Do LLMs diminish diversity of thought?
Put aside the political issues, do Large Language Models too often give “the correct answer” when a more diverse sequence of answers might be more useful and more representative? Peter S. Park, Pilipp Schoenegger, and Chongyang Zhu have a new paper on-line devoted to this question. Note the work is done with GPT3.5.
Here is one simple example. If you ask (non-deterministic) GPT 100 times in a row if you should prefer $50 or a kiss from a movie star, 100 times it will say you should prefer the kiss, at least in the trial runs of the authors. Of course some of you might be thinking — which movie star!? Others might be wondering about Fed policy for the next quarter. Either way, it does not seem the answer should be so clear. (GPT4 by the way refuses to give me a straightforward recommendation.)
Interestingly, when you pose GPT3.5 some standard trolley problems, the answers you get may vary a fair amount, for instance on one run it was utilitarian 36% of the time.
I found this result especially interesting (pp.21-22):
The second, arguably more surprising finding was that according to each of the three distance metrics, our sample of self-reported GPT liberals were still closer to the human conservative sample than it was to the human liberal sample. Also, the L1 distance metric found that self-reported GPT liberals were—among human liberals, human moderates, human conservatives, and human libertarians—closest in response to human conservatives…We thus robustly find that self-reported GPT liberals revealed right-leaning Moral
Foundations: a right-leaning bias of lower magnitude, but a right-leaning bias nonetheless.
The authors seem to think this represents an inability to GPT models to represent the diversity of human thought, on the contrary I think this shows their profundity. In my view many “liberals” (not my preferred term) actually have pretty conservative moral foundations in the Jon Haidt sense, namely, in spite of what they may say the liberals prioritize “in-group, authority, and purity,” rather than worrying so much about actual “harm and fairness.” Just like so many conservatives.
No, GPT does not know all, but sometimes it hits the nail on the head. An interesting paper, even if I part company with the authors on a number of their interpretations.
Via Ethan Mollick.