FBI Considers Using AI Tech to Review Signatures on Seized Mail-In Ballots

Background and Motivation

During its investigation into disputes surrounding the 2020 U.S. presidential election, the Federal Bureau of Investigation (FBI) has initiated an assessment of upgrading forensic signature verification workflows for a large volume of mail-in ballots seized in Fulton County, Georgia. This effort directly responds to legal challenges filed by the Trump campaign contesting election legitimacy and alleging signature inconsistencies.

Current State of Technical Implementation

• The FBI is not developing a proprietary model but exploring integration of existing commercial or open-source signature verification AI tools into its forensic pipeline;
• As reported by ProPublica, proof-of-concept testing involves deep learning–based handwriting comparison systems, potentially leveraging libraries such as OpenCV and TensorFlow, or specialized biometric SDKs (e.g., Topaz Systems or SignDoc);
• The initiative remains at the proof-of-concept (PoC) stage: no deployment in formal evidentiary proceedings, and no validation under court-admissible forensic standards (e.g., ASTM E1789-04).

Compliance and Controversy

• While the U.S. Electronic Signatures in Global and National Commerce Act (ESIGN Act) and state election laws do not explicitly prohibit AI-assisted signature review, they mandate human adjudication as the final determination;
• Experts caution that current AI-based signature matching models exhibit nontrivial error rates (reported False Acceptance/False Rejection Rates ranging from 5% to 12%) under suboptimal conditions — including low-resolution scans, temporal handwriting variation, and non-native script writing — falling short of the ‘high reliability’ threshold required for courtroom evidence;
• The ACLU and Election Integrity Alliance have formally petitioned the Department of Justice for public disclosure of algorithmic transparency, training data provenance, and bias audit reports.