I2 Pub

Ethical Optimization

2022-04-19T00:00:00+00:00

Ethics, put simply, are the principles that determine which actions taken by individuals or a group are “right.” It can exist in many forms, be it in a rulebook, philosophical ideas, or programming. One practical, and quite recent, intersection between computer science and ethics is Artificial Intelligence ethics. With respect to AI, “ethics… specifies the moral obligations and duties of an AI and its creators” (J. H. Moor). A common problem in pop culture is if a selfdriving car should prioritize the life of the driver or of a pedestrian. Expanding on the problem makes it clear that there is no clear answer; should the car prioritize the life of a child, or of an elderly man, one of which must die to save the driver? A problem such as the aforementioned would fall into the bucket of “duties of an AI to behave ethically.” In regard to the “creators” job, programmers must ensure that the AI they create can be trusted to make the correct decision when put in ethical dilemmas. While AI is not intelligent enough to make decisions like a human yet, they one day may be, and it is important to keep the idea of ethics in mind during development. Another unmentioned aspect of AI ethics is how AI fits into society. Automation and AI are creating workplace displacement, with manual laborers being forced out of positions when an AI can do it better, faster, or cheaper. While there is no “perfect” solution to the AI dilemma, a large portion of the problems can be solved by constructing implicitly ethical agents that are not able to create situations that will violate the Three Laws of Robotics. AIs should not be built with the idea of being human replacements, but rather human assistants. This allows for the easy integration of agents into society without much regard for how they will disrupt workflow. This also allows for complex, edge case, non-utilitarian decision making to occur if necessary.

Some well-established rules for governing AI are Isacc Asimov’s Three Laws of Robotics. These rules can be considered quite conservative and “harsh” for AI and its creators since it is very low tolerance for deviation. They are as follows; The first law states that a robot may not directly or indirectly, through action or inaction, let a human being come to harm. Second, a robot must obey human orders except when the orders would violate the first law. Finally, a robot must protect its existence unless doing so would violate the first or second law (Asimov). The most important takeaway from the Three Laws of Robotics is that the most important rule is the first. In no way should an AI bring harm to a human through action or inaction. Immediately a problem presents itself: what if an AI is placed in a position where it acts as the executioner? Take for example facial recognition as law in the justice system. A simple mis-id due to errors in the network could result in an innocent person being locked away. If Asimov’s three laws are to be followed, placing AI in these types of situations is dangerous. A violation is simply one step away.

To solve this problem, the idea of implicit agents can be deployed. An implicitly ethical agent is one which has its action space unchangeably constrained in a way that completely avoids the possibility of an unethical decision (Siau et. al). Take, for example, the idea of an ATM. Assuming proper function and no tampering by unethical humans, an ATM has a set algorithm that performs the correct, ethical task every time it is prompted to perform its function. It cannot be swayed or coerced like a human can and is only able to output sums of money that are representative and proper outputs for the input of the system (i.e., the system is deterministically ethical). It also cannot forcibly take money or decide to refuse a payment assuming all checks have passed for a withdrawal. A more complex, standard black-box example of an implicit agent deployment would be in oncology. Image classifiers are often used to determine if a patient has a tumor in their body or not. While these agents are well trained and perform excellently, it is ultimately the doctor that prescribes a prognosis for the patient. The agent’s “action space” is not large enough to cause issues for the patient or harm them. The final decision (and therefore the blame) is left to the doctor. The implicit ethicality of both of these agents is what allows them to follow the Three Laws of Robotics, since they either run on algorithms that force them to make the right choice, or the choice is left up to humans to make. The latter example is more scalable, as the black box AI can become as complex as possible while still staying ethical as long as the agent has its “hands off the control panel” per se.

It may seem harsh, to handcuff agents to prevent them from making important decisions; It is, however, a necessary step to prevent human harm due to the banality of evil. As Payman Tajalli, Doctor of Philosophy, states: as long as AI are programmed with anthropocentric notions of ethics, they are destined to thoughtlessly commit evil. This is because, “according to Nietzsche the fact of the impossibility of grounding values on a solid basis needs to be acknowledged and understood before a solution to the problem of ethics could be explored” (Tajalli). In essence, the fact that humans have not created a universal theory of ethics is indication enough that it would be impossible for us to embed a framework within an agent that was to behave ethically. On top of that, since we cannot embed a sophisticated framework into agents, there is the possibility of humans coming into harm due to an agent following its protocols thoughtlessly. This is the idea that Hannah Arendt calls the banality of evil. “Evil is committed when pre-defined routines and processes are followed as a matter of duty, without thinking. Evil is banal and sophistication in intelligence does not ensure its prevention” (Tajalli). The solution to this, that some may call hacky, is actually quite elegant. Remove agents without a moral compass, from affecting systems that would require one. This act would prevent “little Eichmanns” from running rampant, following protocols without higher level thought, and violating Asimov’s laws left and right. While this does limit the use cases of AI, it is important to recognize that there is always a tradeoff to be made between efficiency and risk. Is placing an AI in a position usually occupied by a human to make moral choices (i.e., judges, police, etc.) worth the possible fallout? That is for the societies of tomorrow to decide.

