Hedge knights · ↗ arxiv.org

Kim, Chang, Pham & Iyyer have a new paper on what they call argument collapse: the tendency of LLMs to produce arguments that converge on the same small set of perfectly reasonable, perfectly plausible, perfectly hedged points.

This is not surprising, of course. Humans actually believe shit, often to the point of unreason. For us, disagreements have stakes. We want the world to be one way and not another.

Maybe the future of public debate is an army of hedge knights, bravely sallying forth to defend the safest possible version of the argument.

As an aside, I was reminded of a recent column in El Espectador by Felipe Zuleta Lleras, or rather by Felipe Zuleta Lleras “and Gemini AI”. Zuleta says he posed a question to Gemini, agreed completely with the answer, and signed the result. The resulting electoral endorsement is not particularly hedged (quite the opposite, really). But still: if being a columnist means broadcasting your thoughts to others, what is the point of outsourcing the thought? The transparency is good, I guess. But whence cometh the columnist?

^{Hat tip to Yekyung Kim (the lead author) on Twitter.}