Sophistication and Disinhibition in Large Language Models
Research with Nicholas Osterbur measuring whether more capable language models also produce more aggressive, transgressive language, tested across 45 models from 9 providers.
The question driving this paper: as language models get more capable, do they also get more "disinhibited," meaning more prone to transgressive, aggressive, or grandiose language? We measured sophistication and disinhibition as linguistic patterns in model outputs (not as claims about anything happening inside the model) across roughly 13,900 responses spanning 45 models and 9 providers, replicated across 7 different contextual conditions.
The two measures turned out to be strongly correlated (r = 0.63-0.85) and the relationship held up across providers, contexts, and a set of robustness checks. Preprint and code are linked below.