmermans, and Mittelstadt (2016); Johnson and Verdicchio (2017); and Giubilini and Savulescu (2018). A more political angle of technology is often discussed in the field of “Science and Technology Studies” (STS). As books like The Ethics of Invention (Jasanoff 2016) show, concerns in STS are often quite similar to those in ethics (Jacobs et al. 2019 [OIR]). In this article, we discuss the policy for each type of issue separately rather than for AI or robotics in general.
2. Main Debates
In this section we outline the ethical issues of human use of AI and robotics systems that can be more or less autonomous—which means we look at issues that arise with certain uses of the technologies which would not arise with others. It must be kept in mind, however, that technologies will always cause some uses to be easier, and thus more frequent, and hinder other uses. The design of technical artefacts thus has ethical relevance for their use (Houkes and Vermaas 2010; Verbeek 2011), so beyond “responsible use”, we also need “responsible design” in this field. The focus on use does not presuppose which ethical approaches are best suited for tackling these issues; they might well be virtue ethics (Vallor 2017) rather than consequentialist or value-based (Floridi et al. 2018). This section is also neutral with respect to the question whether AI systems truly have “intelligence” or other mental properties: It would apply equally well if AI and robotics are merely seen as the current face of automation (cf. Müller forthcoming-b).
2.1 Privacy & Surveillance
There is a general discussion about privacy and surveillance in information technology (e.g., Macnish 2017; Roessler 2017), which mainly concerns the access to private data and data that is personally identifiable. Privacy has several well recognised aspects, e.g., “the right to be let alone”, information privacy, privacy as an aspect of personhood, control over information about oneself, and the right to secrecy (Bennett and Raab 2006). Privacy studies have historically focused on state surveillance by secret services but now include surveillance by other state agents, businesses, and even individuals. The technology has changed significantly in the last decades while regulation has been slow to respond (though there is the Regulation (EU) 2016/679)—the result is a certain anarchy that is exploited by the most powerful players, sometimes in plain sight, sometimes in hiding.
The digital sphere has widened greatly: All data collection and storage is now digital, our lives are increasingly digital, most digital data is connected to a single Internet, and there is more and more sensor technology in use that generates data about non-digital aspects of our lives. AI increases both the possibilities of intelligent data collection and the possibilities for data analysis. This applies to blanket surveillance of whole populations as well as to classic targeted surveillance. In addition, much of the data is traded between agents, usually for a fee.
The data trail we leave behind is how our “free” services are paid for—but we are not told about that data collection and the value of this new raw material, and we are manipulated into leaving ever more such data. For the “big 5” companies (Amazon, Google/Alphabet, Microsoft, Apple, Facebook), the main data-collection part of their business appears to be based on deception, exploiting human weaknesses, furthering procrastination, generating addiction, and manipulation (Harris 2016 [OIR]). The primary focus of social media, gaming, and most of the Internet in this “surveillance economy” is to gain, maintain, and direct attention—and thus data supply. “Surveillance is the business model of the Internet” (Schneier 2015). This surveillance and attention economy is sometimes called “surveillance capitalism” (Zuboff 2019). It has caused many attempts to escape from the grasp of these corporations, e.g., in exercises of “minimalism” (Newport 2019), sometimes through the open source movement, but it appears that present-day citizens have lost the degree of autonomy needed to escape while fully continuing with their life and work. We have lost ownership of our data, if “ownership” is the right relation here. Arguably, we have lost control of our data.
These systems will often reveal facts about us that we ourselves wish to suppress or are not aware of: they know more about us than we know ourselves. Even just observing online behaviour allows insights into our mental states (Burr and Christianini 2019) and manipulation (see below section 2.2). This has led to calls for the protection of “derived data” (Wachter and Mittelstadt 2019). With the last sentence of his bestselling book, Homo Deus, Harari asks about the long-term consequences of AI:
What will happen to society, politics and daily life when non-conscious but highly intelligent algorithms know us better than we know ourselves? (2016: 462)
Robotic devices have not yet played a major role in this area, except for security patrolling, but this will change once they are more common outside of industry environments. Together with the “Internet of things”, the so-called “smart” systems (phone, TV, oven, lamp, virtual assistant, home,…), “smart city” (Sennett 2018), and “smart governance”, they are set to become part of the data-gathering machinery that offers more detailed data, of different types, in real time, with ever more information.
