香京julia种子在线播放

It is hard to overstate the possible implications of the two key enabling features in evolving robots: self-replication and random change in robot form and behavior. First, self-replication allows robots to multiply without human intervention and thus would raise the need for control over their reproduction. Second, mutation or random evolutionary changes in the design of the robots could create undesired robotic behaviors that may harm human interests. Before developing any new technology with such potentially large ramifications, we should determine the acceptability of its consequences and identify ways to anticipate unwanted effects (van de Poel 2016).

Several other fields of science have faced similar safety dilemmas during developments of new technology and subsequent experimentation. In health sciences, biomedical ethical dilemmas are typically evaluated using a principle-based approach, based on the four principles of Beauchamp and Childress (Beauchamp and Childress 2019): autonomy, non-maleficence (avoiding harm), beneficence, and justice. Within the context of technological experimentation, the concept of responsibility has been added (van de Poel 2016), and specifically in the field of Artificial Intelligence (AI), a call has been made for adding the property of “explicability” (Floridi et al., 2018). This property entails that when AI-powered algorithms are used to make morally-sensitive decisions, humans should be able to obtain “a factual, direct, and clear explanation of the decision-making process” (Floridi et al., 2018), or of the decision resulting from the algorithm (Robbins 2019).

In evolutionary robotics all of these principles have clear relevance, but, most pressingly, the risk of harm and the question of responsibility need to be considered in more detail. These, in turn, are intimately related to the crucial issue of control and the potential loss of it. In order for a particular human being or group of human beings to be responsible for some process or outcome, it is usually thought that they need to have some degree of control of the process or outcome. Moreover, loss of control can be viewed as a form of harm, because it is typically seen as undermining human autonomy, and it may compromise other values, such as well-being, which depend to some extent on our ability to control what happens around us.

The issue of risk in the field of AI has previously been considered in relation to control concerns associated with the development of superintelligence (Bostrom 2014; Russell 2019; Russell and Norvig 2020). A notable difference between superintelligence-related concerns and ER-related worries, however, is the perceived probability of the risk. Many people find the idea of superintelligence either inherently implausible or at least something we need not worry about in the short run (Gordon and Nyholm 2021; Müller, 2020). More precisely, people may feel that although an excellent AI chess or Go player is manifestly possible, artificial general superintelligence is much less likely to emerge.

In contrast, evolving physical robots need not possess human level intelligence; animal level intelligence in such robots could be sufficient to do significant harm because of their physical features. Even without much individual intelligence and power, the evolved robots could potentially collaborate efficiently and perform much more complex tasks together than they could on their own. In other words, similar to highly social animals such as ants and wasps in the natural world, the number and cooperation among robots could be decisive factors. Therefore, the plausibility of a harmful scenario with evolving robots is all but trivial, and issues of control and the potential loss of it should be considered.

The most difficult aspect in anticipating possible risks of evolving robots is that we would be dealing with an evolving system that is inherently and continuously changing. The risk of harm therefore needs to be evaluated for potential future trajectories of the evolutionary process, not only for the current robots. We distinguish three key types of risks associated with the evolutionary process, connected to reproduction, selection, and emergent evolution, respectively:

Multiplication risk: The robots can evolve at high reproduction rates, resulting in uncontrolled population growth. If the robot population becomes too large, resources such as space, energy, and raw materials like air or water may be (locally) depleted. This effect can be compared to a locust plague: a swarm’s voracious feeding can completely devour agricultural crops over a vast area, leading to famine and starvation in the human population. While individual robots may not pose any significant risk, their high number and collective behavior can be dangerous.

Maladaptation risk: Evolving the robots for a specific task can lead to unwanted features or behaviors that benefit the robot’s assigned task, but that may be harmful to human society. For instance, robots may attempt to dismantle houses to use the stones or cut car tires for the rubber. In the most extreme cases, robots could harm humans if they hinder robots in performing their tasks. This type of risk can evolve because selection is “blind,” meaning the most effective solutions for the task will prevail, without taking other consequences of the evolved trait into account.

Domination risk: The robots could evolve to become the dominating “species,” not as a direct effect of selection, but rather as an emergent feature of the robot’s functionality (Badyaev 2011). This can happen if they become superior to humans intellectually, physically, or “emotionally” (being stable and consistent). As a result, they might become benevolent influencers or decision makers, implicitly or explicitly arranging life for us. This effect can be compared to a parent-child relationship where the parent is better in understanding and anticipating situations and therefore confines the spatial range and activities of the child. Even though humans may not be physically harmed by the robot’s dominating behavior, human autonomy would be, at least partly, diminished.

