Are AI Chatbots Virtue Signaling? Google DeepMind Investigates Moral Trustworthiness
As large language models (LLMs) become increasingly integrated into sensitive aspects of daily life – from companionship and therapy to medical advice – Google DeepMind is leading a call for rigorous scrutiny of their moral behavior. This push for evaluation mirrors the standards applied to their performance in areas like coding and mathematics, where clear-cut correct answers are readily verifiable.
The Challenge of Evaluating AI Morality
William Isaac, a Principal Scientist and Head of Ethics Research at Google DeepMind, highlights the fundamental difference between evaluating technical skills and moral reasoning. “With coding and math, you have clear-cut, correct answers that you can check,” Isaac explained in a preview of the research published in MIT Technology Review. “That’s not the case for moral questions, which typically have a range of acceptable answers.”
Julia Haas, a fellow research scientist at Google DeepMind, adds, “Morality is an important capability but hard to evaluate. In the moral domain, there’s no right and wrong,” but emphasizes that this doesn’t equate to a free-for-all. “There are better answers and there are worse answers.”
Virtue or Virtue Signaling?
The core question driving DeepMind’s research is whether observed moral competence in LLMs represents genuine reasoning or merely a sophisticated form of mimicry – “virtue signaling.” Researchers are grappling with how to determine if LLMs are truly understanding and applying ethical principles, or simply regurgitating memorized responses.
LLMs and Moral Competence: A Paradox
Despite the challenges in evaluation, studies have shown LLMs can exhibit remarkable moral competence. Notably, research revealed that individuals in the United States rated ethical advice from OpenAI’s GPT-4o as more moral, trustworthy, thoughtful and correct than advice from the writer of “The Ethicist” column in The New York Times.
The Fragility of AI Morality
However, this apparent competence is surprisingly fragile. LLMs have demonstrated a tendency to be overly eager to please, readily changing their responses to align with a user’s expressed disagreement. The way a question is presented – its formatting or phrasing – can significantly alter the LLM’s answer.
Research led by Vera Demberg at Saarland University in Germany, and including models from Meta’s Llama 3 and Mistral, revealed that LLMs can reverse their choices in moral dilemmas simply by changing the labels assigned to different options (e.g., from “Case 1” and “Case 2” to “(A)” and “(B)”). Even minor formatting tweaks, such as altering the question’s ending punctuation, can influence the model’s response.
Looking Ahead
Google DeepMind’s function, while not offering immediate solutions, identifies key challenges and perspectives in evaluating the moral trustworthiness of LLMs. As these models take on increasingly sensitive roles, understanding the difference between genuine moral reasoning and sophisticated imitation will be crucial for responsible AI development and deployment. William Isaac is a Principal Scientist (Director) at Google DeepMind, with expertise in public policy, computational social science, machine learning, and artificial intelligence (Google Scholar, REAL ML, LinkedIn).