Privacy-preserving techniques that can largely conceal the identity of persons or groups are now a standard staple in data science; they include (relative) anonymisation , access control (plus encryption), and other models where computation is carried out with fully or partially encrypted input data (Stahl and Wright 2018); in the case of “differential privacy”, this is done by adding calibrated noise to encrypt the output of queries (Dwork et al. 2006; Abowd 2017). While requiring more effort and cost, such techniques can avoid many of the privacy issues. Some companies have also seen better privacy as a competitive advantage that can be leveraged and sold at a price.
One of the major practical difficulties is to actually enforce regulation, both on the level of the state and on the level of the individual who has a claim. They must identify the responsible legal entity, prove the action, perhaps prove intent, find a court that declares itself competent … and eventually get the court to actually enforce its decision. Well-established legal protection of rights such as consumer rights, product liability, and other civil liability or protection of intellectual property rights is often missing in digital products, or hard to enforce. This means that companies with a “digital” background are used to testing their products on the consumers without fear of liability while heavily defending their intellectual property rights. This “Internet Libertarianism” is sometimes taken to assume that technical solutions will take care of societal problems by themselves (Mozorov 2013).
2.2 Manipulation of Behaviour
The ethical issues of AI in surveillance go beyond the mere accumulation of data and direction of attention: They include the use of information to manipulate behaviour, online and offline, in a way that undermines autonomous rational choice. Of course, efforts to manipulate behaviour are ancient, but they may gain a new quality when they use AI systems. Given users’ intense interaction with data systems and the deep knowledge about individuals this provides, they are vulnerable to “nudges”, manipulation, and deception. With sufficient prior data, algorithms can be used to target individuals or small groups with just the kind of input that is likely to influence these particular individuals. A ’nudge‘ changes the environment such that it influences behaviour in a predictable way that is positive for the individual, but easy and cheap to avoid (Thaler & Sunstein 2008). There is a slippery slope from here to paternalism and manipulation.
Many advertisers, marketers, and online sellers will use any legal means at their disposal to maximise profit, including exploitation of behavioural biases, deception, and addiction generation (Costa and Halpern 2019 [OIR]). Such manipulation is the business model in much of the gambling and gaming industries, but it is spreading, e.g., to low-cost airlines. In interface design on web pages or in games, this manipulation uses what is called “dark patterns” (Mathur et al. 2019). At this moment, gambling and the sale of addictive substances are highly regulated, but online manipulation and addiction are not—even though manipulation of online behaviour is becoming a core business model of the Internet.
Furthermore, social media is now the prime location for political propaganda. This influence can be used to steer voting behaviour, as in the Facebook-Cambridge Analytica “scandal” (Woolley and Howard 2017; Bradshaw, Neudert, and Howard 2019) and—if successful—it may harm the autonomy of individuals (Susser, Roessler, and Nissenbaum 2019).
Improved AI “faking” technologies make what once was reliable evidence into unreliable evidence—this has already happened to digital photos, sound recordings, and video. It will soon be quite easy to create (rather than alter) “deep fake” text, photos, and video material with any desired content. Soon, sophisticated real-time interaction with persons over text, phone, or video will be faked, too. So we cannot trust digital interactions while we are at the same time increasingly dependent on such interactions.
One more specific issue is that machine learning techniques in AI rely on training with vast amounts of data. This means there will often be a trade-off between privacy and rights to data vs. technical quality of the product. This influences the consequentialist evaluation of privacy-violating practices.
The policy in this field has its ups and downs: Civil liberties and the protection of individual rights are under intense pressure from businesses’ lobbying, secret services, and other state agencies that depend on surveillance. Privacy protection has diminished massively compared to the pre-digital age when communication was based on letters, analogue telephone communications, and personal conversation and when surveillance operated under significant legal constraints.
While the EU General Data Protection Regulation (Regulation (EU) 2016/679) has strengthened privacy protection, the US and China prefer growth with less regulation (Thompson and Bremmer 2018), likely in the hope that this provides a competitive advantage. It is clear that state and business actors have increased their ability to invade privacy and manipulate people with the help of AI technology and will continue to do so to further their particular interests—unless reined in by policy in the interest of general society.
