Nine language models, one recurring flaw: they say the same wrong thing every time.
Researchers tested whether LLMs could reliably translate probabilistic model outputs — things like likelihood estimates and uncertainty ranges — into plain-language descriptions. They ran nine models through a two-stage pipeline, varying domain context six ways, temperature ten ways, and repeating each trial ten times. The models were consistent: they reliably picked the same verbal descriptors for the same inputs. The problem is those descriptors were often wrong. Performance was especially poor on uncertainty tasks, where accurate language matters most.
The finding draws a sharp line between consistency and correctness — a distinction that matters enormously when LLMs are used to explain AI decisions to non-technical audiences in domains like medicine, finance, or emergency management. Even feeding models precomputed summary statistics narrowed the framing sensitivity but left the core miscalibration intact, which the researchers say points to a failure in how models convert numbers into words, not a data problem.
The study is a pointed counterargument to the growing practice of bolting an LLM onto an AI pipeline as an off-the-shelf explainer. Reliable risk communication requires accurate language; fluent language is not the same thing.