Anthropic Explores AI Ethics with Philosophers: Why Character Formation Has Become a Core Issue in AI Alignment

A New Direction in AI Ethics Dialogue: What Anthropic Is Doing

Anthropic recently disclosed that over the past few months, the company has been engaging in a series of conversations with scholars, philosophers, clergy, and ethicists to explore the deep questions raised by AI—starting with the fundamental issue of "how good character is formed."

This initiative signals that frontier AI companies are now more systematically examining the ethical dimensions and societal impact of AI beyond technical development alone. For practitioners focused on AI safety and AI governance, this dialogue sends a noteworthy signal.

Why "Character Formation" Is a Core Issue in AI Alignment

From Technical Alignment to Value Alignment

In the field of AI safety, "alignment" has always been the central topic—how to ensure AI systems behave in accordance with human intentions and values. Alignment research originated from concerns that superintelligence might deviate from human intent, and has become more concrete and urgent with the rise of large language models (LLMs). Current mainstream technical approaches include Reinforcement Learning from Human Feedback (RLHF), Direct Preference Optimization (DPO), and Anthropic's own Constitutional AI. However, all these technical solutions face a common philosophical prerequisite question: Is human feedback itself sufficiently reliable? Are human preferences equivalent to genuine values?

Anthropic's entry point for this dialogue is deeply significant: rather than starting from a technical perspective, they chose a seemingly ancient yet profoundly fundamental philosophical question—"how is good character formed?"

This choice is no accident. When we attempt to make AI exhibit "good behavior," we first need to answer: What is "good"? Where does good character come from? These questions have been debated in human societies for thousands of years, from Aristotle's virtue ethics to Confucian self-cultivation, with vastly different answers from different cultural traditions.

In the Nicomachean Ethics, Aristotle proposed that virtue is not innate but developed through repeated practice and habit—"we become just by doing just acts." This view bears a striking structural resemblance to contemporary machine learning training paradigms: AI systems similarly form behavioral patterns through extensive "practice" (training data and feedback signals). However, virtue ethics also emphasizes "practical wisdom" (Phronesis)—the capacity to make appropriate judgments in specific situations—which is precisely the quality most difficult for current AI systems to acquire. The Confucian concept of "self-cultivation" (修身) similarly emphasizes the gradual nurturing of character, but places greater emphasis on social relationships and role-based responsibilities, creating a productive tension with Western individualistic notions of virtue and offering diverse reference points for AI value alignment across different cultural contexts.

For AI alignment research, this means that technical solutions must be grounded in clear value foundations. Without a deep understanding of "good," any alignment technique may only address the problem superficially.

Why Cross-Disciplinary Dialogue Is Indispensable

The participants Anthropic invited span scholars, philosophers, clergy, and ethicists—a cross-disciplinary composition reflecting a key recognition: AI ethics problems cannot be defined and solved solely by technologists.

• Philosophers provide theoretical frameworks for moral reasoning and value judgment
• Clergy represent the deep wisdom about good and evil, responsibility, and human nature embedded in different faith traditions
• Ethicists translate abstract principles into actionable behavioral guidelines
• Scholars contribute empirical research and critical perspectives from their respective fields

A single-discipline perspective easily creates blind spots, and the scope of AI systems' influence has already far exceeded the technical domain itself.

The Ethics Turn at Frontier AI Companies

Industry Trend: From Rapid Iteration to Deliberate Development

In recent years, as LLM capabilities have advanced rapidly, the ethical pressure facing frontier AI companies has continued to mount. Anthropic's approach can be understood within this broader industry context.

On one hand, AI system decisions are increasingly permeating daily life—from content recommendations to medical diagnosis, from legal consultation to educational tutoring. These application scenarios require AI to be not only technically reliable but also able to withstand scrutiny in value judgments.

On the other hand, global regulatory frameworks are rapidly taking shape. The EU AI Act officially came into effect in 2024, adopting a risk-tiered regulatory framework that classifies AI systems into unacceptable risk, high risk, limited risk, and minimal risk categories, imposing strict transparency, explainability, and human oversight requirements on high-risk AI. The United States issued an Executive Order on Safe, Secure, and Trustworthy Artificial Intelligence in 2023, requiring frontier AI developers to report safety test results to the government. China, the UK, Japan, and other countries are also accelerating the development of their own AI governance frameworks. This global regulatory pressure is externally pushing AI companies to transform ethical compliance from optional to mandatory, and Anthropic's proactive cross-disciplinary ethics dialogue is, to some extent, a forward-looking response to this regulatory trend.

Anthropic's Differentiated Path

Anthropic has always positioned AI safety as the company's core mission. Constitutional AI is a training method proposed by Anthropic in 2022, with the core idea of providing AI systems with an explicit "constitution"—a set of written principles that the model references when self-evaluating and correcting its outputs, while introducing an "AI feedback" (RLAIF) mechanism to reduce dependence on large-scale human annotation. From Constitutional AI methodology to responsible model release strategies, the company has accumulated extensive practical experience at the technical level.

However, Constitutional AI also exposes a deeper dilemma: Where does the content of the constitution itself come from? Who has the authority to establish these principles? Can these principles transcend cultural boundaries? This dialogue with humanities experts extends safety awareness from the technical domain to deeper philosophical and cultural dimensions, seeking answers to these unresolved questions.

The significance of this approach lies in its frank acknowledgment of a fact: technical means alone cannot fully solve AI's value alignment problem. Constitutional AI can set behavioral boundaries, but where do the value judgments behind those boundaries come from? This is precisely where humanities experts can contribute.

From Dialogue to Practice: Challenges and Outlook

Can Dialogue Outcomes Be Implemented

The value of academic dialogue is undeniable, but the real test lies in whether these discussions can tangibly influence the design and deployment of AI systems. Based on past experience, tech companies' ethics boards and advisory mechanisms often face the challenge of being "more form than substance."