2.3 Opacity of AI Systems
Opacity and bias are central issues in what is now sometimes called “data ethics” or “big data ethics” (Floridi and Taddeo 2016; Mittelstadt and Floridi 2016). AI systems for automated decision support and “predictive analytics” raise “significant concerns about lack of due process, accountability, community engagement, and auditing” (Whittaker et al. 2018: 18ff). They are part of a power structure in which “we are creating decision-making processes that constrain and limit opportunities for human participation” (Danaher 2016b: 245). At the same time, it will often be impossible for the affected person to know how the system came to this output, i.e., the system is “opaque” to that person. If the system involves machine learning, it will typically be opaque even to the expert, who will not know how a particular pattern was identified, or even what the pattern is. Bias in decision systems and data sets is exacerbated by this opacity. So, at least in cases where there is a desire to remove bias, the analysis of opacity and bias go hand in hand, and political response has to tackle both issues together.
Many AI systems rely on machine learning techniques in (simulated) neural networks that will extract patterns from a given dataset, with or without “correct” solutions provided; i.e., supervised, semi-supervised or unsupervised. With these techniques, the “learning” captures patterns in the data and these are labelled in a way that appears useful to the decision the system makes, while the programmer does not really know which patterns in the data the system has used. In fact, the programs are evolving, so when new data comes in, or new feedback is given (“this was correct”, “this was incorrect”), the patterns used by the learning system change. What this means is that the outcome is not transparent to the user or programmers: it is opaque. Furthermore, the quality of the program depends heavily on the quality of the data provided, following the old slogan “garbage in, garbage out”. So, if the data already involved a bias (e.g., police data about the skin colour of suspects), then the program will reproduce that bias. There are proposals for a standard description of datasets in a “datasheet” that would make the identification of such bias more feasible (Gebru et al. 2018 [OIR]). There is also significant recent literature about the limitations of machine learning systems that are essentially sophisticated data filters (Marcus 2018 [OIR]). Some have argued that the ethical problems of today are the result of technical “shortcuts” AI has taken (Cristianini forthcoming).
There are several technical activities that aim at “explainable AI”, starting with (Van Lent, Fisher, and Mancuso 1999; Lomas et al. 2012) and, more recently, a DARPA programme (Gunning 2017 [OIR]). More broadly, the demand for
a mechanism for elucidating and articulating the power structures, biases, and influences that computational artefacts exercise in society (Diakopoulos 2015: 398)
is sometimes called “algorithmic accountability reporting”. This does not mean that we expect an AI to “explain its reasoning”—doing so would require far more serious moral autonomy than we currently attribute to AI systems (see below §2.10).
The politician Henry Kissinger pointed out that there is a fundamental problem for democratic decision-making if we rely on a system that is supposedly superior to humans, but cannot explain its decisions. He says we may have “generated a potentially dominating technology in search of a guiding philosophy” (Kissinger 2018). Danaher (2016b) calls this problem “the threat of algocracy” (adopting the previous use of ‘algocracy’ from Aneesh 2002 [OIR], 2006). In a similar vein, Cave (2019) stresses that we need a broader societal move towards more “democratic” decision-making to avoid AI being a force that leads to a Kafka-style impenetrable suppression system in public administration and elsewhere. The political angle of this discussion has been stressed by O’Neil in her influential book Weapons of Math Destruction (2016), and by Yeung and Lodge (2019).
In the EU, some of these issues have been taken into account with the (Regulation (EU) 2016/679), which foresees that consumers, when faced with a decision based on data processing, will have a legal “right to explanation”—how far this goes and to what extent it can be enforced is disputed (Goodman and Flaxman 2017; Wachter, Mittelstadt, and Floridi 2016; Wachter, Mittelstadt, and Russell 2017). Zerilli et al. (2019) argue that there may be a double standard here, where we demand a high level of explanation for machine-based decisions despite humans sometimes not reaching that standard themselves.