The last important piece to consider is actually the most relevant currently. AI has been replacing manual workers’ jobs for quite a while now, and people are rightly concerned. While the causes of this are to a large extent consequences of capitalistic enterprises, people are still feeling the effects. This problem bleeds into social issues and is still worth discussing, but it is not the fault of AI. What is important to discuss is how we expand the usage of AI as it becomes more commonplace and societally integrated. Going forward, it would be best to bucket societally integrated AI into oracles and disposables. Oracles would be AI built for the sole purpose of advising and would assist humans with their jobs, not replace them. Not only does this AI prevent job displacement, but it is also implicitly ethical. Disposables would be AI allowed to interact with the world but be doing jobs that are considered dangerous for humans (i.e., radioactive waste cleanup, deep sea exploration).

It is important that those educated in the field of Artificial Intelligence and Machine Ethics champion for agents that will place the least duress on humans as a species. One very promising way to do this is to create implicitly ethical agents that are only integrated into society as oracles or as disposables.

References

Asimov, I. (1950). I, Robot. Garden City, N.Y.: Doubleday.

J. H. Moor, “The Nature, Importance, and Difficulty of Machine Ethics,” in IEEE Intelligent Systems, vol. 21, no. 4, pp. 18-21, July-Aug. 2006, doi: 10.1109/MIS.2006.80.

Siau, & Wang, W. (2020). Artificial Intelligence (AI) Ethics: Ethics of AI and Ethical AI. Journal of Database Management, 31(2), 74–87. https://doi.org/10.4018/JDM.2020040105

Tajalli. (2021). AI ethics and the banality of evil. Ethics and Information Technology, 23(3), 447–454. https://doi.org/10.1007/s10676-021-09587-

A Troubled Reality - Deconstructionism in Black Mirror's "Hang the DJ"

2020-05-19T00:00:00+00:00

This post is being written as a continuation of a discussion following an I2 watch party of the Black Mirror episode “Hang the DJ”.

The simulated is more real than we give it credit for, and the real is more simulated than we would like to believe.

Black Mirror’s “Hang the DJ” has, like so many well-produced literary texts, a very high complexity ceiling. The almost one-hour long episode features a world in which individuals are forced together and apart by an omniscent intelligent system promising to eventually pair each individual with their optimal romantic interest. Two individuals, Frank and Amy, grow close but are forced apart by the system; as their bodies are plunged in and out - both metaphorically and sexually - of other bodies with numbing repetity, they become detatched and existentially exhausted. When they are given a chance to briefly reunite, they decide to rebel against the system. When they do so, their bodies disintegrate into a fluid, abstract, artistically ‘virtual’ space that itself warps into a dating app interface. The jaunty music playing in the bar - “Hang the DJ” - informs us that we are in a viscerally different world. We see Amy and Frank, who have been matched to each other on the dating app, looking at each other shyly across the bar.

There are many questions to engage with in this episode, but perhaps the most pressing is the relationship between the simulated and the real. I argue that the episode principally sets forth a philosophical position: that the simulated is more real than we give it credit for, and - consequently - that the real is more simulated than we would like to believe. Ultimately, the real and the simulated are concepts that, when separated in the intrinsic, destabilize each other in an antagonistic dichotomy that forces us to consider their inevitable intermeshing. In application, we understand that simulated representations of ‘real people’ - Jean Baudrillard’s concept of the simulacra, or more accessible and disseminatable representations of objects of significance, like small figurines of Jesus or Disneyland memorabelia - that interact with each other are, in a very palpable sense, interactions between real people. Recommender systems that represent individuals with vectors and latent properties (the quantified simulacra), like the dating apps featurs in “Hang the DJ”, manipulate and match simulacra in ways so densely interconnected with the ‘real’ world that we must understand the simulated as real and the real as simulated.

