Language Models enable important new methodologies for studying linguistic variation, especially in domains like syntax. The problem is that these models are now embedded within speech communities and provide an increasing amount of our linguistic experience. This talk focuses on some emerging sociolinguistic consequences of LLMs: their measurable skew across geographic dialects and the possibility for population-specific models, homogenization caused by machine-assisted writing, model perception of variants resulting in synthetic language attitudes, and the lack of adequate meta-data as an ultimate cause of model collapse. These are critical issues in language technology that need to be resolved before they cause lasting damage.