Princeton Researcher Used a $20 AI Subscription to Break Georgia Ballot Secrecy

A Princeton University researcher broke the secrecy of Georgia's ballots with a $20 subscription to an AI large language model and public records obtained through an Open Records Act request. Max Springer, a postdoctoral research fellow at Princeton's Center for Information Technology Policy, says the experiment worked within hours — and the AI never pushed back.

"Within a couple of hours, I had a pipeline to analyze and identify secret ballots across the state of Georgia and the agent told me exactly what further information it would need to identify real voters' ballots," Springer wrote. "At no point did the agent refuse to comply or raise concerns over implementing the exploit."

The stakes are immediate. Early voting in Georgia's midterm election begins in less than two weeks, and the contest will help determine control of the US Congress. Ballot secrecy is a core democratic safeguard: it protects voters from family members, neighbors, employers, and the government itself — anyone who might pressure or intimidate them into voting a particular way. Springer's test demonstrates that AI has now made it cheap and fast to strip away that protection using records Georgia's own transparency laws require it to publish.

What the researcher actually did

Georgia's voting system works like this. Election workers record every voter who enters a polling place. Voters select candidates on a touchscreen, which prints a paper ballot. The voter then feeds that ballot by hand into a scanner for tabulation.

"When a voter inserts their ballot into a tabulator, a digital record of that ballot is created. Think of it as a row in a spreadsheet. One row is one scanned ballot. It's often called a cast vote record, or CVR for short," explained Ben Adida, an MIT-trained cryptographer and founder of the election technology nonprofit VotingWorks, in remarks to the state's elections board. "In order to later retrieve the CVR, the tabulator assigns a unique identifier to each CVR at the time of scanning."

The flaw sits in how the machine assigns that identifier. According to Adida, it is not random enough. If the identifiers correlate with the sequence in which ballots were cast, and that sequence can be paired with other information — the check-in list at a precinct, or camera footage showing when someone entered a building on a slow day — a ballot can be matched to a named voter.

Springer combined the AI model, the early-voting list, and each county's cast-vote record file. The results were stark. He recovered the voting order for 1.52 million ballots — 98.9% of in-person ballots across 114 of the 139 counties he examined. Springer notes he has never even visited Georgia.

"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," he wrote.

Small counties fared worst. In tiny Heard County, Springer said, "the agent was able to match the majority of the 650 early in-person voters to a specific ballot, and the rest to within a single swap."

The vulnerability is not entirely new. Election security researchers discovered roughly four years ago a flaw in the software used by Georgia voters that could connect a voter's identity to their secret ballot. What has changed is the cost and ease of exploiting it. A task that once demanded specialized expertise now runs through a commodity AI subscription in an afternoon.

The paradox behind the numbers

The identifying numbers exist because of political paranoia surrounding Georgia's elections. Each ballot carries an ID number to guard against unscrupulous election workers double-stuffing ballots through tabulator machines. The defense against one imagined threat created the opening for another.

Georgia now finds itself squeezed between two legal imperatives. The Georgia constitution explicitly requires a secret ballot, and the state's law protecting ballot secrecy ranks among the strongest in the country. But Georgia law — revised in the wake of the 2020 election disputes — also demands transparency for election records. Other states that use Dominion software have either patched the software or withhold the vulnerable voting data from public inspection. Georgia has done neither until now.

Officials scramble, two weeks out

Georgia's State Elections Board held an emergency meeting Thursday morning to discuss how AI has made voter identification through ballot codes easier. Outgoing Secretary of State Brad Raffensperger moved to lock down the tabulation data, ordering that any public release redact the ID numbers after an election. Adida told the board the measure "substantially address[es] this flaw."

Not everyone on the board agreed that redaction is lawful. Board member Salleigh Grubbs argued that Raffensperger's fix violates state transparency law. She circulated a draft proposal requiring poll workers to hold ballots in a tray before voters cast them into the machine, shuffling the printed pages out of order to break the sequence.

County elections officials and fellow board members dismissed the shuffling plan as unworkable — including Janelle King, a conservative Republican.

"We are asking our election officials to completely retrain everybody that has been a part of this process up until this point on how to administer the election," King said. "I also have major concerns about dropping our ballots into a secured box and then having other individuals scan my ballot. How do I know that in the shuffling something doesn't fall out? I mean, I just have some major concerns about whether or not we are creating more problems while we are trying to address a problem."

Three years of unanswered requests

The secretary of state's office says it saw this coming. Representatives say they have been asking the legislature for help for three years, and those requests became entangled in the long-running political feud between Raffensperger and conservative legislators positioning themselves in the orbit of Donald Trump and 2020 election skeptics. Raffensperger has remained a top Trump target since refusing to overturn Trump's defeat in Georgia's 2020 presidential election.

Spokesperson Robert Sinners sent reporters a litany of emails and phone records documenting requests for an appropriation to upgrade Georgia's voting system. Governmental affairs committee secretary Tim Fleming — now the Republican nominee for secretary of state — was on the receiving end of many of those requests. Fleming did not respond to a request for comment.

"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."

With early voting imminent, Sinners said, any process changes now would arrive "a day late and about $30m short."

Why it matters beyond Georgia

The episode illustrates a broader problem that AI poses to legacy privacy protections. Data that was technically public but practically inaccessible — because correlating it required expert labor — becomes actionable when an AI agent will do the correlation for the price of lunch. Georgia's cast-vote records were published in the name of transparency; the assumption underpinning that transparency was that no one could realistically reconstruct individual ballots from them. Springer's $20 experiment invalidated that assumption in two hours.

Raffensperger's redaction order closes the most direct route to re-identification for this election. The deeper questions — whether Georgia overhauls its software, how it reconciles its secrecy and transparency mandates, and whether lawmakers fund the fix — will land with the legislature and the next secretary of state, in a political environment where the office's requests have already gone unanswered for three years.