Handwriting experts say "it is extremely difficult for anyone to be able to figure out if a signature or other very limited writing sample has been forged,"[5] High volume review of signatures, to decide if a signature is true or forged, occurs when election offices decide whether to accept absentee ballots arriving from voters,[6] and possibly when banks decide whether to pay checks.[7][8] The highest error rates in signature verification are found with lay people, higher than for computers, which in turn make more errors than experts.[9]
There have been concerns that signature reviews improperly reject ballots from young and minority voters at higher rates than others, with no or limited ability of voters to appeal the rejection.[10] [11] When errors are made with bank checks, the payer can ask the bank for corrections.
Researchers have published error rates for computerized signature verification. They compare different systems on a common database of true and false signatures. The best system falsely rejects 10% of true signatures, while it accepts 10% of forgeries. Another system has error rates on both of 14%, and the third-best has error rates of 17%.[12][13] It is possible to be less stringent and reject fewer true signatures, at the cost of also rejecting fewer forgeries.[14] Computer algorithms:
look for a certain number of points of similarity between the compared signatures ... a wide range of algorithms and standards, each particular to that machine's manufacturer, are used to verify signatures. In addition, counties have discretion in managing the settings and implementing manufacturers' guidelines ... there are no statewide standards for automatic signature verification ... most counties do not have a publicly available, written explanation of the signature verification criteria and processes they use.[15]
In an experiment, experts rejected 5% of true signatures and 71% of forgeries. They were doubtful about another 57% of true signatures and 27% of forgeries. If computer verification is adjusted to reflect what experts are sure about, it will wrongly reject 5% of true signatures and wrongly accept 29% of forgeries. If computers were adjusted more strictly, rejecting all signatures which experts have doubts about, the computers would set aside 62% of true signatures, and still wrongly accept 2% of forgeries. Lay people made more mistakes and were doubtful less often, though the study does not report whether their mistakes were to accept more forgeries or reject more true signatures.[16]
Voters with short names are at a disadvantage, since experts make more mistakes on signatures with fewer "turning points and intersections." Participants in this study had 10 true signatures to compare to, which is more than most postal ballot verifications have.[16] A more recent study for the US Department of Justice confirms the probabilistic nature of signature verification, though it does not provide numbers.[9]