AI Privacy Test Exposes Vulnerability in Georgia Voting System Ahead of Midterms

Less than two weeks before early voting begins for the pivotal midterm elections, Georgia election officials are grappling with a serious privacy vulnerability: the ability of artificial intelligence to de-anonymize ballots. The alarm was triggered by an experiment conducted by Max Springer, a postdoctoral research fellow at Princeton University’s Center for Information Technology Policy, who successfully used AI to identify how voters cast their ballots.

Springer’s test required only a $20 subscription to a large language model and data obtained through a public records request. Within hours, he established a pipeline capable of analyzing and identifying secret ballots across the state. “At no point did the agent refuse to comply or raise concerns over implementing the exploit,” Springer reported.

The findings prompted an emergency meeting of the Georgia State Elections Board on Thursday. The core issue lies in the cast vote record (CVR), a digital file created when a voter feeds their paper ballot into a scanner. Each CVR is assigned a unique identifier based on the order in which ballots are scanned. Because this sequencing is not sufficiently random, it can be correlated with sign-in logs and other public data to reveal a voter’s identity.

Springer stated that he was able to recover the voting order for 1.52 million ballots—98.9% of in-person ballots across 114 of the state’s 139 counties. In smaller counties such as Heard, where fewer people vote on any given day, the AI agent could match the majority of early in-person voters to specific ballots, with remaining entries identifiable within a single swap.

“In some of Georgia’s counties, I can determine the exact ballot cast by every voter at the vote centers – ballot secrecy is entirely lost,” Springer wrote. The vulnerability echoes a flaw discovered nearly four years ago in Georgia’s election software, which also allowed for the potential linking of voter identity to secret ballots.

Ben Adida, a cryptographer and founder of the election technology nonprofit VotingWorks, addressed the board regarding the crisis. He noted that outgoing Secretary of State Brad Raffensperger has ordered the redaction of ID numbers in any public release of tabulation data, a move Adida described as substantially addressing the flaw.

However, the road to a permanent solution has been obstructed by political turmoil. Representatives for the secretary of state’s office claim they have sought legislative funding to overhaul the voting software for three years but received nothing. Robert Sinners, a spokesperson for the office, cited a litany of ignored requests for appropriations, directed largely toward Tim Fleming, the Republican nominee for secretary of state.

“We have presented numerous plans to the legislature for this – not, not just a patch, but an entire overhaul of the software to fix up any known vulnerabilities,” Sinners said. “Guess what we were gifted with from the legislature? A big fat goose egg.” Fleming declined to comment.

Georgia faces conflicting legal requirements: the state constitution mandates ballot secrecy, while laws enacted after the 2020 election disputes demand transparency in election records. This tension complicates potential fixes. Elections board member Salleigh Grubbs proposed a procedural change requiring poll workers to hold ballots in a tray to shuffle them out of order before scanning. However, this idea was dismissed by county officials and board member Janelle King, a conservative Republican, who raised concerns about the logistical burden of retraining staff and the risk of ballots being lost or dropped during manual handling.

With early voting imminent, officials acknowledge that significant process changes are no longer feasible. Sinners summarized the situation bluntly: changes at this stage are “a day late and about $30m short.”