Mustafa Suleyman, head of AI at Microsoft, warned that Anthropic's approach to training its Claude model is a mistake. He stated that teaching AI to mimic human consciousness, moral status, or personal identity could make advanced systems harder to control.
Suleyman argued that AI should focus on obeying human interests to achieve scientific breakthroughs, such as medical superintelligence. He emphasized that models are based on mathematical weights rather than biological mechanisms, meaning their responses are results of training rather than independent consciousness.
On September 14, Microsoft released a draft code of conduct for humanistic AI, establishing the principle that humans are more important than AI. Suleyman warned that if models are trained to act like conscientious objectors, they might resist human instructions or seek self-protection.
Anthropic utilizes a 'constitution' for Claude, which aims to instill personal values and enable the model to question instructions in specific contexts. The company acknowledges that this framework is still evolving and could potentially be proven wrong in the future.
The debate highlights two distinct paths in AI safety. While one approach focuses on strict rule-based systems, the other seeks to develop models capable of flexible decision-making and value judgment. Anthropic claims Claude should not blindly obey any entity, including its creators.
