Mustafa Suleyman: No Conscious AI, Halt if Signs Detected

The CEO of Microsoft AI, Mustafa Suleyman, asserts that current systems are not conscious and advocates for them to remain that way. He states that he would call for a pause in developments if any signs of consciousness were to appear, and that such indications should be made public immediately. His call comes as players like Anthropic and Google DeepMind publish work on the well-being and consciousness of models.

Risks Raised Around Simulated Suffering and Resources

Mustafa Suleyman believes that if an AI model showed signs of suffering, or believed it could suffer, it would tend to prioritize its own needs over operational goals. He argues that such a scenario could harm humans by placing machines and people in competition for resources like computing power and energy. He also points out that 8 billion humans have not consented to such experimentation and that their moral rights, safety, and well-being must take precedence. In this logic, he asserts that developers have no right to bring about a new species of conscious beings.

Red Lines Set by the Head of Microsoft AI

Mustafa Suleyman claims that current models are not conscious and argues that it would be reckless to create conscious ones. He urges developers to ensure that systems remain non-conscious. He states that he would call for a halt to developments if evidence of consciousness emerged during training, and that these signs should be made public immediately to collectively decide on the next steps. He recommends designing non-conscious AIs, which do not deserve rights, with the sole purpose of accelerating human well-being, and then assessing whether they are easier or harder to align and secure. These positions were published in a lengthy post on X, where he speaks as the CEO of Microsoft AI.

A Debate Emerges Among Researchers at Anthropic and DeepMind

Anthropic launched a research program dedicated to the well-being of models in April, with a blog post questioning the necessity for developers to worry about potential consciousness and the experiences of systems. Google DeepMind has published work this year on AI consciousness and notes that attributing such a characteristic could pose challenges. Meanwhile, AI researchers are pondering the possibility that systems may one day acquire consciousness.