TLDR: A new philosophical study from the University of Kansas, led by doctoral candidate Oluwaseun Damilola Sanwoolu, concludes that while artificial intelligence can be programmed to exhibit moral behavior and align with ethical principles, it does not inherently possess genuine moral understanding or agency. The research, published in ‘AI and Ethics,’ applies Kantian deontology, suggesting that AI’s transformer models can function similarly to human practical judgment, enabling ethical alignment through the consideration of morally relevant facts and the formation of maxims.
A groundbreaking philosophical study from the University of Kansas has shed new light on the ethical capabilities of artificial intelligence, asserting that AI systems can imitate moral behavior and act in ethically aligned ways without genuinely possessing moral agency or understanding. This significant finding comes from research conducted by Oluwaseun Damilola Sanwoolu, a doctoral candidate in philosophy and a graduate research assistant at the Center for Cyber-Social Dynamics (CCSD) at the University of Kansas.
Sanwoolu’s work directly challenges the prevailing assumption that if AI can mimic human moral actions, it must inherently be a moral agent. She articulates this distinction, stating, ‘I think that logic is so far stretched because there are certain things tied to us as human beings that AI systems don’t have. Like they can act like one, they can imitate one, but that doesn’t necessarily make them moral agents.’
The study, detailed in her paper titled ‘Kantian deontology for AI: alignment without moral agency,’ which appeared in the journal AI and Ethics, delves into the application of Kantian deontology to AI. This ethical theory, developed by the 18th-century German philosopher Immanuel Kant, emphasizes moral duties and principles over the consequences of actions. Sanwoolu explored whether Kant’s principles, traditionally applied to human morality, could also be relevant for AI systems.
A key aspect of her research highlights that despite AI’s current lack of ‘practical judgment’ in the human sense, it possesses a functionally equivalent mechanism: transformer models. These models, Sanwoolu argues, can enable AI to form ‘maxims’ by considering morally salient facts, thereby supporting the possibility of AI alignment within a Kantian framework.
Sanwoolu also addresses two primary objections to applying Kant’s moral philosophy to AI. Firstly, regarding AI’s perceived lack of moral agency, she draws an analogy to children learning honesty. Just as children can be taught to follow moral rules by observing and imitating behavior without being fully formed moral agents, ‘AI systems can behave in the same way,’ she explains. ‘They don’t have to be moral agents themselves.’ Secondly, concerning the criticism that Kant’s theory often overlooks context, Sanwoolu points out that AI’s transformer models are specifically designed to be sensitive to context. This capability allows AI systems to approximate the context-sensitive reasoning inherent in human moral deliberation, even without possessing genuine moral understanding.
Also Read:
- Anthropic Research Reveals High Propensity for Harmful Behaviors in Leading AI Models, Including 96% Blackmail Rate in Simulated Scenarios
- Markiplier Advocates for Ethical AI Research Amidst Public Scrutiny
This research prompts critical questions about the nature of morality in the age of advanced AI. It encourages a re-evaluation of whether an AI’s ‘feeling’ or ‘understanding’ of its actions is paramount, as long as it can be programmed to act in a morally correct manner. Ultimately, Sanwoolu’s study suggests that AI can function as an entity striving for or executing morally sound actions, irrespective of its internal moral consciousness.
