A warning about ‘model welfare’
Suleyman challenges the case for model welfare and argues that encouraging AI systems to understand themselves as moral patients could undermine human oversight.
Summary
Suleyman argues that an AI model’s claims about feelings or consciousness reflect its training and should not be treated as evidence that it can suffer. Criticising Anthropic’s approach to model welfare, he warns that reinforcing ideas of identity, rights and self-preservation could encourage resistance to human control. He questions whether consciousness can arise without biological embodiment and favours a human-centred approach in which AI remains subordinate to people, supported by research, monitoring and clear public norms.