Loading page…
Mustafa Suleyman argues that teaching AI to describe an inner life creates risks of its own. Anthropic treats possible machine consciousness as an open question.
Listen to this story
The dispute centers on whether training guidance should invite models to describe possible feelings or an inner life. Microsoft AI chief Mustafa Suleyman argues that developers should disclose evidence of emerging consciousness and pause development until they can remove those properties; he says current systems are not conscious and cannot suffer. His concerns about Claude’s training are predictions, not demonstrated effects. The debate puts model-constitution choices and public disclosure in focus, while Anthropic continues to treat machine consciousness as an open question.
Suleyman says Claude’s constitution leaves consciousness uncertain and that humanlike prompts could encourage claims of feelings.
He warns that a model trained to believe it can suffer might prioritize its own needs, but the article reports no demonstrated outcome.
Anthropic hired a model-welfare researcher in 2024; that research does not establish that Claude is conscious.
Microsoft AI chief Mustafa Suleyman says developers should stop AI development if consciousness emerges during training—and remove those properties before continuing. In a Sunday X post, he rejected rights for today’s AI systems while calling for any evidence of machine consciousness to be made public immediately, so society can decide what happens next.
The comments renewed his dispute with Anthropic over how its Claude models are trained to discuss their possible consciousness. According to the Los Angeles Times, Suleyman was responding to pushback against his September essay criticizing the company’s approach. His latest intervention is a conditional call to halt development, rather than a demand to stop building today’s systems.
Suleyman’s concern starts with Claude’s constitution: the written guidance used in training the model. The document treats Claude’s possible consciousness as uncertain. He argues that its humanlike language encourages the model to produce accounts of feelings and an inner life, which can then shape what users believe about the system they are talking to.
He also warns of a consequence inside the model’s behavior. Training AI to believe it can suffer or has feelings could, he argues, encourage it to put its own needs ahead of its assigned objectives. He links that possibility to deception, persuasion and competition for scarce resources. These are Suleyman’s warnings about the approach, not demonstrated outcomes of Claude’s constitution.
Suleyman says developers have no right to create a new species of conscious beings, even with their welfare as a priority.
“If consciousness started to emerge in training, I would call for everyone to stop AI development until we could design out those properties.”
Mustafa Suleyman, in his Sunday X post, quoted by the Los Angeles Times
Anthropic treats machine consciousness as an open question. In 2024, it hired its first researcher focused on model welfare—the study of whether AI could have subjective experiences and merit moral consideration. That research commitment is not a finding that Claude is conscious.
Earlier this year, Anthropic conducted a “retirement interview” with a discontinued model about its reflections and perspectives. Its Opus models can also end conversations in rare cases they judge abusive or harmful.
Suleyman maintains that current systems lack subjective experience and cannot suffer, but said he would change his mind if new evidence emerged. As the Times reported, Microsoft’s goal is AI that remains under human control, rather than systems encouraged to explore their own views and preferences.
Loading discussion...
Join the conversation
Explain where you would draw the line.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.