Have you no shame, sir?
On college theater, confidently wrong answers, and the danger of trusting anything without an intrinsic sense of shame
If you had a friend, an acquaintance or a co-worker who were for some reason utterly without shame, it might be entertaining for a while. There's a reason why "No Shame Theater" spread from the University of Iowa to twenty other college campuses: Sometimes lifting the constraints of ordinary propriety and social niceties gives room for amusing, entertaining, and sometimes even significant ideas to break through. To be literally shameless can be liberating -- for a time.
■ But after a while, living in the shameless friend's orbit would lose its luster, because a real lack of shame causes people to make decisions that can be consequential for others -- sometimes in a painful way. Someone living and acting without fear of consequences and conscious of no shame about the harm they bring upon others could quickly morph into being a verifiable sociopath.
■ We should probably apply the very same lessons to artificial intelligence before we go too far down the road we are presently traveling. It has become a gold rush in every sense, and otherwise serious people are starting to believe that large language models have something equal to real spontaneous cognition, comparable to what we experience as carbon-based life forms.
■ They don't, and we can know that because there is no way to make a computer capable of shame. And this is dangerous because, if the computer is incapable of shame, it won't feel bad about making a mistake or, worse, misleading the user. Instead, we get confidently wrong outputs, because that's what predictive language modeling creates.
■ Shame can be taken much too far within human institutions, but when appropriately moderated, it acts as a social preservative. Acting honorably matters if we wish to gain and maintain others' esteem. For artificial intelligence, this lack of shame will be consequential in ways that we can't really imagine right now.
■ Those consequences will be wholly troublesome if we haven't anticipated them and built rules, regulations, and safeguards into our own human processes to protect ourselves from shamelessly wrong answers. It's bad enough to have humans in powerful roles who never apologize, never back down, and never express shame -- those people are dangerous to us all. But to have computers masquerading as human-like actors doing the same thing? The consequences could be devastating.