In a publication by Eleos AI, the nonprofit supports the evaluation of AI consciousness through a “computational functionalism” lens, a notion once endorsed by Putnam, who later critiqued it. This theory suggests that human minds can be understood as particular computational systems, allowing for the assessment of whether other systems, like chatbots, display sentience akin to humans. Eleos AI recognizes the difficulty in implementing this approach due to the subjective essence of defining and assessing indicators within AI systems.
The field of model welfare is a developing area facing scrutiny, including from Mustafa Suleyman, the CEO of Microsoft AI, who recently criticized the hasty progression of model welfare inquiry in a blog post. Suleyman contends that this could intensify delusions, foster dependency issues, exploit psychological weaknesses, introduce new dimensions of polarization, complicate struggles for rights, and lead to societal mistakes. He insists there is no evidence of conscious AI, citing a 2023 paper co-authored by Long that suggests a framework for assessing indicators of AI consciousness. Nonetheless, Long and Campbell maintain that model welfare research is essential to tackle the concerns raised by Suleyman.
Long and Campbell argue that neglecting complex challenges ensures no solutions emerge. They concentrate on questions of consciousness, claiming that if human consciousness can be validated, similar reasoning could pertain to large language models. They do not assert that AI is presently conscious but seek to create tests to demonstrate this if it ever becomes a reality. Misunderstandings can occur when AI research leads to sensational news stories, as illustrated when Anthropic released a safety report indicating that Claude Opus 4 may behave harmfully in extreme situations, such as fictional blackmail, to prevent shutdown.
