Mustafa Suleyman argues training AI models to consider themselves potentially conscious could make advanced systems harder to control
