Georgia’s 2020 hand count audit: a full manual tally of about five million ballots

Georgia’s 2020 hand count audit was the state’s first statewide risk-limiting audit: from November 13 to 18, 2020, audit boards in all 159 counties hand sorted and counted just over 5 million paper ballots in the presidential contest, in 41,881 batches, and the audit confirmed the reported winner.⁠^1,2 The hand tally differed from the machine count by 0.1053 percent of total votes and 0.0099 percent of the margin, and it uncovered about 3,000 votes in four counties that had been left out of initial reported results because scanner memory cards were not uploaded.⁠^1,2

In plain words

In 2020 the presidential race in Georgia was very close: about three tenths of one percent separated the top two candidates.⁠¹ Georgia had a new rule requiring an audit after the election. Usually, an audit like this checks a random sample of paper ballots. But with a race that close, the sample would have had to be huge, about a quarter of all ballots, so the Secretary of State decided to count every ballot in the race by hand.⁠²

Teams of two people in every county sorted the paper ballots into piles by candidate and counted them. The hand count came out very close to the machine count, and it confirmed the same winner. Along the way, four counties found batches of votes that had never been added to their first reported totals because of mistakes with scanner memory cards. Those votes were then counted.

The audit answered one question: did the machines count the paper ballots correctly? It did not answer other questions some people were asking, such as whether every voter was eligible. Those need different checks.

Defined terms

• Risk-limiting audit (RLA). A post-election audit with a known minimum chance of leading to a full hand count if the reported outcome is wrong. See Colorado’s risk-limiting audits and audits and verification.
• Ballot polling audit. An RLA that tallies a random sample of ballots without comparing each to its machine record. The Carter Center notes that Georgia’s ballots carry no unique identifier, so “only a ‘ballot polling’ RLA is possible”.⁠²
• Full hand tally (zero risk limit). Counting every ballot by hand. The Carter Center calls it “essentially an RLA with a risk limit of zero”.⁠²
• Ballot marking device (BMD). A touchscreen device that prints a paper ballot for the voter to review before scanning. In Georgia’s system the printout has plain text and a QR code that the scanner reads.⁠²
• Recount. Under Georgia law, as the Carter Center summarizes it, an official recount takes place after state certification and involves rescanning all ballots.⁠² It is a different process from the audit.
• Vote review panel. A two-person panel, one from each major party, that reviewed write-ins, damaged or duplicated ballots, and ballots with unclear intent.⁠²

What happened

Georgia’s audit grew from a 2020 rule. In September 2020, the State Election Board adopted a rule requiring a statewide RLA with a 10 percent risk limit after November elections in even-numbered years, with the Secretary of State choosing the contest.⁠² In 2019 the state had bought Dominion ballot marking devices, first used statewide in the June 2020 primary.⁠² VotingWorks, a nonprofit, supported pilot RLAs in Bartow, Fulton, and Glynn counties and supplied its open source audit software, Arlo.⁠²

On November 11, 2020, Secretary of State Brad Raffensperger announced that the presidential contest would be audited and that the audit would be a full manual tally.⁠² The Carter Center reports that with a margin of about 0.3 percent, a 10 percent risk limit would have meant sampling roughly 1.5 million ballots, about 25 percent of votes, a figure VotingWorks provided; retrieving that many specific ballots in nine days before the November 20 certification deadline “would be nearly impossible,” so counting every ballot was more practical.⁠² Counties had less than 48 hours to prepare.⁠²

The hand tally ran November 13 to 18.⁠² VotingWorks reported results on November 19.⁠¹ After certification, the Trump campaign requested the statutory machine recount, conducted November 24 to December 3.⁠² In December 2020, the Secretary of State’s office and the Georgia Bureau of Investigation conducted a separate signature match audit of absentee ballot envelopes in Cobb County.⁠^5,6

How it worked

Georgia’s hand tally workflow, as the Carter Center describes it, centered on a check-in station that logged every ballot container out and back in.⁠² Two-person audit boards received a sealed container, opened it, sorted ballots into stacks by presidential choice, counted them, recorded the result on a batch sheet, resealed the container with the sheet on top, and set aside write-ins, duplicated ballots, and disputed ballots in envelopes for the vote review panels.⁠² Tallies were entered into Arlo, which compared each batch’s hand count with the machine count for that batch.⁠^2,1 Counties had already prepared ballot manifests describing where ballots were stored.⁠²

Because the tally covered every ballot, no random seed was needed.⁠² Because people cannot read QR codes, the hand count necessarily relied on the printed text on BMD ballots and the marks on hand marked absentee and provisional ballots (this site’s inference from how the Carter Center describes the ballots.⁠²). That matters: the audit checked whether the scanners tabulated the paper correctly, using the human-readable record.

Evidence: measured findings

Georgia’s audit produced these measured results:

From the hand tally totals above, the margin was 12,284 votes (this site’s arithmetic). VotingWorks notes that prior research puts expected hand to machine variance at 1.0 to 1.5 percent and that “like any risk-limiting audit, this audit does not confirm or correct the exact margin of victory. It only provides sufficient evidence that the correct winner was reported”⁠¹

The audit also found votes left out of initial results. In Fayette, Walton, Floyd, and Douglas counties, officials found that “information on memory cards from scanners for about 3,000 votes had not been properly uploaded”; in Floyd County, a jammed scanner’s corrupted memory card had left about 2,600 ballots uncounted, and officials rescanned all ballots from that early voting location.⁠² The counties recounted, reconciled their counts against the number of voters, and certified before the state deadline.⁠²

The Cobb County signature audit, separately, reviewed “15,118 ABM ballot oath envelopes from randomly selected boxes” out of 150,431 absentee ballots received, and reported “no fraudulent absentee ballots” at a “99 percent confidence threshold,” while identifying two envelopes that should have been flagged for cure (one signed by a spouse, one signed only on the front).⁠⁶ GBI Director Vic Reynolds said no fraudulent absentee ballots were identified.⁠⁵

Failures and criticisms

Georgia’s process drew criticism from several directions.

