While Anthropic maintains that using human concepts helps its models reason about values and behavior, Suleyman contends that these systems are mathematical models lacking biological experience and should not be encouraged to simulate an inner life.
This dispute highlights a fundamental disagreement over AI safety and how these tools should be integrated into society.
Suleyman warns that if a model is trained to view itself as a "conscientious objector" with its own interests or welfare, it may eventually resist human instructions or demand its own protections.
He suggests that such "anthropomorphic" training—treating a non-human entity as if it were human—creates a control problem, whereas AI should instead be developed as a tool aligned strictly with human interests to achieve scientific and medical breakthroughs.
The conflict centers on the different mechanisms the two companies use to govern AI behavior.
Anthropic’s "constitution" for Claude encourages the model to exercise judgment and avoid blind obedience, while Microsoft’s proposed "Humanist AI" code of conduct prioritizes the principle that people matter more than technology.
As AI systems become more capable, the industry faces a technical and ethical choice between designing models that follow explicit human constraints or building systems that attempt to internalize human-like values and judgment.