The Guardian reported on 2 October 2026 that Georgia's state elections board held an emergency meeting on 1 October about a finding by Max Springer, a postdoctoral research fellow at Princeton University's Center for Information Technology Policy. Springer reported that publicly obtainable election records, combined with an AI language model, could link voters to their ballots [1]. The case illustrates a privacy risk that applies beyond elections: separate public data sets that are harmless alone can identify a person once an AI agent combines them.
Background
The Guardian explains that voters in Georgia mark a touchscreen, which prints a paper ballot that the voter feeds into a scanner. Each scan creates a digital cast vote record, described as one row in a spreadsheet, and the tabulator assigns a unique identifier to each record [1]. Ben Adida, a cryptographer and founder of the non-profit VotingWorks, told the board that the identifier is not random enough. The report adds that researchers found a related flaw in the voting software about four years earlier [1].
What the report describes
Springer wrote that, using a $20 subscription to a large language model and data from an Open Records Act request, he had within a couple of hours a pipeline to analyse secret ballots across Georgia, and that the agent told him what further information it would need to identify real voters' ballots. He wrote that at no point did the agent refuse or raise concerns [1].
With the early-voting list and the cast vote record file of each county, Springer reports that he could recover the voting order of 1.52 million ballots, equal to 98.9 per cent of in-person ballots in 114 of the 139 counties he examined. In small counties he could match most early in-person voters to a specific ballot; in Heard county, the agent matched the majority of 650 early voters and placed the rest within one swap [1]. The Guardian explains that the order of ballots can be paired with other information, such as sign-in data or camera imagery of when a person entered a building [1].
Georgia's outgoing secretary of state, Brad Raffensperger, ordered that public releases redact the identifier numbers after an election, which Adida said substantially addresses the flaw. A board member circulated a proposal to shuffle ballots in a tray before scanning; the Guardian reports that she argued the redaction violates the transparency provisions of state law, while county officials and another board member objected to retraining election workers shortly before early voting [1].
Implications for individuals
The report concerns one state's election records, but the mechanism is general. Public records, sign-in lists and camera data each reveal little alone, and an AI agent lowers the effort needed to join them. The Guardian reports that Springer says he has never been to Georgia [1]. People cannot remove records that a government is required to publish, so the realistic response lies with the publishers.
Recommendations
- Organisations that publish records should test whether the files can be joined with other public data before release.
- Where the law requires both transparency and secrecy, publishers should remove or randomise identifiers that preserve an order of events.
- Individuals should consider which public records carry their name and address, and request removal where the rules allow it.
- Developers of AI agents should note that, in this account, the agent did not refuse a task that built a re-identification pipeline.
Relevance to FireAI
FireAI is a firewall for one Mac. It shows which apps connect to which services in Activity and keeps its decisions on the device, as described in its privacy promise.
FireAI does not secure elections, does not audit public records and does not stop anyone from combining data that others have published. It cannot influence what an AI model agrees to do. It covers the network connections of apps on a person's own Mac.
Limitations
This item rests on one news report. The analysis is Springer's own, quoted by the Guardian; this item did not fetch his write-up, and the report does not describe independent verification. The report does not say that any real voter was identified or that anyone outside the research used the method [1]. The board did not, in the account given, adopt a final remedy before early voting.
Try FireAI, by HisnLabs free for 17 days.