My hypothesis is signature verification where it’s minimally trained individuals doing the comparison seems quite open for inaccuracies. I couldn’t find any large scale studies though looking into this.