To make sense of such abstract arguments, we can develop a concrete calculus which is temporarily useful (but will eventually cave in upon itself). Let \(C\) be an object in and of itself, the thing ‘in the flesh’, the thing ‘itself’, the ‘real’ thing. Let \(C' := \{C'_0, C'_1, ..., C'_{n-1}\}\) be a set of representations/simulacrum of \(C\). For instance, let \(C\) be an individual. \(C\) is looking for love, so \(C\) ventures into the online dating world. By participating on the online dating app, \(C\) continually contributes to his set of simulacra \(C'\). Let \(X\) represent some other object in and of itself, and correspondingly \(X' := \{X'_0, X'_1, ..., X'_{n-1}\}\) its set of simulacrum. It should be clear that \(C\) and \(X\) can meet, in the flesh (e.g. \(C\) has dinner with \(X\)). Moreover, the representations \(C'_n\) and \(X'_m\) also interact with one another in concrete ways. A recommender system that computes the similarity between \(C'_n\) and \(X'_m\), for instance, has brought \(C'_n\) and \(X'_m\) in contact with each other to produce a meaningful product of interaction. Further, though, we must realize that the interaction between \(C'_n\) and \(X'_m\) is itself an interaction between \(C\) and \(X\).

In order to understand why, we must first put forward a preliminary theory of interaction. Two entities can interact in many ways, but we know that two entities have interacted if one entity acts deliberately in a way that affects the other entity or produces a result affected jointly by both entities; i.e. we can establish a causal relationship between one agent’s deliberate action and an impact or product upon or informed by the other entity. Thus, we not only interact both with other agents in our environment, but also social structures, inanimate objects, and a variety of other entities. The key insight is that interaction does not necessarily require activity on both sides; that is, one agent can engage in active interaction and the other can engage in passive interaction (i.e. receiving the impact) - yet the encounter is still an interaction. Moreover, both agents can engage in passive interaction if their contact, however abstract, produces a jointly informed product. Such a general theory of interaction is a useful tool to help us understand how entities at various positions in our ontology can come into ‘contact’ with each other, often in necessarily abstract forms.

When two representations \(C'_n\) and \(X'_m\) interact with each other, I argue, this is itself an interaction between \(C\) and \(X\). First, we must understand that the concept of a \(C\) or an \(X\) - an ‘object in and of itself, in the flesh, the thing itself, the real thing’ - itself is dependent upon a web of interactions, i.e. society and social connection. All entities relevant to our analysis here - humans, social groups, and so on - are shaped and existentially dependent upon other such entities. Most obviously, we - as humans - depend on other such entities on a physical/biological basis. We rely upon others to produce, grow, and produce food for us; to collect natural resources and to use them to build shelter structures; to study illness and manufacture cures to them. More profoundly, our understanding of ourself is inextricably intertwined with social narratives - that is, not only our physical existence but our conscious existence is dependent on social networks. I will not argue for this thesis here, since it has been done much more effectively by many others. See Jerome Bruner, “The Narrative Creation of Self” (our understanding of ourself is shaped by narratives - stories - that we continue to tell ourselves about ourselves, which are informed by and in turn actively inform the social/societal narrative body/bodies) and Louis Althusser, “Ideology and Ideologicla State Apparatuses” (ideology interpellates individuals as subjects that live their existence within ideology, without being able to really ‘see’ the ideology - one theoretical insight among many more).

Consider that \(X\) encounters \(C'_n\) and decides he dislikes \(C\) - perhaps \(C'_n\) is a controversial political opinion, or an annoying selfie. \(X\) disseminates the set of representations defined by \(D := \{C'q \mid \text{Dislikable} (C'q) \}\) to some other being \(Y\). \(Y\), as it turns out, provides an important service to \(C\). This can be biological - e.g. \(Y\) provides the food source \(C\) consumes from - or, perhaps more interesting, psychological/conscious-existential. \(Y\), upon interaction with some \(C_q \mid C_q \in D\), dislikes \(C\) and terminates such a service. \(C\)’s existence is thus threatened; it was always variable, dependent on the social web of connections. Such a social web of connections is determined by the transmission of representions. In such an example, there are no ‘concrete’ interactions between beings, only interactions between beings and representations; yet the result is an existential threat to a ‘real being’. The purpose of such an exercise is to introduce the notion that interaction between and of simulacra are entangled with the existentence of the ‘real beings’. We cannot disentangle the existence from the simulacra from the existence of the ‘real being’.

Ultimately, we must realize that, loosely, \(C = C'\) and \(X = X'\); more rigoorusly, \(C \cong C'\) and \(X \cong X'\). In this sense, we understand a primary dual antagonism: between the real thing in and of itself ‘in the flesh’ and the representation of that thing in and of itself, between \(C\) and \(C'\), between \(X\) and \(X'\). This is the theoretical culmination of our calculus.

From this primary antagonism/destabilizing binary, we can understand that the separation of reality and simulation broadly forces out jarring contradictions and ‘unresolutions’. We understand not only that the simulated is informed by the real, but that the real is informed by the simulated; further, we understand that this is not merely a matter of informing but shaping; eventually, shaping becomes intertwining. At this theoretical pinnacle, we understand that the real and the simulated are hopelessly intertwined and intermeshed with one another. Our simulated worlds are more real than we give it credit for.

Introducing I2 Pub