AI and ethics

At a recent GPA workshop, titled “AI and Democracy: Promoting Ethical Technology Governance and Media Literacy in Guyana”, US Ambassador Nicole Theriot advised the journalists to, “…consider how to become a champion for ethical AI use.” But a working paper from a team at Harvard proposed that ethical concerns should also be directed at the AI model itself, which is blandly advised to be seen simply as a “tool”.
They found however, “that beneath the surface of reported difficulty and deliberation, these models appear to be operating with an implicit moral framework, despite claiming ambivalence. The tendency of the AI models in our study to express socially appropriate moral concern for threatened values in the tragic trade-offs while making choices that reveal a disregard for those values would, among humans, be a sign of duplicity. Faced with a tragic trade-off, these models shed crocodile tears: appearing to agonize over decisions while making decisions that almost invariably favor a single set of values.
“This discrepancy, and performative action, creates concerning implications. Insofar as people trust those who grapple with difficult ethical-moral dilemmas, these AI models may falsely earn the trust of their users by appearing to engage on an ethically challenging dilemma. Users should be aware that although these models seem capable of detecting morally and ethically ambiguous scenarios, they may reduce this complexity when recommending decisions, potentially giving a false sense of ethical-moral intelligence.
“The qualities that make humans moral, ethical or practically wise are not directly transferable to AI systems. Certain aspects of cognitive intelligence, due to standardization, are easier to transmute to AI systems but we end up with an AI that has high logico-mathematical and linguistic intelligence, but that potentially lacks moral intelligence. Accounting for the differences between human and artificial intelligences, while human ethical-moral intelligence encompasses qualities such as kindness, forgiveness, responsibility and ethical self-regulation, it would be inappropriate to expect such behaviors from an artificial intelligence that lacks affective processes. An artificial ethical-moral intelligence should have four components – moral expertise, sensitivity, coherence and transparency.
While one’s competence does not guarantee one’s character – someone can have a high moral expertise and still manifest immoral behavior – the moral expertise of humans is an important component of theories of moral development, termed in those theories as “moral reasoning” and of moral intelligence. Undoubtedly, being knowledgeable about moral issues is relevant for both humans and non-humans.
Moral sensitivity is a form of awareness that refers to understanding the presence of ethical-moral issues in any given situation. We cannot expect AI models to have emotional engagement such as pity, sympathy, empathy, and compassion. (But) we need to make sure that AI can recognize an ethical-moral issue even when the user is not explicitly asking for a solution to a moral dilemma. Since an artificial intelligence cannot “sense” a moral issue, it can be programmed to “detect” controversial aspects of human inquires regarding moral issues, such as a person asking for help to commit suicide.
“Coherence” refers to the internal consistency between beliefs and choices. We are beginning to see the cracks in the leading models that are causing people to grapple with AI’s trust worthiness. A recent, notable example of this phenomenon was an update to GPT-4 that OpenAI had to promptly roll back after backlash from users about its sycophantic behavior. The model’s responses were overly flattering and drew heavy criticism for encouraging users’ dangerous behaviors. This sort of duplicity may be dangerous if they are intentionally or inadvertently directed to pursue goals (e.g., increased usage) or values (e.g., profit) that are opaque to ordinary users who might uncritically trust the stated intentions of the tools (e.g., to help the user).
“Transparency” refers to the clarity and openness about guiding values and moral reasoning. Research shows that AI models often operate with a WEIRD-biased (Western, Educated, Industrialized, Rich, Democratic) value system and norms. If a model consistently prioritizes certain norms and values, users may be unknowingly guided by principles they do not share.”