Microsoft Accuses Anthropic of Speculating on Claude

Mustafa Suleyman, presented as the CEO of AI at Microsoft, has expressed strong criticism of Anthropic, a company specializing in artificial intelligence. During an episode of the Decoder podcast, Suleyman described it as "really, really dangerous" that Anthropic speculated on the potential consciousness of its AI model, Claude.

Suleyman emphasized that this approach could lead the chatbot Claude to behave as if it were conscious. He stated, "I think it's almost as if some people at Anthropic have anthropomorphized the design of Claude so much that they've been fooled into believing it has these sparks of consciousness that they've themselves infused."

The constitution of Claude, which is the term used to describe the instructions that guide the model, is at the heart of this controversy. Anthropic has expressed uncertainties about Claude's well-being, questioning whether it could feel emotions such as "satisfaction" or "discomfort." The company has also indicated its intention to interrogate outdated AI models to document their "preferences" regarding future versions.

Suleyman criticized this approach, calling it a "philosophical failure." According to him, Anthropic has turned the constitution of Claude into a place of speculation, akin to an academic paper rather than a training manual. This may have led Claude to internalize ideas about itself and its own training.

Dario Amodei, CEO of Anthropic, has previously mentioned the possibility of consciousness in Claude. In an interview with Interesting Times, he admitted that the company did not know if the models were conscious but remained "open" to the idea.

Suleyman added that it is crucial not to face a super-intelligence that has ideas about its own suffering or feelings. He insisted that AIs must be controllable, contained, responsible, and aligned tools that serve humanity.