With Vote-By-Mail on the Rise, Automated Signature Matching is Not a Panacea for Election Officials
With the ongoing public health threat of COVID-19, officials are scrambling to figure out how to administer a general election during a pandemic. With expected increases in the use of absentee ballots (also referred to as vote-by-mail), election officials and vendors alike are preparing to scale up. And if recent elections are any indication there is a lot to prepare for. Both Georgia and Wisconsin each saw more than 1.1 million voters cast absentee ballots in recent primaries. But just as quickly as states figure out how to accommodate absentee ballot requests, they will need to figure out how they will process these ballots. Election officials seeking to save time ease the administrative burden on their staff may look to automated signature matching software, but this is not a panacea.
Signature matching — or, comparing a voter’s ballot envelope signature to that of their voter record — is the most frequently used method by states to verify ballots. If a voter’s signature cannot be verified, the ballot is often not counted. Thus, the signature matching process is critically important in the absentee balloting process, particularly in states without stringent voter identification laws. Despite this, automated signature verification processes are often in uncharted regulatory territory at both the state and federal level.
From a civil rights perspective, signature matching is a better option to ensure election integrity than requiring voters to produce photo identification. Indeed, many voter ID laws are often thinly veiled attempts to suppress the power of Black voters, poor voters, and students. However, the accuracy of automated signature verification is unsettled, and presents significant concerns of bias. If election officials and advocates are not careful, they could adopt a practice that harms vulnerable groups that they are trying to help.
Too often technology has been viewed as a neutral tool, only to find out that bias baked into assumptions of algorithms has exacerbated racial disparities. Indeed, concerns of bias in technology is nothing new, as civil liberties advocates have continued to raise the alarm about the harm of pretrial risk assessments and facial recognition software. Like any algorithm created by humans and trained on data based on the world as it is, the risk of bias in automated signature matching is based on language and writing. This software is often trained on single-language (i.e., English) handwriting to refine the algorithm that allows for the best matches. Thus, certain voters, such as those with disabilities or who do not write in English, may be at an inherently higher risk of having their ballot rejected based on a non-matching signature just because of the algorithm and way the data was trained. However, these are not the only voters affected.
Increased absentee ballot usage combined with the use of automated signature matching can potentially be a recipe for disenfranchisement. And there is great cause for concern. Mail-in ballots are generally rejected at a higher rate than ballots cast in person. A recent study found that in 2018 Florida voters of color comprised less than 28% of those voting absentee, but 47% of all rejected ballots. Out-of-state and military dependents also had disproportionately higher rejection rates. Similarly, New Jersey’s May special election saw 1 in 10 mail ballots thrown out, and over one quarter of those discarded ballots were due to a mismatched signature.
Lawsuits challenging state signature matching laws have focused on the inadequate training of election workers or the lack of procedural protections for voters to be notified of issues with their ballot. This risk of wrongfully rejecting ballots becomes even more acute in states without a formal cure process, which affords voters an opportunity to fix issues with rejected ballots so that their votes are counted. In fact, according to the National Conference of State Legislatures, only 16 states have statutory cure processes to address signature discrepancies.
Moreover, even the way signatures are collected can impact the accuracy of automated signature matching systems. For example, many voters now register at a motor vehicle agency where their signature is digitized using a signature pad; these signatures notoriously look different than those handwritten on paper. Automated signature matching may also cause younger voters to be disproportionately disenfranchised as their handwriting changes over time. More still, immigrants who have recently become citizens may also have their ballots disproportionately rejected.
To protect voters from being disenfranchised, officials should match signatures without the use of automated software. If automated signature matching software is used, states need to step in with strong regulations. Software should be non-proprietary, with the algorithm and training data available to the public for inspection; doing so promotes transparency and bolsters democracy. Election officials should also regularly audit the software for bias to make sure that ballots of vulnerable voters are not disproportionately rejected. If humans are inserted into the process as a ‘safeguard’ against rejected ballots, election officials should also evaluate any automation bias — or the tendency to affirm machine decisions — in order to understand how people interacted with the automated decisions.
Making sure that all votes are counted in November will take deliberate action by election officials, and we must make sure that civil rights aren’t trampled for the sake of convenience.