An Algorithmic Failure Beneath the Secret Ballot

Max Springer, a postdoctoral fellow at Princeton’s Center for Information Technology Policy, demonstrates that a long-known flaw in some ballot scanners’ anonymization can be reversed to recover the order in which ballots were cast and, when combined with public voting logs, to deanonymize votes. The scanners assign deterministic “random” identifiers to cast-vote records (CVRs) that can be algorithmically reversed. Using only public CVR files and early-voting lists, and without touching machines or private systems, Springer employed AI coding agents to deshuffle Georgia’s May 2026 primary ballots and recover the in-person cast order in 114 of 139 counties - about 1.52 million ballots, or 98.9% of in-person ballots analyzed. In many places he could uniquely identify voters’ ballots; with additional public audit logs and check-in timestamps he matched nearly all early voters in small counties like Heard and large vote centers such as Ball Ground in Cherokee County.

The write-up argues that large language models collapse the technical barrier that once slowed exploitation, making speedy, automated attacks feasible with minimal expertise. A software fix was certified in March 2023 (Dominion version 5.17+), but slow adoption leaves jurisdictions exposed. The practical risk is stark: publicly visible running ballot counts, CCTV, or poll watchers can supply timing data that links people to ballots. The clear recommendations are urgent patching of affected scanners before the next election, reconsideration of vulnerability disclosure practices given AI-enabled exploitation, and preserving auditable public data while correcting privacy-algorithm deployment.