> The current argument is not “is artificial sentience generally impossible,” it is “are the machine learning systems we are currently building and operating sentient.”

The argument is 'what is sentience?'

>And in fact it is supported by tons of material evidence: for example they don’t act sentient at all if you decline to provide a prompt

This falls into the category of 'appearance of sentience'. Based on this logic all that would be required would be a perpetually self-prompting InstructGPT instance.