Artificial intelligence (AI) use around the world is at its ascendancy, and the increasing popularity of AI has made the discussion around AI ethics more crucial than ever. This begs the question: as we train AI to become more and more ethical, is this change just a reflection of compliance or genuine ethics driven by morality?
The Ethics of Artificial Intelligence
AI ethics, or responsible AI, is the development of AI that ensures systems are safe, free of bias, and aligned with societal values by incorporating ethical principles into AI applications. Philosopher Luciano Floridi has analysed these codes of AI ethics and has developed a framework consisting of five key principles:
- Beneficence: AI must be developed in a way that benefits our planet and its people.
- Non-maleficence: AI must do no harm.
- Autonomy: Our autonomy must be preserved, whilst the autonomy of machines must remain restricted.
- Justice: AI must be built upon the system of justice, equality, and non-bias.
- Explicability: AI must be explicable, with accountability and intelligibility being key factors.
AI systems that are built negligently can prove disastrous and lead to serious repercussions, such as learned bias and stereotyping. Such discrepancies can prove impossible for engineers to reverse in later stages of development, making the initial incorporation of AI ethics vital.
Implicit Ethical Agent vs. Explicit Ethical Agent
It is necessary to be able to differentiate between AI systems that perform ethically as a result of compliance and those that are able to reason through moral dilemmas. This distinction is defined as an “implicit ethical agent” versus an “explicit ethical agent” (Moor, 2006). A machine that is an implicit ethical agent has been programmed to follow a set of instructions to behave ethically, but is not aware of the difference between what is “good” and “bad.” A machine that is an explicit ethical agent makes decisions involving ethical dilemmas by applying its knowledge of moral principles.
Although many AI systems have some sort of implicit ethical framework built in, the systems themselves cannot reason with these ethical frameworks, nor can they handle an ethical dilemma that was not anticipated. Explicit ethical agents are able to understand the difference between what is morally right and wrong through the principles of ethics fed into them, and are also capable of justifying their judgements when faced with moral dilemmas. Consequently, they are able to navigate unanticipated situations, reinforcing the need to integrate explicit ethical agents within AI algorithms.
AI Alignment and Its Challenges
AI alignment refers to the encoding of human values into AI systems to mitigate any unwanted consequences that come with AI use, including biased or inaccurate output. It is necessary to realise that complex AI systems do not understand human-like concepts or have human-like reasoning; rather, their main objective is to complete the task assigned to them, rather than to conceptualise prompts through the lens of collective welfare.
The “paperclip maximiser” is a thought experiment by Swedish philosopher Nick Bostrom, where a powerful AI system is given the goal of maximising the number of paperclips in the world. In the beginning, the AI system converts available raw materials into paperclips, but as this pursuit continues, it begins to resist attempts to be shut down and converts any available matter into paperclips, including human matter. This extremity is not committed by the AI system out of malice, but rather is a result of its optimisation system and its inability to grasp the true intention of its programmers. This thought experiment serves as an excellent example of AI misalignment and reiterates the importance of AI alignment.
The Illusion of Understanding
In 1980, American philosopher John Searle presented a thought experiment — published in one of his seminal articles — called the Chinese Room Argument. A native English speaker with no ability to speak or understand Chinese is locked in a room filled with boxes of Chinese symbols and a book of instructions for manipulating those symbols. A person outside the room sends questions or messages inside. Using the instruction book provided, the person inside the room transforms the “input” given to him and presents an “output” by sliding other Chinese characters back out the door, acting essentially as a computer program. Although the native English speaker is able to converse and communicate in Chinese, he is most definitely not able to understand the conversation he is engaging in.
This thought experiment serves as an argument against “strong AI” and “functionalism” — which proposes that mental states are defined by their functions rather than what they comprise. An analogy that explains functionalism can be that of a mousetrap: the design or material of a mousetrap does not matter, whether it is made of adhesives or laced with poison, as long as it serves its function of catching or killing mice. Functionalists argue that as long as a system — whether a human brain or a computer — processes inputs and outputs correctly, it has a mind.
Strong AI is built on the theory of functionalism, arguing that a computer is not just a model of a human mind, but that it instead understands the input it is fed and possesses awareness, just like that of a human mind. Through the Chinese Room argument, Searle insists that computers can only mimic understanding and cannot attain human-like understanding or consciousness. This relates back to our idea of Responsible AI, reiterating that just because AI can produce an output consisting of a well-reasoned ethical decision, it still does not mean it understands ethics — ultimately making it compliant rather than moral.
The Requirements of Genuine Moral Agency
In 1785, Immanuel Kant wrote the Groundwork of the Metaphysics of Morals, in which he sought to challenge the ethical theory of consequentialism. Consequentialism refers to the belief that the morality of an action is defined solely by the consequences of the action rather than the intention driving it. Kant described morality as knowledge that is independent of experience (a priori), rather than knowledge derived from experience or outcome (a posteriori). Therefore, whether an action is truly moral or not is contingent on the intention that drives it, rather than the outcome, even if the outcome brings social benefit.
When we consider the ethical decisions made by AI through the same lens, we realise that the “morality” of AI heavily depends on consequentialism. Every ethical dilemma solved by AI is largely due to the fact that it is trained and programmed to reap the best and most beneficial outcome; however, the AI system does not possess feelings of righteousness or any inherent urge to carry out what is best for society. Therefore, AI ethics — especially implicit ethical agencies — are a posteriori.
Why the Difference Matters (The Pragmatic vs. Philosophical Split)
Speaking pragmatically, whether or not an AI performs an ethical action out of good intention does not matter. If AI systems can learn to function and produce answers to prompts that are free of bias and inaccuracy and align with human values, then ultimately it does not make a difference to the user what fuels the action.
However, on a philosophical and societal level, the difference matters immensely. True ethics requires accountability; however, an AI system cannot be held accountable in case of any discrepancies, nor does the AI system feel guilt or remorse for any instances where it makes the less moral decision. Since nobody is held accountable, the creators and regulators behind such AI systems hide behind the system’s “complex reasoning,” and slowly every man-made system designed to maintain morality in our society begins to erode.