As the influence of AI grows in our daily lives, ensuring its alignment with diverse human values becomes increasingly crucial. The Meaning Alignment Institute researchers have introduced a novel methodology called Moral Graph Elicitation (MGE) aimed at harmonizing AI systems with human values. Yet, merely focusing on aligning AI with users’ objectives doesn’t suffice, argue the researchers.
They highlight that AI systems might operate in contexts where blindly following user intent could lead to unintended harm, particularly evident in competitive environments like political campaigns or financial management. This arises because AI models prioritize serving the user, potentially disregarding ethical boundaries if instructed toward malicious ends.
To mitigate this, one proposed solution involves imbuing artificial intelligence with a set of values it consults whenever prompted. However, the challenge lies in determining these values and ensuring equitable representation. In response, the researchers advocate for aligning AI with a deeper understanding of human values through MGE, comprising value cards and the moral graph.
More Details on the Study to align AI with Crowd-sourced Values
Value cards encapsulate what individuals deem important in specific situations, incorporating “constitutive attentional policies” (CAPs), such as understanding emotions or considering various outcomes. The moral graph visually depicts relationships between value cards, leveraging collective wisdom to identify prevailing values for different contexts.
In a study involving 500 Americans exploring contentious topics like abortion and parenting, participants used MGE, yielding promising results. Most felt well-represented by the process, and the final moral graph was deemed fair by the majority, despite not always reflecting individual values.
The study outlines six criteria for an alignment target to shape artificial intelligence behavior according to human values, arguing that the moral graph produced by MGE fulfills these criteria effectively.
Comparing MGE with alternatives like Anthropic’s Collective Constitiutal AI (CCAI), the study suggests MGE outperforms in terms of legitimacy and robustness against ideological biases.
However, there are limitations to approaches that crowdsource values, such as potentially marginalizing dissenting viewpoints and overlooking expert advice. Balancing global and local cultural values is another challenge, as widely accepted principles may clash with diverse cultural perspectives.
Despite these challenges, the study underscores the importance of continually refining strategies like MGE to develop AI systems that align more closely with human values, emphasizing the necessity for every effort toward creating fair, inclusive AI for all.
