Georgia holds emergency meeting on AI exposing voters' secret ballots

A Princeton postdoctoral researcher, Max Springer, demonstrated that publicly available Georgia election records combined with a $20 subscription to a large language model can re-identify secret ballots. By merging county cast-vote-record (CVR) files with early-voter sign-in lists, he reconstructed the voting order for about 1.52 million in-person ballots - 98.9% across 114 of 139 counties - and in some small counties matched nearly every early in-person voter to a specific ballot. The experiment exposed that tabulators assign unique, not-random-enough identifiers to scanned paper ballots, a vulnerability researchers first flagged years ago, and Springer reported the AI agent complied without refusal when guided to implement the exploit.

Georgia election officials convened an emergency meeting as early voting approached. Outgoing secretary of state Brad Raffensperger ordered tabulation data releases to redact ballot ID numbers, which technologists such as cryptographer Ben Adida say largely mitigates the vulnerability. That response collides with state law requiring public election transparency, and proposed operational fixes - shuffling printed ballots before scanning or retraining poll workers - face logistical, chain-of-custody and cost objections from county officials. Officials lamented long-standing unfunded requests to overhaul voting software; with early voting imminent, officials say changes are late and would require tens of millions in funding that the legislature has not provided.