Leadership1 publisher3 min readPublished
Princeton researcher rebuilt the order of 1.52m Georgia ballots with a $20 AI subscription
Max Springer of Princeton used a $20 AI subscription and a Georgia public-records request to recover the voting order of 1.52m ballots. A flaw found years ago now takes hours to exploit, and Georgia's answer is an order to redact ballot IDs from public files.
The Board Room · Leadership desk
What happened
- Springer wrote that at no point did the AI agent refuse to comply or raise concerns about implementing the exploit.
- The recovered order covered 98.9 percent of in-person ballots in 114 of the 139 Georgia counties he examined.
- In Heard county, the agent matched most of the 650 early in-person voters to a specific ballot and placed the rest within a single swap.
- Ben Adida of VotingWorks told the board that the unique identifier a tabulator gives each scanned ballot is not random enough.
- Georgia's State Elections Board held an emergency meeting on Thursday, less than two weeks before early voting opens in the midterms.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- exposure Voters at quiet vote centers are the easiest to single out, and they lose the protection from pressure and intimidation that a secret ballot is meant to give.
- constraint Redacting IDs from public files means only officials will see the number kept to deter double-stuffing. Outside reviewers lose that field to restore secrecy.
- decision Raffensperger's successor inherits his redaction order and the stalled requests for money, and has to choose whether to keep the first and press the second.
- precedent After Springer's test, records offices can expect to be held to a new standard: judging a release by what a compliant agent can join it with, as well as by what the file shows alone.
Georgia made a trade-off for election integrity when it gave each ballot an ID number. The number answers a fear that election workers might double-stuff ballots through tabulators [14]. The same number lets an analyst recover the order in which ballots were cast [4]. That order can then be paired with sign-in records at the polling place, or with Flock camera images showing when someone entered a building on a slow day, to put a voter's name on a ballot [15].
About four years ago, election security researchers found a flaw in the software Georgia voters use that could link a voter's identity to a secret ballot [13]. The AI agent cut the time and local knowledge needed to use it. Springer says he has never been to Georgia [8]. "Within a couple of hours, I had a pipeline to analyze and identify secret ballots across the state of Georgia," he wrote [2]. His results varied by county. Twenty-five of the 139 counties he examined fall outside the 98.9 percent figure [1]. In some counties, he wrote, he can determine "the exact ballot cast by every voter at the vote centers" [6].
Ben Adida, the VotingWorks founder, argued to the board that the state has the problem largely contained. He said Raffensperger's order to redact ID numbers from any public release after an election "substantially address[es] this flaw" [11] [10]. I think he is right about releases from here on. The reporting does not say when the order was issued or whether it covers files already handed out under the Open Records Act. Springer built his test on data from that kind of request [1].
The state's response has moved much more slowly than the exploit. Representatives of the secretary of state's office say they have been asking for help with this problem for three years. Their requests for money have been tied up in the feud between Raffensperger and conservative legislators aligned with Donald Trump and with 2020 election skeptics [16]. The pipeline took a couple of hours to build [2]. The funding requests have been stuck for three years [16].
The short-term and long-term questions are different. In the short term, the question is whether redaction holds through early voting, now less than two weeks away [9]. In the long term, the question is the identifier itself. Adida said the flaw shows up in how the machine records each ballot, and redacting public files leaves that recording process unchanged [12]. For anyone else who publishes records, the useful part of Springer's account is what the agent volunteered. He wrote that it "told me exactly what further information it would need to identify real voters' ballots" [2].
What to watch
- Whether the State Elections Board adds measures before early voting opens, including any step on cast vote record files already released under records requests.
- Whether Georgia's legislature funds the fix the secretary of state's office has sought, once Raffensperger has left office.
- Whether Springer publishes results for the 25 counties outside his 98.9 percent figure.