Human alignment is a specification error. Getting ai to align with human interests and needs can have catastrophic consequences. Based on my background in the History of Science and Ideas as well as my experience working with people and companies on biases, I have come to believe that human alignment is a specification error. Human interests and needs, as we express them and as we pursue them individually and as societies, are biased and short term. Science has identified over a hundred systemic cognitive biases that lead us to make the wrong decisions - while thinking we have the correct information and are being rational in our conclusion. There is both historical and present evidence that humans looking out for what they believe are human interests can mean devastation for whole species, for other humans and even for life in general. Human alignment attempts to make ai safe for humans - but given what humans can do to other humans and to the biosphere that we depend on, we may be shooting ourselves in the foot. Alignment needs to be with something more physically grounded than human interests and needs as we express them.

I'm curious to hear your take on this perspective.