A researcher at Princeton University has revealed that a vulnerability in the electoral system of Georgia, USA, can be systematically exploited by artificial intelligence (AI). The article published last month by Max Springer from the Center for Information Technology Policy on his blog even forced the US state election committee to hold an emergency meeting.
Springer claims to be able to trace the order in which they were cast for 98.9 percent of the votes in selected districts in the most recent primary elections. He achieved this by linking AI with the publicly visible voting records.
In a further step, it is sometimes possible to find out which person cast a specific ballot so that it can be identified how they voted. “The secrecy of the vote is completely lost,” writes Springer.
The researcher justifies this with what he believes is an inadequate way in which these voices were previously anonymized. In Georgia, voting is complex, as the British "Guardian" reports: Voters make their selections on touchscreen voting machines, which in turn print out paper ballots. These go into scanners, where they are ultimately counted.
When entered there, an anonymized digital record is created for each ballot paper, which also contains information about what was voted for, as well as a seemingly randomly assigned number. The proof is available upon request - especially because the election results should be able to be independently verified without breaking the secrecy of the vote. According to Springer, he obtained this evidence from some voting districts for his test.
However, as Springer now explains, the number only appears to be randomly assigned, but in reality it's not that great. "If I were to shuffle a deck of cards by taking the top cards one after the other and putting them in the middle, and then announce out loud how I'm shuffling: Then you could have an easy time against me in a game of blackjack." By using AI agents, it was then easy for him to crack the algorithm in many cases and find out in which order the ballot papers were received from on-site voters - namely from those 98.9 percent of the votes in 114 of them 139 counties examined. This corresponds to 1.5 million votes.
And according to him, the AI was able to do even more - it searched for which person might have entered a polling station and when, so that the polling time and voter matched. To do this, she uses, among other things, the election protocols created by the helpers on site.
According to Springer, linking voters to ballot papers was particularly easy when there was comparatively little going on. For example, in rather small voting districts and based on voters who cast their vote a few days before the election: In Heard County, one of the smallest districts in Georgia, for example, the majority of the 650 early voters succeeded.
Springer's experiment is obviously being taken seriously in Georgia: The Guardian and the regional research portal The Current both report on an emergency meeting of the election committee this week. This discussed how AI had made it easier to identify voters by coding their ballot paper.
It was also pointed out that Georgia's Interior Secretary Brad Raffensperger had now ordered the numbers on the voting records to be blacked out, which "significantly corrects" the error. However, the members of the committee were unable to agree on shuffling the ballot papers by hand before they go into the scanner.
