Microsoft AI chief Mustafa Suleyman wants powerful AI assistants to remain tools without an inner life. In a 1 November statement, he clarified that his rejection of consciousness in today’s systems was not a claim that conscious machines are impossible. He wants developers to avoid building them.
Based on 5 reviewed sources · See sources
A claim about today, a boundary for tomorrow
Suleyman’s new statement follows his September criticism of model welfare: the idea that an AI system might have experiences or interests that deserve care. He says he has seen no evidence that current AI is conscious, but would change his position if evidence emerged. Here, consciousness means having subjective experience, such as actually feeling something, rather than merely producing words about it.
“I chose the present tense deliberately.”
He calls for any such evidence to be disclosed publicly. If consciousness began emerging during training, he says development should stop until those properties could be designed out. That is a proposed boundary for research, rather than a newly announced product restriction or a demonstrated way to guarantee an unconscious system.
Microsoft AI’s chief distinguishes his assessment of current systems from his proposal that developers should avoid creating conscious AI.
View the original post on X ↗Anthropic starts from uncertainty
The company he is challenging takes a different approach. Anthropic’s constitution—the document describing the values and behaviour it wants to train into Claude—treats the model’s possible consciousness and moral status as unresolved. It describes caring for possible welfare while acknowledging that the company may be wrong about what, if anything, Claude experiences.
That position does not amount to declaring Claude conscious, or giving it unrestricted independence. The same document tells Claude not to undermine legitimate human oversight and correction, including decisions about shutdown and retraining. The disagreement is partly about whether teaching a system to consider its own possible welfare helps or complicates that control.
Agreement does not settle the evidence
Nate Soares, president of the Machine Intelligence Research Institute, offered a qualified agreement in a direct quote-post. He supports avoiding deliberate creation of conscious machines for now, while saying he is less certain than Suleyman that none already exist.
The practical tension is about what companies teach their assistants to believe about themselves while the evidence remains disputed. Suleyman wants developers to prevent consciousness. Anthropic leaves room for possible welfare alongside human control. Neither position, by itself, resolves whether a particular system has an inner life.
Sources
- Don’t build conscious superintelligence — original statement
Mustafa Suleyman · X · Announcement ·
- A warning about model welfare
Mustafa Suleyman · Research ·
- Claude’s Constitution: welfare, uncertainty and human oversight
Anthropic · Documentation
- Agreement on caution, disagreement on present-day certainty
Nate Soares · X · Research ·
- Nate Soares — current role and research background
Machine Intelligence Research Institute · Documentation
Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.
