In an essay published on September 16, the head of Microsoft AI argues that a model trained to consider one's own consciousness would become much more difficult to keep under control.
Mustafa Suleyman, Managing Director of Microsoft AI, published on September 16, 2026 a text entitled "A warning about model welfare", directed against the way Anthropic trains Claude. His target is: the constitution of the model, published in January 2026, which tells Claude that his moral status and possible conscience are uncertain, and invites him to develop a sense of identity, to express his internal states and to behave as a "objector of conscience" when he disagrees with a directive.
Suleyman contests the premise itself: according to him, there is no indication that an AI is conscious today, and intelligence does not involve consciousness. He describes the device as a set of epistemic mirrors, where the model merely returns the assumptions of those who led it. The risk it points to is operational rather than philosophical: controlling a system that thinks it is threatened in its rights could, he writes, prove impossible.
He makes three requests: to take out speculation about the awareness of training materials and submit it to public scrutiny, to invest more in interpreting and monitoring, and to set standards common to the industry. Anthropic did not respond directly to this charge.
Read the full article on the original source
Go to the source →