LLMs infer ethnicity from names more accurately than BISG
A withdrawn paper shows models can profile race without training data
The facts
- Researcher Noah Dasanaike reports LLM classification hit up to 84.7% accuracy on Florida and North Carolina voter files with self-reported race, versus 68.2% for Bayesian Improved Surname Geocoding (BISG)
- Six models were tested, including Gemini 3 Flash, GPT-4o, and open-source DeepSeek v3.2 and GLM-4.7; adding metadata such as party registration pushed accuracy to 86.7%
- The method was validated outside the US: Lebanese voter rolls with religious sect (64.3%), Indian MPs from reserved constituencies (99.2%), Indian land records with caste (74.0%), and aggregate population distributions across India, Uganda, Nepal, Armenia, Chile and Costa Rica
- The paper reports LLM classification reduces BISG's income bias, which systematically misclassifies minorities in wealthier neighbourhoods as White, and that small transformer models fine-tuned on LLM labels can match BISG accuracy for free, local deployment
- The arXiv listing shows the paper withdrawn as of 15 October 2026, superseded by a merged submission, arXiv:2610.19048, combining this paper with arXiv:2601.21138
A paper posted to arXiv reports that large language models can infer a person's ethnicity from their name and a few public fields. They do it more accurately, the author says, than the US government's standard tool for the job, Bayesian Improved Surname Geocoding, or BISG.
The author, Noah Dasanaike, used stratified samples from Florida and North Carolina voter files that include self-reported race. LLM-based classification reached up to 84.7% accuracy on balanced samples, against 68.2% for BISG. It needed no extra training data to do it.
The method, Dasanaike reports, generalises well beyond US race categories. Tests on Lebanese voter registration recovered religious sect at 64.3% accuracy. Indian MPs from reserved constituencies were classified at 99.2%, and Indian land records yielded caste classification at 74.0%. Aggregate validation against full-count voter rolls in India, Uganda, Nepal, Armenia, Chile and Costa Rica reportedly recovered known population distributions, wherever naming conventions were distinctive.
Enabling extended reasoning added one to three percentage points in some contexts. Adding metadata such as party registration pushed accuracy to 86.7%. The paper also says the approach cuts the income bias built into BISG, which tends to misclassify minorities in wealthier neighbourhoods as White.
The paper further claims that small transformer models, fine-tuned on labels generated by an LLM, can exceed BISG's accuracy while running locally at no cost. That would put large-scale ethnicity inference within reach of any team with a voter file or land registry and a GPU.
The arXiv record shows the submission withdrawn as of 15 October 2026. A note says it has been superseded by a merged paper, arXiv:2610.19048, combining it with a companion submission, arXiv:2601.21138. No reason for the withdrawal is given in the listing.
Why it matters
If the underlying results hold in the merged version, any developer with name and address fields already has a cheap, no-training-data route to racial and religious profiling that beats the census-era statistical tools privacy reviews were built around. That shifts the compliance question from "do we have demographic data" to "can our LLM infer it anyway" — a much harder thing to audit or restrict.
Primary source
arXiv cs.CL — arxiv.orgWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.