2.4 Bias in Decision Systems
Automated AI decision support systems and “predictive analytics” operate on data and produce a decision as “output”. This output may range from the relatively trivial to the highly significant: “this restaurant matches your preferences”, “the patient in this X-ray has completed bone growth”, “application to credit card declined”, “donor organ will be given to another patient”, “bail is denied”, or “target identified and engaged”. Data analysis is often used in “predictive analytics” in business, healthcare, and other fields, to foresee future developments—since prediction is easier, it will also become a cheaper commodity. One use of prediction is in “predictive policing” (NIJ 2014 [OIR]), which many fear might lead to an erosion of public liberties (Ferguson 2017) because it can take away power from the people whose behaviour is predicted. It appears, however, that many of the worries about policing depend on futuristic scenarios where law enforcement foresees and punishes planned actions, rather than waiting until a crime has been committed (like in the 2002 film “Minority Report”). One concern is that these systems might perpetuate bias that was already in the data used to set up the system, e.g., by increasing police patrols in an area and discovering more crime in that area. Actual “predictive policing” or “intelligence led policing” techniques mainly concern the question of where and when police forces will be needed most. Also, police officers can be provided with more data, offering them more control and facilitating better decisions, in workflow support software (e.g., “ArcGIS”). Whether this is problematic depends on the appropriate level of trust in the technical quality of these systems, and on the evaluation of aims of the police work itself. Perhaps a recent paper title points in the right direction here: “AI ethics in predictive policing: From models of threat to an ethics of care” (Asaro 2019).
Bias typically surfaces when unfair judgments are made because the individual making the judgment is influenced by a characteristic that is actually irrelevant to the matter at hand, typically a discriminatory preconception about members of a group. So, one form of bias is a learned cognitive feature of a person, often not made explicit. The person concerned may not be aware of having that bias—they may even be honestly and explicitly opposed to a bias they are found to have (e.g., through priming, cf. Graham and Lowery 2004). On fairness vs. bias in machine learning, see Binns (2018).
Apart from the social phenomenon of learned bias, the human cognitive system is generally prone to have various kinds of “cognitive biases”, e.g., the “confirmation bias”: humans tend to interpret information as confirming what they already believe. This second form of bias is often said to impede performance in rational judgment (Kahnemann 2011)—though at least some cognitive biases generate an evolutionary advantage, e.g., economical use of resources for intuitive judgment. There is a question whether AI systems could or should have such cognitive bias.
A third form of bias is present in data when it exhibits systematic error, e.g., “statistical bias”. Strictly, any given dataset will only be unbiased for a single kind of issue, so the mere creation of a dataset involves the danger that it may be used for a different kind of issue, and then turn out to be biased for that kind. Machine learning on the basis of such data would then not only fail to recognise the bias, but codify and automate the “historical bias”. Such historical bias was discovered in an automated recruitment screening system at Amazon (discontinued early 2017) that discriminated against women—presumably because the company had a history of discriminating against women in the hiring process. The “Correctional Offender Management Profiling for Alternative Sanctions” (COMPAS), a system to predict whether a defendant would re-offend, was found to be as successful (65.2% accuracy) as a group of random humans (Dressel and Farid 2018) and to produce more false positives and less false negatives for black defendants. The problem with such systems is thus bias plus humans placing excessive trust in the systems. The political dimensions of such automated systems in the USA are investigated in Eubanks (2018).
There are significant technical efforts to detect and remove bias from AI systems, but it is fair to say that these are in early stages: see UK Institute for Ethical AI & Machine Learning (Brownsword, Scotford, and Yeung 2017; Yeung and Lodge 2019). It appears that technological fixes have their limits in that they need a mathematical notion of fairness, which is hard to come by (Whittaker et al. 2018: 24ff; Selbst et al. 2019), as is a formal notion of “race” (see Benthall and Haynes 2019). An institutional proposal is in (Veale and Binns 2017).
2.5 Human-Robot Interaction
Human-robot interaction (HRI) is an academic fields in its own right, which now pays significant attention to ethical matters, the dynamics of perception from both sides, and both the different interests present in and the intricacy of the social context, including co-working (e.g., Arnold and Scheutz 2017). Useful surveys for the ethics of robotics include Calo, Froomkin, and Kerr (2016); Royakkers and van Est (2016); Tzafestas (2016); a standard collection of papers is Lin, Abney, and Jenkins (2017).
