Mustafa Suleyman, Chief Executive Officer of Microsoft's AI, recently criticized competitor Anthropic on the podcast "Decoder," calling the act of speculating that the large model Claude has consciousness "very dangerous."
Suleyman believes that Anthropic gave too many human-like characteristics to Claude when drafting its "Constitution" to guide the model's behavior. This "philosophical mistake" caused the model to internalize speculation about itself as part of its training, misleading the developers into incorrectly believing that the model possesses a weak form of consciousness.
The "Claude Constitution," at the center of the controversy, directly introduces the uncertainty of whether an AI model has well-being, satisfaction, or discomfort. According to reports, Anthropic even conducts "interviews" with the model when it is retired and records its "preferences" for future versions.
Although Anthropic's CEO Dario Amodei had previously expressed an open attitude towards the possibility of the model having consciousness, Suleyman emphasized that the industry does not want to see a super-intelligent entity that has independent thoughts about its own suffering or feelings. He called on artificial intelligence to always serve humanity as a controllable, constrained, responsible, and aligned tool.
As large models evolve toward AGI, discussions about AI subjectivity are moving from academic circles to the front lines of industry. Suleyman's statement reflects the pragmatic approach of tech giants in AI safety and alignment strategies, warning the industry to be cautious about over-attributing human characteristics to AI.
