FBI looked at using AI to review 2020 ballot signatures

FBI Explores Use of AI to Compare Signatures on Seized Fulton County Mail Ballots

The FBI has examined the possibility of applying artificial intelligence tools to compare signatures on mail-in ballot envelopes collected from Fulton County, Georgia, as part of an ongoing review of 2020 election materials. The discussion occurred internally as recently as late June, though the current status of any such project remains unclear.

In January 2026, the FBI executed a search warrant at Fulton County facilities and seized approximately 700 boxes of election records, including roughly 150,000 mail-in ballots. About 116,000 of those ballots were cast for Joe Biden. The warrant was supported by an affidavit containing specific allegations that a federal judge reviewed before authorizing the seizure. The materials include ballot envelopes whose signatures are now under consideration for further analysis.

Agency communications reviewed by ProPublica indicate that one proposed method would compare signatures on the ballot envelopes against signatures appearing on voter registration forms or other records such as driver’s licenses. An FBI technology specialist familiar with the internal discussions stated that existing image-comparison technology could produce results of varying reliability depending on the size of the dataset and the thresholds chosen for identifying potential discrepancies. The specialist noted that the bureau has consulted experts inside and outside government and considered both internal capabilities and commercial AI products.

At the same time, the FBI directed 260 analysts from field offices across the country to support the Fulton County review, according to an internal memo. Some of the assigned work has included cross-checking a list of approximately 175,000 voter records against commercial databases to assess, among other factors, whether individuals remain living. Reports from The New York Times and MS Now described the reassignment; MS Now additionally reported that certain analysts declined the assignment and were removed from their positions.

Signature comparison has long been used in election administration and fraud investigations, though its accuracy when applied at large scale continues to be examined. Research has documented that signatures can vary because of age, health conditions, or differences in writing conditions. Studies have also found that rejection rates under signature-matching rules have sometimes been higher for certain demographic groups, including voters of color, younger voters, older voters, and those without party affiliation. A 2009 National Academy of Sciences report on forensic sciences concluded that the scientific foundation for handwriting comparisons required further strengthening.

At the same time, vendors that supply signature-verification systems to banks and election offices maintain that modern software has improved substantially. David Gerber, senior vice president at ParaScript, stated that current technology can achieve high accuracy when paired with trained human review of a portion of flagged cases. He emphasized that both the software and the procedures governing its use determine overall performance.

Linton Mohammed, former president of the American Society of Questioned Document Examiners, expressed skepticism that general-purpose AI systems would reach the level of reliability required for forensic document examination, noting that signatures differ from biometric identifiers such as DNA or fingerprints. Boston University professor Max Palmer, who has studied mail ballot signature processes, observed that the limited number of comparison signatures available in many election records introduces substantial variability that can reduce the strength of any automated analysis.

Internal FBI discussions have included consideration of these limitations. Some staff members have pointed out that signature comparisons in past voter-fraud cases have typically involved smaller numbers of documents and multiple reference signatures, conditions not fully replicated in the current dataset. Agency leaders have nevertheless continued to explore the approach.

The review of Fulton County materials follows earlier complaints from conservative activists who alleged that local officials did not consistently apply signature-verification rules during the 2020 election. Those allegations contributed to the decision to seek the search warrant. Separate reviews conducted by an independent monitor and the States United Democracy Center examined the same claims and concluded that evidence of widespread procedural violations was not present.

President Trump has repeatedly stated that irregularities occurred in Georgia’s 2020 election, which he lost by 11,779 votes. He has specifically cited mail-in ballot procedures in Fulton County. The White House referred questions about the current FBI work to the agency, which did not provide comment.

Whether any AI-assisted signature analysis proceeds will depend on decisions about data thresholds, available reference signatures, and the legal standards applied to any resulting findings. The bureau has not announced a final determination on the method.