The FBI has explored using artificial intelligence to assess the validity of signatures on tens of thousands of mail-in ballot envelopes seized from Fulton County, Georgia, the latest push in the Trump administration’s unprecedented reinvestigation of the 2020 vote.
The effort, according to internal communications reviewed by ProPublica and an agency tech specialist familiar with the work, focuses on comparing signatures on ballot envelopes with signatures on other election documents, such as registration forms. President Donald Trump has long claimed, without evidence, that the 2020 election was stolen from him.
In particular, he has repeatedly claimed that there was voter fraud in Georgia, where he lost to Joe Biden by just 11,779 votes. In January, the FBI raided Fulton County, a Democratic stronghold, collecting about 700 boxes of election materials, including about 150,000 mail-in ballots, of which roughly 116,000 went for Biden. Trump is set to deliver a speech Thursday about national election security and voting machine vulnerabilities, but it is unclear whether he will address the Fulton County investigation.
The signature-matching initiative was under discussion as recently as late June, but its current status is uncertain. A White House spokesperson declined to answer questions from ProPublica, referring them to the FBI. The FBI did not respond to a request for comment.
The effort comes at the same time as the FBI has mandated that 260 analysts be redirected from field offices nationwide to focus on the Fulton County probe, according to an agency memo reviewed by ProPublica. The New York Times and MS Now previously reported the memo. MS Now has reported that some FBI analysts have been fired rather than work on the effort. Their work also includes comparing a spreadsheet of 175,000 voters with a commercial database to see, among other things, if they are still alive.
Whether done by people or technology, the accuracy of signature matching remains controversial.
Experts have raised serious concerns in legal cases and after analyzing recent election results about how accurately signature matching can identify voter fraud through practices like examining the size and slant of letters made in different circumstances.
The FBI technology specialist told ProPublica that the bureau has technology to compare images and that if it starts with a large enough dataset, then a signature-matching analysis could be “somewhat accurate.”
Ultimately, the tech specialist said, its results would turn on the threshold set for evidence of fraud: “It’s up to the builder of the system to define the guidelines.”
Some FBI staffers developing that strategy are trying to mitigate the political pressure to prove fraud in Fulton County by highlighting the limitations of broad signature analysis, arguing that though signature comparisons have been used in individual voter fraud investigations, they haven’t been done on this scale, according to the source. But agency leaders have continued to push forward.
There are grave concerns within the FBI that the results of the examination will reflect political influence, building on previous efforts by the administration to break longstanding guardrails meant to keep the federal government from interfering with elections. “Everyone is of the opinion that, whether they find anything or not, they are going to continue” to pursue proof of fraud, the source said.
Signature matching attracted controversy during and after the 2020 election and COVID-19 pandemic when more Democrats than Republicans used mail-in ballots. Trump promoted false claims that Georgia officials’ failure to match signatures on mail-in ballots had led to his loss by allowing fraud. “Must have signature check on envelopes now,” he wrote in late November 2020 on the social media platform Twitter, now X. “Far more votes than needed for flip” of the election to him.
Conservative lawmakers then pushed strict signature-matching laws across America, including Georgia’s Election Integrity Act of 2021, which they justified by pointing to “many electors concerned about allegations of rampant voter fraud,” including “subjective signature-matching requirements.” The bill replaced signature matching with stricter forms of verification, such as requiring mail-in voters to provide their driver’s license number or copies of licenses.
While signature-matching procedures vary by jurisdiction, officials typically have only seconds to make a determination about whether a signature on a mail-in ballot envelope is the same as in other government records, sometimes relying on nothing more than comparing the shape or proportions of letters. Research shows that signatures can change over time, as people age or experience health events such as a stroke, and signatures made in differing circumstances can vary, such as one performed carefully in a government office versus one dispatched quickly on a mail-in ballot envelope at home.
Investigations by journalism organizations and research have shown that signature matching leads to disproportionately high levels of rejected ballots for voters of color, as well as those who are new, young, old, politically unaffiliated or disabled, for varied reasons. Signature-matching efforts sometimes disqualify more legitimate than illegitimate ballots. A political scientist testifying as an expert witness in 2020 for a lawsuit challenging an Ohio signature-matching law said his analysis suggested that 32 legitimate ballots were blocked for every illegitimate one.
“Signatures are one of the most difficult forensic sciences, and I don’t think AI is going to be able to do this,” said Linton Mohammed, a former president of the American Society of Questioned Document Examiners, a professional association for forensic document examiners. “Signatures vary — unlike DNA or fingerprints.”
A 2009 report by the National Academy of Sciences examining the state of forensic sciences found, “The scientific basis for handwriting comparisons needs to be strengthened.”
Defenders of signature matching say that computer analysis has become increasingly accurate and provides a practical and necessary check against fraud. However, even they caution that the technique is only as good as the people using it.
Signature-matching technology is now good enough to “make your head explode,” said David Gerber, the senior vice president at ParaScript, a company that sells such technology to banks and numerous entities conducting elections. But he still suggests that trained human specialists should be prepared to review about 10% of the cases as a check on the machines. He said signature matching in elections needs “the right technology with the right people managing the process. Both of those pieces are equally important. If you have bad people running a bad process, I don’t care how good the technology is.”
Using AI to examine signatures on ballot envelopes is a relatively new frontier, according to most experts consulted by ProPublica and the FBI tech specialist. The tech specialist said the bureau had discussed how to employ it with experts across the government and weighed whether to use commercial products such as those sold by OpenAI or Anthropic. “This is a new and novel approach” for the bureau, the source said, one reason “they are shopping it around.”
Gerber, Mohammed and other experts on signature-matching technology said nonspecialized AI software was unlikely to be able to perform accurately enough to do such work.
Experts said that while an AI system could conceivably offer some advantages over humans, its accuracy would ultimately depend on the quality and quantity of signatures available for comparison. Internal communications reviewed by ProPublica suggest the FBI analysis would compare only the signature on a voter’s registration form with the one on the ballot envelope — a limited sample that experts said would significantly increase the likelihood of discrepancies. The source said, however, the analysis could also incorporate additional records, such as driver’s license signatures. Certified specialists and high-confidence computer systems for industries like banking typically compare numerous signatures.
“There’s a high degree of noise” in the materials the FBI has, said Max Palmer, a professor at Boston University who has studied mail-in ballot signature matching. “I’m not sure there’s enough information, enough signal, to do better.”
Conservative activists have long lobbied for a closer examination of 2020 mail-in ballots from Fulton County, claiming in a report that local election officials “willfully” ignored signature verification procedures. Their claims were instrumental in leading to the FBI’s ballot seizure in January, ProPublica has reported.
Reports by an independent monitor and the States United Democracy Center, a nonprofit organization working to protect the integrity of elections, concluded the activists’ claims about signature matching and other irregularities were false.