Lawrence – Can artificial intelligence really demonstrate morality?
Currently, the field of AI ethics is debating whether these machines should be treated as moral agents capable of making ethical decisions.
“We're now asking, 'If these systems can act like humans who are moral agents, then perhaps these systems are moral agents,'” he said. Oluwaseun Damilola Sanwoolu, a PhD candidate in philosophy and a graduate research assistant at the Center for Cybersocial Dynamics (CCSD) at the University of Kansas.
“I think logic has expanded to this point because AI systems are sometimes connected to us as people we don't have.
Such questions are explained in a new paper entitled “Cantiandentology for AI: Alignment without Moral Institutions.” Although AI currently has no practical judgment, Sanwoolu shows that there is a functionally equivalent mechanism (transformer model) that can form a saying that takes into account morally prominent facts. Therefore, this supports the claim that AI alignment is possible within the Kantian Framework.
This study is published in Journal AI and Ethics.
Developed by the German philosopher Immanuel Kant (1724-1804), Cantian dentology is an ethical theory that focuses on moral obligations and principles rather than the outcome of actions, in order to determine whether it is right or wrong. It emphasizes the use of reason to identify universal moral law.
“In the beginning, I wasn't a Kantian scholar. I didn't even really like the guy. But when I started thinking about one of my research goals at CCSD, how technology can be ethically shaped, I realized I was reading more Kant.
Sanwoolu is tackling two major challenges to applying Kant's moral philosophy to AI. The first objection is that AI cannot meet Kant's standards regarding moral institutions. Second, Kant's theory does not explain the context when the principles of his theory are applied.
“From the standpoint of the initial objection, I admit AI systems are non-moral agents. But can this make us behave in a way that mimics human agents using the Cantian system without being moral agents themselves? “I think it's viable,” she said.
For comparison, she points out that when teaching children honest values, adults don't necessarily say, “This is the moral code we live in.” We simply model honest behavior towards them. They are not yet fully formed moral agents, but they can see what people are doing and imitate it.
“AI systems can act the same way,” she said. “They don't have to be moral agents.”
Second, Kant's theory is very strict and does not take into account the context.
“Kant would say something like, 'Don't break your promise.' And people argue that if you make a promise under obsession, Kant is not blind to the context.
Ultimately, she assumes that AI can act like someone who wants to achieve morality or perform morally correct actions.
“People want to avoid harming AI. What does that mean?” she said. “Is an AI system harmful or useful when helping people commit suicide? That's where ethical systems and ethical frameworks come into play. It tells you, 'Because we'll do this.' ”
Nigerian native Sanwulu is currently in his fourth year at KU. Her research focuses on AI's ethics and technology philosophy and is overseen by John Simmons, KU's Professor of Philosophy and Director of the Centre for Cybersocial Dynamics.
“Ethics gives AI systems context what they want to do,” she said. “Maybe there is no consensus on what the correct and correct ethical theory is, but I think this framework actually works for AI tuning.”