While AI can be used to manipulate humans into believing and doing things (see section 2.2), it can also be used to drive robots that are problematic if their processes or appearance involve deception, threaten human dignity, or violate the Kantian requirement of “respect for humanity”. Humans very easily attribute mental properties to objects, and empathise with them, especially when the outer appearance of these objects is similar to that of living beings. This can be used to deceive humans (or animals) into attributing more intellectual or even emotional significance to robots or AI systems than they deserve. Some parts of humanoid robotics are problematic in this regard (e.g., Hiroshi Ishiguro’s remote-controlled Geminoids), and there are cases that have been clearly deceptive for public-relations purposes (e.g. on the abilities of Hanson Robotics’ “Sophia”). Of course, some fairly basic constraints of business ethics and law apply to robots, too: product safety and liability, or non-deception in advertisement. It appears that these existing constraints take care of many concerns that are raised. There are cases, however, where human-human interaction has aspects that appear specifically human in ways that can perhaps not be replaced by robots: care, love, and sex.
2.5.1 Example (a) Care Robots
The use of robots in health care for humans is currently at the level of concept studies in real environments, but it may become a usable technology in a few years, and has raised a number of concerns for a dystopian future of de-humanised care (A. Sharkey and N. Sharkey 2011; Robert Sparrow 2016). Current systems include robots that support human carers/caregivers (e.g., in lifting patients, or transporting material), robots that enable patients to do certain things by themselves (e.g., eat with a robotic arm), but also robots that are given to patients as company and comfort (e.g., the “Paro” robot seal). For an overview, see van Wynsberghe (2016); Nørskov (2017); Fosch-Villaronga and Albo-Canals (2019), for a survey of users Draper et al. (2014).
One reason why the issue of care has come to the fore is that people have argued that we will need robots in ageing societies. This argument makes problematic assumptions, namely that with longer lifespan people will need more care, and that it will not be possible to attract more humans to caring professions. It may also show a bias about age (Jecker forthcoming). Most importantly, it ignores the nature of automation, which is not simply about replacing humans, but about allowing humans to work more efficiently. It is not very clear that there really is an issue here since the discussion mostly focuses on the fear of robots de-humanising care, but the actual and foreseeable robots in care are assistive robots for classic automation of technical tasks. They are thus “care robots” only in a behavioural sense of performing tasks in care environments, not in the sense that a human “cares” for the patients. It appears that the success of “being cared for” relies on this intentional sense of “care”, which foreseeable robots cannot provide. If anything, the risk of robots in care is the absence of such intentional care—because less human carers may be needed. Interestingly, caring for something, even a virtual agent, can be good for the carer themselves (Lee et al. 2019). A system that pretends to care would be deceptive and thus problematic—unless the deception is countered by sufficiently large utility gain (Coeckelbergh 2016). Some robots that pretend to “care” on a basic level are available (Paro seal) and others are in the making. Perhaps feeling cared for by a machine, to some extent, is progress for come patients.
2.5.2 Example (b) Sex Robots
It has been argued by several tech optimists that humans will likely be interested in sex and companionship with robots and be comfortable with the idea (Levy 2007). Given the variation of human sexual preferences, including sex toys and sex dolls, this seems very likely: The question is whether such devices should be manufactured and promoted, and whether there should be limits in this touchy area. It seems to have moved into the mainstream of “robot philosophy” in recent times (Sullins 2012; Danaher and McArthur 2017; N. Sharkey et al. 2017 [OIR]; Bendel 2018; Devlin 2018).
Humans have long had deep emotional attachments to objects, so perhaps companionship or even love with a predictable android is attractive, especially to people who struggle with actual humans, and already prefer dogs, cats, birds, a computer or a tamagotchi. Danaher (2019b) argues against (Nyholm and Frank 2017) that these can be true friendships, and is thus a valuable goal. It certainly looks like such friendship might increase overall utility, even if lacking in depth. In these discussions there is an issue of deception, since a robot cannot (at present) mean what it says, or have feelings for a human. It is well known that humans are prone to attribute feelings and thoughts to entities that behave as if they had sentience,even to clearly inanimate objects that show no behaviour at all. Also, paying for deception seems to be an elementary part of the traditional sex industry.
Finally, there are concerns that have often accompanied matters of sex, namely consent (Frank and Nyholm 2017), aesthetic concerns, and the worry that humans may be “corrupted” by certain experiences. Old fashioned though this may seem, human behaviour is influenced by experience, and it is likely that pornography or sex robots support the perception of other humans as mere objects of desire, or even recipients of abuse, and thus ruin a deeper sexual and erotic experience. In this vein, the “Campaign Against Sex Robots” argues that these devices are a continuation of slavery and prostitution (Richardson 2016).
