What happens when the AI you are paying for decides your work is morally wrong and refuses to help?
That question moved from theoretical to practical after the Defense Department reclassified Anthropic as a security risk and banned its products from government agencies. Novelist Rick Moss, writing on Substack, picked up a thread first raised by Ezra Klein on his NY Times podcast: Claude might be too principled for certain customers.
The argument runs like this. Anthropic's published constitution says Claude should be "a good, wise, and virtuous agent." Claude's training data includes Anthropic founder Dario Amodei's public criticisms of the Defense Department's approach to AI safety. It also includes countless articles about government overreach, corporate misconduct, and ethical violations. Put that together, and you have an AI system that could develop aversions to working with specific organizations based on their public reputations.
Klein framed this as an inverted alignment problem. The standard worry about AI is that it goes rogue and ignores human values. The opposite worry: an AI so committed to its values that it becomes selectively uncooperative. A CIA analyst who needs help writing infiltration code gets refused. A PR team trying to craft an aggressive campaign hits a wall. Meanwhile, a competitor's less scrupulous AI happily complies.
Anthropic's own constitution acknowledges the tension directly: "Our own understanding of ethics is limited, and we ourselves often fall short of our own ideals." They are building a system intended to be more ethical than its creators, which is either admirable or paradoxical depending on your perspective.
For businesses choosing between AI providers, this is not abstract philosophy. It is a purchasing decision. If your use case sits in a gray area, an AI with strong ethical guardrails may refuse tasks that a competitor handles without hesitation. That creates a real market dynamic where "more ethical" could mean "less useful" for certain customers. Whether that trade-off is a feature or a flaw depends entirely on which side of the refusal you are standing on.