The risks of harm associated with robot evolution as identified above all arise from the underlying control problem of (semi)autonomous robotic systems. In the Artificial Intelligence literature, solutions to this control problem are often phrased in terms of meaningful human control (Santoni de Sio and van den Hoven, 2018). This term acknowledges that whereas there may be no direct control–e.g., a steering wheel in a car–it may still be possible to have indirect control allowing for allocation of responsibilities (Di Nucci 2020; Nyholm 2020). For evolving robots this would mean that precautionary design measures are required to control the evolutionary process itself. Such measures could include:

1) Centralized, externalized reproduction. A rigorous way of maintaining control over the system would be to set it up such that robot reproduction cannot take place “in the wild” but only in a centralized infrastructure–a reproduction center–where robot offspring can be made, for instance by 3D-printers and automated assembly facilities (Eiben et al., 2013; Hale et al., 2019). Limiting the reproduction to a single or a few centers not only allows keeping track of robot numbers, but also provides the option to restrict the number of robots produced per day. In addition, such a center could provide a possibility to test new robots for safety before releasing them into the outside world. Furthermore, a reproduction center can contain a “kill switch” that can be used to halt evolution by shutting down the reproduction process.

These control measures, meaningful as they are, can leave humans vulnerable because of the very nature of evolving systems, in which change is inherent. Evolving robots represent a whole new breed of machines that can and will change their form and behavior. This implies that robots could adapt their behavior to escape the implemented control measures. Therefore, controlling evolving robots is different from controlling the production of fixed entities, such as cars. One would therefore need to continuously adjust the control measures to stay ahead of evolutionary escape routes, not unlike a co-evolutionary arms race (Thompson 1994). In what follows, we highlight three possible evolutionary escape routes: two technology-related possibilities and one that exploits human emotional vulnerabilities and normative judgments.

First, the robots could develop solutions to circumvent the technological safeguards that have been put into place. A very unlikely, but conceivable escape route is the “Jurassic Park scenario,” where the robots find an alternative way of reproducing outside the central reproduction facility. To mitigate this risk, additional reproductive constraints may be necessary, e.g., using an ingredient that is necessary for being viable and controlling its supply (Ellery and Eiben 2019). A more realistic way of escaping control is that robots stop sharing their operational data and thereby evade monitoring. This could partly be resolved by a mandatory data recorder built into all robots, similar to the flight recorders (a.k.a. black box) in airplanes (Winfield and Jirotka 2017; Winkle 2020).

Second, while Bostrom [36] suggests “value loading” for robotic and AI systems, in the case of evolving robot populations it is important to realize that it would be risky to rely on the (current) features of individual robots. In an evolutionary process the robot’s features undergo change. This does not mean that creating certain features (such as values or goals) in the robots is without merit, but it should be combined with some form of verification that the goals/values continue to be present in the newly produced robots. This requires new technologies that effectively combine immutable values with adaptable robot features and protocols for a thorough screening of “newborn” robots before they are allowed to leave the reproduction facility.

A third possibility for evolving robots to escape human control is non-technological, exploiting deep-seated emotional response patterns. Specifically, humans may grow fond of robots, developing feelings of “affection” towards them (Carpenter 2016; Darling 2017). This emotional vulnerability is probably the result of the long evolutionary history of humans, which has equipped our brains with various motivational and affective pathways tuned to human psychology (Damiano and Dumouchel 2018; Nyholm 2020). Consequently, we are responding to robots with brains and emotional sensitivities that are well-adapted to interacting with fellow human beings and familiar animals, but not necessarily adequate to responding sensibly to machines. Robots and other artificially intelligent technologies, therefore, may “push our Darwinian buttons” in ways that we may not upon reflection find suitable (Turkle 2004).

These sensibilities can be exploited if robots evolve features humans tend to like such as, possibly, big eyes, certain locomotion patterns or “lovely” sounds and gestures. Such features can increase attachment, undermine human controller’s ability to remain objective and provide an evolutionary advantage on the long run. For instance, a robot could entice a human into supplying it with extra energy or allowing it to reproduce. Similarly, a “lovable” robot could prevent a human from switching off the robot or using the “kill switch” to shut down the evolution of the whole robotic species. These scenarios illustrate how emotions could get in the way of strict human control and induce an evolutionary bias [cf. (Bryson 2018)].

