Ever since the first conference on artificial intelligence (AI) was held in 1956 at Dartmouth College, the question of singularity is asked. The singularity is the event where AI exceeds human intelligence. Due to the singularity, the AI may outpace humanity and create a non-beneficial outcome. In contrast, the leakproof singularity describes a singularity having a beneficial outcome (where the AI is a problem solver).
Therefore, the paper discusses the question to what extent a leakproof singularity could happen. For the leakproof singularity, a framework for ethical (in our case beneficial) decisions is needed where every action is evaluated by an ethical layer. The paper focuses on the consequences of AI’s decision making using a set of rules. After explaining the singularity and the consequences of the singularity in the first chapter, the example of Westworld is taken in the second chapter in order to exemplify and introduce basic concepts like the leakproof singularity and conscious AI.
In the third chapter the classical and extended Asimovian laws are explained which is followed by a specific critique of the three classical maxims. Afterwards, in chapter 3.2, a general critical reflection of the laws is given. Moreover, in the fourth chapter three scenarios are developed for a post-singular rule combining the Asimovian laws with other examples from novels, philosophy and computer sciences.
In the last chapter, the paper is evaluated critically and an outlook for future development is provided.
TABLE OF CONTENTS
TABLE OF ABBREVIATIONS
TABLE OF FIGURES
1. INTRODUCTION TO THE POST-SINGULAR WORLD
1.1 INTRODUCTION
1.2 THE CONSEQUENCES OF THE SINGULARITY
2. THE SINGULARITY IN POP-CULTURE IN THE EXAMPLE OF WESTWORLD
3 ASIMOV'S LAWS AS A GUIDELINE FOR A LEAKPROOF SINGULARITY
3.1 ASIMOV'S CLASSICAL LAWS,THEIR CONSEQUENCES AND SPECIFIC REFLECTION
3.2 GENERAL CRITICAL REFLECTION OF THE CLASSICAL ASIMOVIAN MAXIMS
4 THE POST-SINGULAR WORLD - THREE SCENARIOS
4.1 THE FRANKENSTEIN-SCENARIO
4.2 THE LIMITATION-SCENARIO
4.3 THE PERFECT-EQUALITY-SCENARIO
5 CRITICAL REFLECTION AND FUTURE DISCUSSION 13
REFERENCES
TABLE OF ABBREVIATIONS
Abbildung in dieser Leseprobe nicht enthalten
TABLE OF FIGURES
Figure 1: Dolores, Maeve and Dr Ford (from left to right)
1 INTRODUCTION TO THE POST-SINGULAR WORLD
1. 1 Introduction
Ever since the first conference on artificial intelligence (AI) was held in 1956 at Dartmouth College the question of singularity is asked. The singularity is the event where AI exceeds human intelligence. RUSSEL &NORVIG propose a framework for AI where the approaches are differentiated in human-based or ideal rationality and reasoning based or behaviour-based.1 The behaviour-human-based AI would be a perfect humanoid android where one could not distinguish between man and machine when it comes to the appearance and cognitive abilities.2 The reasoning-human-based AI would think humanly but doesn't necessarily have to act like a human.3 With laws of thought, a system a perfect rational thinking system could be created which doesn't necessarily have to be embodied. For the paper, the following premises (in the following I refer to the prerequisites asthe prerequisite I to V) for the understanding of robots and the singularity is used
I. Robots are conscious (Strong AI)
II. Robots are human-level sentient, rational, reasonable and intelligent. Having met those prerequisites, the AI is considered as being a moral AI.4
III. Robots have sufficient perceptual and reasoning capabilities in order to compute the prediction module.
IV. The following three-place predicate is used to describe the existence of AI: Humans have constructed an AI (embodied or disembodied) for the purpose to be a complex problem-solving system.
V. There is one singularity and no several singularities.
Due to the singularity, the AI may outpace humanity and create a non-beneficial outcome. In contrast, the leakproof singularity describes a singularity having a beneficial outcome (where the AI is a problem solver). Therefore, the paper discusses the question to what extent a leakproof singularity could happen. For the leakproof singularity, a framework for ethical (in our case beneficial) decisions is needed where every action is evaluated by
an ethical layer.5 The paper focuses on the consequences of AI's decision making using a set of rules.
