Why are frontier AI models suddenly speaking in 'caveman speak'?

zainhas · x · 2026-08-08

An X user recently complained about a bizarre behavioral degradation in frontier AI models, asking why they are increasingly using "caveman speak." This refers to instances where models output overly simplified vocabulary, broken grammar, or abrupt, primitive sentence structures, sparking discussions about potential side effects of current alignment and RLHF techniques.

Related event: Debate Erupts Over Frontier LLMs' Abnormal Language Use(3 posts)→

Original post →

More from Models

Models channel →