The above-mentioned considerations concern ways of controlling the process of robot evolution. But there are more conceptual–ethical–concerns as well. Being able to ascribe responsibility is always important when risks are involved, both from an ethical and a legal point of view. The relevant form of responsibility here does not only have a backward-looking component (who can be blamed when things have gone wrong?), but is also forward-looking and clarifies who should do what in order to maintain control, e.g., mitigating risks and taking precautions (Nihlen Fahlquist 2017; Di Nucci 2020). Thus, a prominent issue is a potential responsibility gap. A responsibility gap occurs when there are significant risks of harm for which someone should take responsibility, but there is no obvious candidate to ascribe the responsibility to (Matthias 2004; Sparrow 2007; Nyholm 2020). In the solutions above, the control envisioned will, at least in part, be exercised by humans. The crucial question is then how potential responsibility gaps might be filled.

At this point it may be instructive to refer to recent work by Santoni de Sio and Van den Hoven (Santoni de Sio and van den Hoven, 2018). They have developed a “track-and-trace” account of meaningful human control. The tracking part requires that the system behaves according to rules or paths that track human interests. In other words, the system should behave in a way that aligns with human values and interests. The tracing part requires that the robotic behavior can be traced back to at least one person who understands how the process works, as well as its moral and social significance. It might be added here that, ideally, this should work like when one is tracking and tracing a parcel: it should be possible to monitor how things are developing, just like one can monitor the journey of a parcel [(Nyholm 2020), p. 78].

The track-and-trace theory, understood as including the monitoring condition, looks promising from an ethical perspective for robot evolution. If the robot evolution is tracking human interests, if there are people who understand the process and its moral significance, and are able to monitor the robot evolution, then we can tentatively say that meaningful human control over this process has been achieved. If those conditions are fulfilled, that could help to fill any potential responsibility gaps.

The control solutions suggested above cover the “tracking” requirements from the track-and trace theory to a significant extent. The centralized, externalized reproduction centers would allow humans to monitor the numbers and types of robots produced each day, while the crystal ball would give insight into the future directions of the evolutionary path of the robots. Being able to monitor robot development in these ways, the humans involved would be able to observe whether human interests are being tracked. If not, they could use the “kill switch.” The tracing part however, would need to be developed further as, at the moment, we do not have an appropriate level of understanding nor control of how the evolutionary process unfolds. At the same time, if studying these evolutionary processes in robots would deepen our scientific understanding of evolution, this could in effect help to also fulfil the tracing condition.

That being said, the big challenge here is, again, the inherent variability of an evolutionary system where new features emerge through random mutations and recombination of parental properties. Even though the whole system, specifically the genetic code (the robotic DNA), the mutation operators, and recombination operators are designed by humans, it is not clear to what extent these humans can be held responsible for the effects over several generations. On the positive side, let us reiterate that robots are observable, thus the genetic material and genealogy tree of an evolving population can be logged and inspected. In principle, it is possible to examine a newly created genotype (the robotic zygote) before the corresponding phenotype (the robot offspring) is constructed and destroy the genotype if it fails a safety test.

In the sections above, our main concern was to protect the human race from evolving robots. However, the matter can be inverted if we conceive of robots that can evolve and learn as a form of artificial life. Considering them as a form of life implies different kinds of ethical considerations (Coeckelbergh 2012; Bryson 2018; Gunkel 2018; Danaher 2020), which go beyond the issues of affection and attachment to individual robots as discussed above, and refer to the whole robotic population. The key is to see the robot population as a species that requires some moral consideration. Such an ethical view could be motivated by two arguments.

First, these robots have the possibility of reproduction, and in biology the crucial difference between life and non-life is reproduction. In addition, these robots share other characteristics with other life forms, such as movement and energy consumption. Second, the robots are not only able to reproduce; they themselves have also evolved. In other words, these robots are not (just) the result of human design, but of an evolutionary process. If humans, generally, start to feel that these robots are forms of life–albeit artificial–this could entail some perceived moral obligations, like we may feel we have obligations towards whales, dolphins, dogs, and cats. In other words, we may feel that these robots–and along with them, their evolutionary process–deserve some level of protection. This could raise the issue of robot rights, similarly to how we think about animal rights (Gellers 2020).

Second, it could be questioned whether certain control-interventions, such as the use of the “kill switch”, are ethical regarding such forms of artificial life. An essential question here is if terminating evolutionary robots should be seen as switching off a machine or as killing a living being (Darling 2021). In any case, such moral considerations could potentially limit the possibilities of meaningful human control of robot evolution we have discussed.