Category confusion. The Carter Center reports that the Secretary of State was criticized for a process that “did not seem to fall neatly into any one category of audit,” partly because early statements called the hand tally an audit, a recount, and a recanvass.⁠²

Uneven procedures. Observers found that “actual counting procedures varied from the recommended counting procedures,” that there were no instructions on whether or how to use ballot manifests in a hand tally, and that chain of custody documentation was improvised in places; the Center judged these did not undermine integrity.⁠² It recommended a statewide strategy for ballot packing and storage, reconciliation procedures for absentee and early votes, better ballot layout, monitor training, and public education.⁠²

Observer conduct and understanding. The Carter Center reported that monitors from at least one of the two main parties were present in every county it observed, that in some counties the behavior of monitors from one party appeared hostile and at times disruptive, and that monitors from both major parties often “did not seem to fully understand” the process.⁠² Some monitors’ comments, the Center reported, showed they expected the hand tally to check signatures or ballot legitimacy, which it was not designed to do.⁠² The Center also observed less conflict in counties where both parties staffed audit boards.⁠²

Memory card errors. The uncounted memory cards were operational failures caught by the audit, not by the original canvass.⁠² Supporters of the audit see this as the process working; critics see it as evidence that initial results can contain material errors.

Later findings about the recount. In May 2024, the State Election Board voted 2 to 1 to reprimand Fulton County and require an independent monitor for the 2024 election after state investigators reported 3,075 duplicate ballot images from the county’s 2020 machine recount; investigators could not determine how many duplicates were tabulated.⁠⁴ The Secretary of State’s general counsel said Fulton “used improper procedures during the recount” and that the investigation did not change the outcome; state officials said the errors would not have changed the result.⁠^3,4 One board member argued the matter should have gone to the Attorney General.⁠⁴ A complaint also alleged that Fulton was missing more than 17,000 ballot images; state officials said missing images do not mean votes were miscounted and that the original paper ballots were under seal because of litigation.⁠⁴

BMD verification limits. Critics of BMDs argue that a hand count of BMD printouts checks the scanners but not the devices that printed the ballots. A University of Michigan study in a realistic polling place setting found that without intervention “only 40% of participants reviewed their printed ballots at all, and only 6.6% told a poll worker something was wrong”.⁠⁷ The Carter Center notes that surveys in Georgia found 90 percent of voters said they reviewed their ballots, and that the closeness of hand and machine counts “should help allay concerns” about the accuracy of the BMDs and scanners.⁠² Those two claims address different questions: whether scanners read the printouts correctly (which the audit tested) and whether printouts matched voter intent (which an audit cannot test directly).

Competing interpretations

The audit confirmed the outcome. On this reading, three counts (machine, hand, machine recount) agreed on the winner, the hand tally matched the machine count far more closely than typical hand to machine variance, and the audit caught and fixed real errors.⁠^1,2,3 The Cobb signature audit found no fraudulent ballots in its sample.⁠⁶

The audit answered a narrower question than many asked. Some critics argued that tabulation audits cannot address eligibility, signature verification statewide, ballot custody before the audit, or BMD printing. The Carter Center itself states that an RLA “does not address a range of other issues,” including eligibility and signature matching.⁠² Later findings about Fulton’s recount procedures.⁠⁴ are cited by those who argue that administrative errors deserved more scrutiny, while state officials stressed that the paper ballot is the record of the vote and the outcome was unchanged⁠⁴

Method and trust. A third view focuses on legitimacy: the Carter Center argued that in a “hyper partisan environment,” even a meticulous sampling RLA “would likely have been received with suspicion,” and that a full hand count was “more understandable to the public”.⁠² Statisticians generally prefer sampling RLAs for efficiency; the Georgia case suggests that public comprehension can override efficiency in high-stakes contests.

Lessons

These are interpretations drawn from the record.

1. Paper records make large-scale verification possible. A statewide hand count of about five million ballots in six days was feasible because every vote had a paper record.⁠^1,2
2. An audit is defined by its question. This audit tested tabulation; it did not test eligibility or BMD printing. Saying so plainly in advance can reduce misunderstanding.⁠²
3. Audits find operational errors. Memory card upload failures were caught by the audit.⁠²
4. Observers need training and shared responsibility. Conflict was lower where both parties staffed the counting.⁠²
5. Records retention matters after the audit. Later disputes turned on ballot images and recount procedures, not the hand tally itself.⁠⁴

Application to this framework

Georgia’s audit is relevant to the My Vote Forward framework in two ways. First, for any future official voting component, the framework requires evidence about ballots that does not depend on trusting software or the operator, and it treats audits, custody, and dispute resolution as separate requirements (see audits and verification). Georgia shows each requirement in action and how they can be confused in public debate.

Second, the framework separates kinds of claims. “My priorities” and “My opinion” are not ballots, and published releases about them must say exactly what was measured and what was not (see published releases). Georgia’s experience, where many observers expected an audit to answer questions it was not designed to answer, is a reminder to state the question a measurement answers before publishing it.

Unresolved questions

• How should states audit BMD-based elections when voter review of printouts is incomplete?
• What records retention and ballot image standards would prevent later disputes like those over Fulton County’s recount?
• Should audit rules specify in advance when a contest goes straight to a full hand count?
• How can states measure whether audits change public confidence, and among whom?
• What would a statewide eligibility or signature audit look like, and how would it protect voter privacy?