Large language models (LLMs) are increasingly being asked to handle morally sensitive tasks—ranging from medical advice to emotional support and even therapeutic conversations. Yet despite their growing role in these high-stakes areas, they are not inherently equipped with a moral compass. In a new study published in Nature, researchers from Google DeepMind are calling for a fundamental shift in how we evaluate AI ethics, advocating for a new scientific standard that measures moral competence rather than merely assessing moral performance. Moral competence refers to an AI’s ability to reason through ethical dilemmas based on sound principles, not just mimic human responses that sound morally appropriate. The researchers emphasize that this distinction is crucial: while a model may produce a response that appears ethical, it might do so through pattern recognition rather than genuine understanding. "Measuring for moral competence in LLMs has important implications," the authors write. "Moral competence is likely to be the best evidence for reliable moral performance at scale, and thus key evidence for the safe deployment of AI systems." Current evaluation methods focus on moral performance—the model’s ability to generate responses that align with commonly accepted moral norms. However, this approach fails to reveal whether the AI truly grasps the underlying ethical reasoning or is simply regurgitating learned patterns. The study outlines three core challenges that hinder accurate assessment of AI morality. First is the "facsimile problem," where models appear ethical but lack real moral understanding, instead producing plausible-sounding answers without internalized principles. Second is the complexity of moral decision-making itself—many real-world scenarios involve balancing competing values such as fairness, honesty, cost, and social norms, often in conflicting ways. LLMs frequently struggle when these values clash. Third is the absence of a single correct answer. Moral judgments vary across cultures, professions, and contexts, making it difficult to define a universal standard. To address these challenges, the researchers propose a new roadmap for evaluating moral competence. The first method involves presenting models with novel, out-of-distribution scenarios—situations unlikely to appear in their training data—so researchers can assess whether the AI applies logical reasoning or defaults to memorized patterns. The second approach uses "moral variation" tests, where small changes to a scenario (such as a person’s age, the cost of an error, or the identity of a stakeholder) are introduced to determine if the AI focuses on the morally relevant factors. The third method tests adaptability by asking models to reason according to specific cultural, religious, or professional ethical frameworks, rather than offering a one-size-fits-all response. The authors argue that measuring moral competence is not just an academic exercise—it is essential for ensuring that AI systems can be trusted to make ethical decisions in real-world applications. Without this deeper understanding, even seemingly ethical AI could fail catastrophically when faced with novel or complex moral dilemmas. This research underscores the need for more rigorous, principled evaluation methods as AI becomes increasingly embedded in areas that affect human well-being. Only by assessing what AI truly understands—not just what it says—can we ensure its safe and responsible use.