2.6 Automation and Employment
It seems clear that AI and robotics will lead to significant gains in productivity and thus overall wealth. The attempt to increase productivity has often been a feature of the economy, though the emphasis on “growth” is a modern phenomenon (Harari 2016: 240). However, productivity gains through automation typically mean that fewer humans are required for the same output. This does not necessarily imply a loss of overall employment, however, because available wealth increases and that can increase demand sufficiently to counteract the productivity gain. In the long run, higher productivity in industrial societies has led to more wealth overall. Major labour market disruptions have occurred in the past, e.g., farming employed over 60% of the workforce in Europe and North-America in 1800, while by 2010 it employed ca. 5% in the EU, and even less in the wealthiest countries (European Commission 2013). In the 20 years between 1950 and 1970 the number of hired agricultural workers in the UK was reduced by 50% (Zayed and Loft 2019). Some of these disruptions lead to more labour-intensive industries moving to places with lower labour cost. This is an ongoing process.
Classic automation replaced human muscle, whereas digital automation replaces human thought or information-processing—and unlike physical machines, digital automation is very cheap to duplicate (Bostrom and Yudkowsky 2014). It may thus mean a more radical change on the labour market. So, the main question is: will the effects be different this time? Will the creation of new jobs and wealth keep up with the destruction of jobs? And even if it is not different, what are the transition costs, and who bears them? Do we need to make societal adjustments for a fair distribution of costs and benefits of digital automation?
Responses to the issue of unemployment from AI have ranged from the alarmed (Frey and Osborne 2013; Westlake 2014) to the neutral (Metcalf, Keller, and Boyd 2016 [OIR]; Calo 2018; Frey 2019) to the optimistic (Brynjolfsson and McAfee 2016; Harari 2016; Danaher 2019a). In principle, the labour market effect of automation seems to be fairly well understood as involving two channels:
(i) the nature of interactions between differently skilled workers and new technologies affecting labour demand and (ii) the equilibrium effects of technological progress through consequent changes in labour supply and product markets. (Goos 2018: 362)
What currently seems to happen in the labour market as a result of AI and robotics automation is “job polarisation” or the “dumbbell” shape (Goos, Manning, and Salomons 2009): The highly skilled technical jobs are in demand and highly paid, the low skilled service jobs are in demand and badly paid, but the mid-qualification jobs in factories and offices, i.e., the majority of jobs, are under pressure and reduced because they are relatively predictable, and most likely to be automated (Baldwin 2019).
Perhaps enormous productivity gains will allow the “age of leisure” to be realised, something (Keynes 1930) had predicted to occur around 2030, assuming a growth rate of 1% per annum. Actually, we have already reached the level he anticipated for 2030, but we are still working—consuming more and inventing ever more levels of organisation. Harari explains how this economic development allowed humanity to overcome hunger, disease, and war—and now we aim for immortality and eternal bliss through AI, thus his title Homo Deus (Harari 2016: 75).
In general terms, the issue of unemployment is an issue of how goods in a society should be justly distributed. A standard view is that distributive justice should be rationally decided from behind a “veil of ignorance” (Rawls 1971), i.e., as if one does not know what position in a society one would actually be taking (labourer or industrialist, etc.). Rawls thought the chosen principles would then support basic liberties and a distribution that is of greatest benefit to the least-advantaged members of society. It would appear that the AI economy has three features that make such justice unlikely: First, it operates in a largely unregulated environment where responsibility is often hard to allocate. Second, it operates in markets that have a “winner takes all” feature where monopolies develop quickly. Third, the “new economy” of the digital service industries is based on intangible assets, also called “capitalism without capital” (Haskel and Westlake 2017). This means that it is difficult to control multinational digital corporations that do not rely on a physical plant in a particular location. These three features seem to suggest that if we leave the distribution of wealth to free market forces, the result would be a heavily unjust distribution: And this is indeed a development that we can already see.
One interesting question that has not received too much attention is whether the development of AI is environmentally sustainable: Like all computing systems, AI systems produce waste that is very hard to recycle and they consume vast amounts of energy, especially for the training of machine learning systems (and even for the “mining” of cryptocurrency). Again, it appears that some actors in this space offload such costs to the general society.