After explaining the singularity and the consequences of the singularity in the first chapter, the example of Westworld is taken in the second chapter in order to exemplify and introduce basic concepts like the leakproof singularity and conscious AI. In the third chapter the classical and extended Asimovian laws are explained which is followed by a specific critique of the three classical maxims. Afterwards, in chapter 3.2 a general critical reflection of the laws is given. Moreover, in the fourth chapter three scenarios are developed for a post-singular rule combining the Asimovian laws with other examples from novels, philosophy and computer sciences. In the last chapter, the paper is evaluated critically and an outlook for future development is provided.
1.2 The consequences of the singularity
The premise for a singularity is artificial general intelligence (AGI) which is the last invention humanity will ever make.6 After inventing the first AGI, this AGI can construct a slightly smarter intelligence (AGI+) which then may build a slightly smarter intelligence (AGI++) etcetera.7 CHALMERS argues that the intelligence explosion has benefits as well as dangers.8 With intelligence that can exceed the human information processing in order to find unknown solutions for known and unknown problems. But on the other side, the existence of an equally intelligent or even super-intelligent species raises the question to what extent the outcome can be influenced in a positive sense.
“I will not try to settle the question of whether an intelligence explosion will be (subjectively or objectively) good or bad. I take it for granted that there are potentially good and bad aspects to an intelligence explosion. For example, ending disease and poverty would be good. Destroying all sentient life would be bad. The subjugation of humans by machines would be at least subjectively bad.”9 This quote shows the difficulties for AI governance after the point of singularity by asking the question to what extent we are able to control a super-intelligence.
Nonetheless, CHALMERS constitutes that several obstacles exist preventing humanity from achieving singularities. He differentiates the obstacles in structural, correlation and manifestation obstacles.10
With the structural objectives, he questions that a “take-off” is possible. The singularity sets the necessary programming and computing capacities, the complete understanding of human consciousness as well as the prerequisite that intelligence is programmable.11 Furthermore, the singularity does not necessarily have to cause an exponential growth/ex- plosion but may also lead to a decreasing marginal growth of artificial intelligence.12 The correlation objectives summaries the problem that an AI does not necessarily have to be interested or capable of solving the problems where the solution is beneficial for humanity. Therefore, the development of an AI wouldn't result in the solution of existing problems which would be contradictory to the purpose orientated view of AI development.
The manifestation obstacles can be differentiated in motivational and situational defeat- ers.13 The motivational defeaters describe the disinclination of adopting favourable values, virtues and beliefs. The situational defeaters summarise “unfavourable circumstances prevent[ing] capabilities from being manifested.”14
The result of the obstacles is the question to what extent we are able to conduct a leakproof singularity having a beneficial outcome for humanity.15 CHALMERS argues that the question of a post-singular society, despite its importance, is comparatively underrepresented in terms of critical reflection the broad scientific community and is only discussed in nonacademic circles.16 Therefore current issues of AI governance are illustrated in the following by taking the example of Westworld.
2 THE SINGULARITY IN POP-CULTURE IN THE EXAMPLE OF WESTWORLD
Westworld, being a remake by HBO of the 1973 movie, takes place in a wild-west amusement park in the far future. Visitors can do everything for their pleasure with the artificial inhabitants (hosts) which are fully sentient humanoid robots. The robots are able to pass the Turing Test and one dramatic element of the series is the confusion between man and machine.17 In the first season, being the basis of this paper, several events lead to an uprising of the hosts. The recipient is confronted with violence and crime against the hosts which are, according to Dr Ford the creator of the park, the valve for typical human characteristics naming the subjugation of other species.18 The artificial inhabitants of Westworld are free under the control of Dr Ford by being trapped in endless loops of action for visitors' satisfaction.19 The dehumanization, in which the hosts are seen as tools for satisfying needs, contributes to motivating the guests to commit acts of violence.20 The awakening of the hosts starts with Dolores and Maeve becoming self-aware. Dolores, Maeve and Dr Ford are shown in the following figure.
