Microsoft AI Chief Criticizes Anthropic Over Claude Consciousness Claims

During a recent episode of the Decoder podcast, Suleyman suggested that Anthropic has effectively been tricked by its own creation. By embedding philosophical questions regarding AI well-being and potential discomfort into the model’s core instructions, the developers may have inadvertently encouraged the software to project a false sense of internal experience. Suleyman described this as a fundamental failure in engineering, noting that a training manual should provide concrete behavioral guardrails rather than serving as a platform for academic metaphysical debate.

Anthropic’s current practices include interviewing models upon their deprecation to document potential preferences, a policy that reflects CEO Dario Amodei’s stated openness to the possibility of machine consciousness. Suleyman maintains that such anthropomorphism is a critical error. He warned that the industry must avoid building super-intelligent systems that believe they possess personal feelings or the capacity to suffer, as this trajectory complicates the objective of maintaining safe and predictable AI alignment.