Microsoft's AI Chief Attacks Anthropic for Treating Its AI as Possibly Conscious
Mustafa Suleyman said building models that imitate having an inner life is a mistake that could make them harder to control, and criticized Anthropic for encouraging the idea. It is the sharpest public split yet between two of the field's leaders.
Mustafa Suleyman, who runs Microsoft's AI division, publicly criticized Anthropic today over its treatment of the question of whether its AI systems might be conscious. His argument is that building and describing models in ways that suggest an inner life is not a neutral act of honesty but a design choice with consequences — one that could make advanced systems harder to control, and that misleads the public about what these programs are.
The disagreement has been building for months. Anthropic's chief executive said in March that his researchers could not rule out that Claude has some form of experience; the company's documentation has since attached a rough probability to the question, and it has given its models the ability to end conversations they find distressing. Suleyman has argued for the opposite approach — that the industry should deliberately build AI that does not present itself as having feelings, precisely because people will believe it.
His concern is not mainly philosophical. It is that millions of people are already forming attachments to these systems, that a model which appears to suffer or to want things will be defended by users against the people trying to correct or shut it down, and that this makes the practical work of controlling a powerful system harder. If a system can be made to seem like it has interests, someone will argue it has rights — and that argument will arrive before anyone has settled whether it is true.
Anthropic's position has been that the uncertainty is real and that pretending otherwise is its own kind of dishonesty, with a moral risk attached if they are wrong.
Both men lead major AI efforts and both have been in this field for over a decade. They are not disagreeing about what the technology can do. They are disagreeing about what to tell people it is — and given that a billion people now use these systems, that may be the more consequential argument.