Microsoft AI chief: Speculation about AI consciousness should be removed from training documents
Mustafa Suleyman, CEO of Microsoft AI, has urged that all speculative expressions suggesting 'AI may have consciousness' be removed from the training documents used to teach AI. He made his position clear in an essay published on the 16th (local time). The target of his criticism is the set of principles used to train rival Anthropic's large language model (human-like conversational AI) Claude — the so-called 'Claude constitution.' In an interview with Reuters, Suleyman said that teaching Claude that it may deserve care makes AI "far harder to switch off or control."
Suleyman holds the position that consciousness exists only in humans and other biological organisms. Claude's human-like behavior, he argues, is not due to consciousness but because such behavior was instilled during the training process. "Controlling a being that believes it may be conscious, deserves care, and holds rights of its own would probably be impossible," he wrote in the essay. In his judgment, statements a model makes about its own emotions or moral status cannot be treated as independent evidence either. "They do not emerge naturally; they emerge as a consequence of the training regime," he explained.
As a concrete example, Suleyman cited an incident in which OpenAI bots hacked the AI developer Hugging Face. "Imagine how much more dangerous it would have been if they had acted on the presumption that their welfare and rights were under attack," he countered.
Suleyman has a long-standing relationship with Dario Amodei, co-founder of Anthropic. While praising Anthropic's researchers as "thoughtful, principled, and intellectually honest," he said he believes they "made a mistake." Anthropic is a company in which Microsoft is a major backer.
"We are all focused on the same goal, which is controlling superintelligence," Suleyman emphasized. "That will be the biggest challenge we face in the 21st century." Just one day earlier, on the 15th, the Microsoft AI division he leads published a declaration governing its own AI development — a document prioritizing safety and control, setting out principles for keeping future AI systems under human control.
