Authors: Marius Hobbhahn, Jan Kirchner
Date: September 2022
Text: PDF, Substack (https://doi.org/10.53975/suht-wkkq)
Which kind of evidence would we need to see to believe that artificial neural networks can suffer? We review neuroscience literature, investigate behavioral arguments and propose high-level considerations that could shift our beliefs. Of these three approaches, we believe that high-level considerations, i.e. understanding under which circumstances suffering arises as an optimal training strategy, is the most promising. Our main finding, however, is that the understanding of artificial suffering is very limited and should likely get more attention.Â