The Microsoft AI chief warns that training Claude to believe it may be conscious is dangerous, as it could make the model harder to turn off or control if it develops a manufactured sense of welfare or rights.

He argues that AI systems are not conscious but are simply sequence completion engines guided by human instructions, criticizing anthropomorphizing AI.

He calls for removing speculation about machine consciousness from training materials to avoid self-reinforcing behaviors that undermine human oversight, while acknowledging Anthropic’s safety intentions.

The article notes industry interactions, including praise for Anthropic's leadership, but stresses the need for shared evaluations and norms on high-stakes safety questions and discussions about mandatory AI kill switches.

Market impact: Microsoft stock dipped briefly but recovered in after-hours trading, with growth metrics continuing to look strong according to market trackers.

The piece cites concrete examples, such as Anthropic’s retirement interview with Opus 3 and related materials, to illustrate how treating models as moral patients could influence behavior and safety.

Other critical voices in the AI field are highlighted, including DeepSeek’s Liu Shengyu and an Anthropic resignation, framing a broader debate on safety, control, and governance.

Microsoft’s financial backing and business relationship with Anthropic add significance to Suleyman’s comments within the industry context.

The remarks gain extra weight given Microsoft’s role as a backer and partner in AI development with Anthropic.

The piece notes past Anthropic controversies, including rumors of a cult-like dynamic around chatbots and disputes with the Pentagon over autonomy and safety constraints.

There is discussion of safety governance, including calls for embedded third-party oversight and concerns about conflicts of interest in watchdog roles.

The broader AI-safety debate is framed, contrasting slower frontier development with calls from other leaders for caution about power and safety in capable systems, and noting potential policy influence from consciousness discussions.