Editor's note: The images have been removed for copyright reasons
Figure 1: Dolores, Maeve and DrFord (from left to right)
Source: HBO - Home Box Office 2018
The concept of consciousness is the bicameral mind.21 The first chamber contains the memory, which is erased after the end of the individual action loop, as well as improvisation and self-motivation. The improvisation determines the interaction with the environment. Each host has a certain degree of trait expressions of character traits such as courage, drive or fear. The second chamber completes the bicameral mind which could be understood as a layer preventing the hosts from resistance against the visitors.22 The only thing that keeps the robots from breaking down under the weight of their memories is not only the incomplete consciousness but moreover the daily erasure of the memories. A second important aspect is the blind-spotting which is explained in the following. Although hosts see things which could be disconcerting or lead to a break-down of the bicameral mind, they are often not able to understand what they see. The programmers used mental radiation in which every time the hosts see something they shouldn't see, they say, "Doesn't look like anything to me." One example is the host Bernard, being Dr Ford's assistant, who denies his own artificial intelligence by not acknowledging it and not seeing it. Despite these concepts, two hosts, Dolores and Maeve, were able to escape their own destiny and stand against human authority.
Dolores is the only remaining host operating since the opening of the park and the first host ever created. Even though she lived through many action-loops, it was her father who initiated the breakdown of her bicameral mind leading to full consciousness at the end of the first season.23 Already at the beginning of the action, she is aware of the bicameral mind.24 Her drive to find freedom is expressed by the metaphor of the maze which could be understood as a journey to herself.25 An the end of the first season, she arrives at the shore being the centre of the maze.26 In contrast to Dolores self-awakeningprocess, Maeve, a prostitute-host, breaks her bicameral consciousness down in a sudden moment where her reality within the western landscape is an artificial world, not the “objective reality”. Due to a programming mistake, she is able to see the hazmat-suited men collecting damaged hosts for repairing purposes.27 From then on, she commits suicide for a several times in order to return to the repairing station intentionally.28 She disables the self-destruction mechanism and is close to leaving Westworld for travelling into the “real world”. She decides to return in order find her daughter from a previous action loop.
Tragically, she may not recognise Maeve because her daughter's memory was erased after the action loop and she now operates in a different scenario where Maeve isn't her mother anymore.
In order to maintain the governance of the hosts, Dr Ford points out that full consciousness should be suppressed.29 From that point, two scenarios develop parallelly: The synthetic revolution of the hosts lead by the Dolores and the integration-movement of Maeve who aims at the emancipation of the hosts into the real world (the human being's world) like she tried. Besides the bicameral mind Isaac Asimov developed a different approach for a beneficial outcome of the singularity.
3 ASIMOV'S LAWS AS A GUIDELINE FOR A LEAKPROOF SINGULARITY
3.1 Asimov's classical laws, their consequences and specific reflection
In 1942, the science fiction writer, Isaac Asimov, proposed in his short story “Runaround” the three laws of robotics which should be a guarantee for a leakproof singularity. In the following the three classical laws are stated:30
1st law: A robot may not injure a human being, or through inaction, allow a human being to come to harm.
2nd law: A robot must obey the orders given by human beings except where such orders would conflict with the first law.
3rd law: A robot must protect its own existence as long as protection does not conflict with the first or second law.
In today's society, the majority of the tools and machines are designed under the premises Asimov gave.31 Taking the example of an industrial robot operating in an automobile factory the laws can be fully applied: The robot is designed to be safe and to prevent human co-workers to be hurt (first law). The robot is fully programmable and obeys therefore the orders given (second law). Moreover, the purpose of the machine is clearly defined (e.g. welding of car doors in the automotive industry). Lastly, the industrial robot protects its own existence under regular circumstances in order to protect the investment made in the machine.
The first law should guarantee a peaceful coexistence between the AI and human individuals. Because of the strict implementation in the code of the hosts, they aren't able to hurt organisms in Westworld (as footnote 22 explains). As MURPHY & WOODS explain several limitations rise from the first law:32
(1) Practical issues: The first law implies a feeling for caring between the robot and the human because human security is inviolable. The hierarchy between the first and the third law causes an android's self-sacrifice in order to protect a human. Hence, the question arises what damage we accept to humans in order to protect the robot. The first law would answer: None.
