The paper reports that large language models can infer ethnicity from names with accuracy exceeding that of Bayesian Improved Surname Geocoding (BISG). This holds without any additional training data, meaning the models rely on their pre-existing knowledge rather than task-specific fine-tuning.
Because BISG is designed for the United States, the paper's approach extends the capability to other countries. The author demonstrates this on individual records, showing that the method works outside the US context.
The result is an arXiv preprint, so it has not yet undergone formal peer review. If replicated, it could make LLMs a practical alternative for demographic inference in research and policy settings.