Princeton researcher recovers Georgia ballot order using AI, public records

PHOENIX: A Princeton University postdoctoral fellow reconstructed the sequence of nearly 99% of ballots cast in person during Georgia's primary election using a $20 artificial intelligence tool and election records that were already public, according to research released this week. Max Springer did not access voting machines or confidential files, and he did not travel to Georgia.

By pairing the reconstructed ballot order with another public dataset, Springer identified the voting selections of roughly 1% of early in-person voters. He wrote that at some polling locations he determined how every voter cast a ballot. Ballot order by itself does not show how a person voted.

The work exploits a flaw disclosed in October 2022 by University of Michigan researchers. Ballot scanners assign each electronic record a number intended to obscure when ballots were cast, but the researchers found those numbers follow a predictable pattern that can be reversed. The flaw is called DVSorder. Dominion released updated software to address it, and the update received federal certification in March 2023.

Maricopa County uses the same family of Dominion ballot scanners at the center of the Georgia issue. County records show Maricopa upgraded its Dominion election management software before the 2024 election. The Arizona Secretary of State's latest equipment inventory shows the county will run an even newer version in November 2026. The state's 14 other counties use equipment from other manufacturers.

Bill Gates, a former Maricopa County supervisor who now directs Arizona State University's Mechanics of Democracy Lab, said ballot secrecy underpins public confidence in elections. "It is public information whether or not someone has voted, but how they voted, which particular candidates they voted for, whether they voted in favor of a proposition or not, that is something that is secret information," Gates told Arizona's Family's Generation AI. He added that if people thought others could learn who they voted for, "it might discourage some people from voting."

Gates said the same exposure could enable intimidation, coercion or vote-buying if a person's choices can be verified. He said Arizona election officials were already weighing how to release detailed results for public examination while keeping selections private, a tension that predates generative AI. He said officials now need their own AI expertise: "Just the same way that we had cybersecurity experts who could deal with the hackers, now we have to have experts in AI who can respond, protect information."

The midterm elections are scheduled for Nov. 3, 2026.