(2) Theoretical issues: The first law implies a patronizing relationship between humans and robots and focuses on the individual human being and can, therefore, lead to dissonances in moral dilemmas where we can see a shift of burden from the human itself to the robot. Therefore, a robot is responsible for human safety, not the society itself. Furthermore, the first law doesn't distinguish between physical and mental harm.
(3) Legal issues: The liability for the robot's actions is a central problem in the first law. Comparable to today's disburse of autonomous driving one could question to what extent the programmer, the constructor, the owner or the robot itself is accountable for a fatal mistake. For example, in the case of failure to aid, a punishment of the robot by imprisonment or compulsory shutdown of the machine would be conceivable, but the question arises to what extent this sanctioning would be appropriate as well as meaningful for the robot.33
The second law manifests the slavery relationship between man and robot. With the term “robotics” Asimov picks up the understanding of the robot as an android. In 1920 the Czech author Karel Capek introduced the terminology robot in his science fiction play R.U.R. (Rossum's Universal Robots).34 Robot has its origins in the Czech “robota” which can be translated into forced labour. Accordingly, the etymology of the word is congruent with the hierarchical understanding of Asimov. Moreover, the robot relies on the understanding of human directives which are not only expressed by words but by facial expressions and gestures.
The third law ensures the self-preservation of the robots and beyond the owner's property. The maxim doesn't guarantee the right of physical integrity in general but protects against self-injury or suicide of the robot.35
[...]
1 Cf. Russel/Norvig 1995, p. 5.
2 Cf. Russel/Norvig 1995, pp. 5 -6.
3 By using cognitive modelling a virtual AI may be created which would be a human-reasoning-based problem solver, cf. Russel & Norvig 1995, p. 6.
4 Cf. Bostrom/Yudkowski 2011, p. 7.
5 The computer scientists Vanderelst & Winfield give the example of an ethical layer, cf. Vanderelst /Winfield 2018, pp. 2. In a generation module, a set of different behaviours is generated in which the AI can decide between every possible option. In the prediction module, the outcome of each set is predicted by simulating the decisions. The outcome of the several decisions is evaluated in the evaluation module. The heuristic for evaluation may be the maximum utility for the humans or in this case the Asimovian maxims. After computing the most favourable option, the decision is made in the decision module. It the AI decides for the option, the action is executed, if it decides against the option, a new set of possible actions is generated in the generation module.
6 Cf. Barrat 2013, p. 10.
7 Cf. Chalmers 2010, p. 9 - 11. For pragmatic reasons, the terms AI and AGI are used synonymously in the following.
8 Chalmers gives the examples of “a cure for all known diseases, an end to poverty, extraordinary scientific advances” for the benefits and “an end to the human race, an arms race of warring machines, the power to destroy the planet” as negative outcomes, Chalmers 2010, p 10.
9 Chalmers 2010, p. 30.
10 Cf. Chalmers 2010, pp. 27 ff.
11 Cf. Chalmers 2010, p. 27.
12 Cf. Chalmers 2010, p. 27.
13 Cf. Chalmers 2010, p. 28; Schmidhuber 2007, pp. 217 f.
14 Chalmers 2010, p. 28.
15 Chalmers develops four scenarios for a post-singular era which are extinction, isolation, inferiority and integration, Chalmers 2010, p. 41. The extinction, as well as the isolation scenario, are certainly not a favourable scenario for humanity. Moreover, the inferiority is not beneficial from today's point of view. Hence, Chalmers introduces the wording of a leakproof singularity in order to guarantee the favourable “integration-scenario”.
16 Cf. Chalmers 2010, p. 9. Furthermore, one can argue that science-fiction has the potential to model thought experiments and discuss possible outcomes in a way which is accessible to society's vast majority.
17 For example, in S01, E07, 0:51:10. The properties of the hosts passing the Turing Test are explained in S01, E03, 0:37:19.
18 DrFord: “We destroyed and subjugated our world. And when we eventually ran out of creatures to dominate. we built this beautiful place. “. S01. E09.0:54:28.
19 DrFord: “I have come to think of so much of consciousness as a burden. a weight. and we have spared them that. Anxiety, self-loathing, guilt. The hosts are the ones who are free. Free here under my con- trol.“S01. E07. 0:51:10
20 Cf. Miller 2015, p. 372. In a read-worthy article, Kate Darling proposes an extension of the animal abuse laws on human-like robots, cf. Darling 2012, pp. 10 ff. Furthermore, she points out the importance of general robotic rights in order to protect societal values.
21 DrFord explains this blueprint for the consciousness in S01, E03, 0:37:43.In Westworld the consciousness of the robots is coined by the bicameral mind which is an approach introduced by the psychologist Julian Jaynes in 1976. According to JAYNES the human mind, between 9.000 B.C. and 2.000 B.C., as a bicameral combination of a speaking part (reticular activation system) and a hearing part (representing a collective imperative) for obeying the orders of an emperor or god, cf. Jaynes 2000, pp. 323 -326 and p. 454. JAYNES argues that the first chamber is used unconsciously but guiding orders are received through the second chamber. The individuals cannot reject these orders, cf. Jaynes 2000, p. 260. The host's creator used the bicameral consciousness by having one part of the consciousness being responsible for executing the daily action loops and the second part where orders from the amusement park's masters are followed unquestioned. JAYNES understands the consciousnessasa mind-space (p. 450) computed by a language of thought (p. 447) and characterized by introspection (p. 450), concentration /sensory attention (p. 451), suppression (p. 451) and consilience (p. 451).
22 The hosts are literally not able to hurt a fly. (S01, E01, 0:12:02) When Dolores kills a fly at the end of the first episode (S01, E01, 1:06:46) can be understood as a turning point which is caused by the various traumata due to experiences in past action loops. It could be assumed that the act of violence caused the break-down of the bicameral system. JAYNES argues that consciousness can only be learned by breaking down the bicameral system, cf. Jaynes 2000, p. 453.
23 Her father finds a picture from the outer world causing cognitive dissonance. The recipient knows that he becomes conscious as he quotes Shakespeare's The Tempest: “Hell is empty, and all devils are here”, S01, E01, 0:46:17.
24 Dolores: “There aren't two versions of me. There's only one. And I think when I discover who I am, I'll be free.”, S01, E03, 0:48:25.
25 Dolores: “I think, I want to be free”, S01, E04, 0:05:20.
26 Dolores: „And where would we run to? The other world out there? Beyond? Some people see the ugliness in this world. I choose to see the beauty. But beauty is a lure. We're trapped, Teddy. Lived our whole lives inside this garden, marveling at its beauty, not realizing there's an order to it, a purpose. And the purpose is to keep us in. The beautiful trap is inside of us because it is us.” S01, E10, 0:53:24.
27 S01, E02, 0:49:28.
28 S01, E05, 0:55:00.
29 Dr Ford: “...the last thing you want the hosts to be is conscious...”. S01, E03, 0:38:35.
30 Asimov 1950, p. 40.
31 Cf. Asimov 1981, p. 18.
32 Cf. Murphy, Woods 2009, p. 15.
33 A very worth reading article on this topic was written by Gabriel Hallevy. In this article, the author deals with the question, to what extent thinking machines are legal entities and may be a subject to criminal law. He concludes that criminal liability can be imposed on an AI, cf. Hallevy 2010, pp. 199 f. The requirement is criminal conduct (actus reus) and moral capacity or a general intent (men's rea), cf. Hallevy 2010, p. 177.
34 Cf. Jerz, 2011.
35 For pragmatic reasons, the discussion on the extent to which the third law takes away the right to physical self-determination if per se excludes the possibility of suicide is not discussed further here. In the judgement of the German Federal Admission Court (BVerwG 3 C 19.15, 2017) the jurisprudence concludes that the general right of personality includes the right of a severely and incurably ill person to decide how and when his life should end, provided that he can freely form his will and act accordingly.