← All posts

"We counted it by hand" is a reassurance, not a proof

Georgia's historic hand count of 5 million ballots is held up as proof elections work — but the same method was off by a dozen votes in Antrim County, and some races are decided by fewer than that.

It is just past midnight on November 5, 2020, and Antrim County, Michigan has a problem that everyone can see. The county has just reported Joe Biden leading in one of the most reliably Republican counties in the state. The numbers are wrong — obviously, visibly, embarrassingly wrong. And they were published anyway.

The clerk identified the error the next morning: a last-minute programming change hadn't been applied to all the memory cards. Human error. The fix was straightforward, and the certified result was never in doubt. But the story doesn't end there.

Six weeks later, on December 17, auditors hand-counted every single presidential ballot in the county — all 15,700 of them — to put the matter to rest. When they finished, the hand count differed from the machine tabulation by about a dozen votes.

Twelve votes. In a county where the error was obvious.

Now ask yourself: what happens when the error is subtle, the race is close, and a dozen votes is the entire margin?

The gold standard has a tolerance range

"We counted it by hand" carries enormous psychological weight. It conjures images of careful humans, eyes close to paper, immune to the glitches and manipulations we fear from machines. It sounds like bedrock. It sounds like proof.

It isn't — or more precisely, it isn't only that. It's a method with a tolerance range. And that tolerance range has a name: irreducible human error.

Hand counts introduce their own category of mistakes. A ballot lands in the wrong pile. A counter miscalls a mark. Two people read the same ambiguous pen stroke differently. A ballot with ink that didn't quite reach a bubble goes to adjudication, where a small panel of officials makes a judgment call that is, by definition, subjective. Multiply any of those possibilities across millions of ballots — or even across a tray of 150 — and the cumulative drift is not zero.

In Antrim County, with only ~15,700 ballots in a county-level race that had already been corrected, the hand count was still off by about a dozen. The Michigan Department of State noted the hand count affirmed the accuracy of the result — and they're right that it did, in the sense that no one believed those 12 votes changed the outcome. But that framing elides the harder question.

What is a 12-vote spread actually telling you?

What "a tenth of a percent" hides

Georgia's 2020 hand count is the most cited example of manual auditing at scale in American history. After the presidential election, with an extraordinarily narrow margin in a state of roughly five million voters, Georgia conducted its first statewide risk-limiting audit — which, because of how tight the race was, became a full hand tally of all 5 million presidential ballots. All 159 counties. All 41,881 batches. Done in less than six days.

The hand count confirmed the machine-tabulated outcome, with a variation of roughly a tenth of one percent between the two tallies.

That sounds extraordinarily precise. And for a race decided by tens of thousands of votes, it is more than sufficient.

But stop and do the arithmetic. A tenth of a percent of five million is five thousand ballots. The variation between Georgia's machine count and its hand count — the gold standard of audits — sits in a range that could, theoretically, absorb an entire medium-sized county's worth of discrepancy.

Now apply that to a different question: if a Georgia statewide race were decided by, say, 800 votes, would the hand count settle it? Statistically, it confirms the likely winner. But does it prove it, in the sense that no one can identify a specific, checkable mechanism by which an error in the adjudication pile might have moved the number? That is a different and harder question — and the honest answer is no.

"A hand count confirmed it" and "the Secretary of State says it's fine" are reassurances. They are not the same thing as proof that anyone, anywhere, can independently verify.

The three places human error enters a hand count

To understand why, you need to understand where the slippage actually happens. It's not random chaos — it's three specific, structural points of failure.

1. The physical sort. Counters work in teams, moving quickly. A ballot travels from scanner bin to table to pile to bundle. At each handoff, misallocation is possible. It doesn't require bad intent — a ballot with a smudge on one corner that makes it harder to see the candidate's column, a counter who's been at a folding table since 6 a.m. These are the dozen votes in Antrim County.

2. The ambiguous mark. Election law in virtually every jurisdiction defines what counts as a valid mark for a candidate. A filled oval is easy. A check mark in an oval — valid in most states. A tiny dot. A mark that's been partially erased. A cross instead of a fill. These ballots don't get auto-counted; they go to adjudication, where human beings make a call. That call is made by officials, in real time, often under time pressure, sometimes with partisan observers watching and sometimes not. It is not a coin flip, but it is not a formula either.

3. The batch-level summary. Even a hand count usually produces batch subtotals, not one giant number. Those subtotals get recorded on sheets, transferred to a spreadsheet, added up. Every transcription is a point where a digit can be transposed. This is how perfectly accurate individual counting can produce an inaccurate aggregate.

None of these failure modes requires fraud. All of them produce uncertainty — uncertainty that, in a race decided by a handful of votes, cannot be resolved by doing the same hand count a second time.

When "close enough" isn't

Here is the tension at the heart of election auditing: the closer the race, the more you need the audit to be precise, and the more the audit's own imprecision starts to matter.

Colorado pioneered the risk-limiting audit in 2017 — a genuinely important step forward. Unlike a fixed-percentage spot check, a risk-limiting audit scales the sample size to the margin: if a race is close, you examine more ballots, because more evidence is required to be statistically confident. This is sound methodology, and it represents a real improvement over older audit regimes.

But a risk-limiting audit is built on a mathematical model that asks: given the reported margin, how many ballots would we need to hand-examine to achieve a stated level of confidence that the reported winner actually won? The answer depends on what error rate you assume the hand-count itself has. Most RLA frameworks assume hand-counting is accurate. If it isn't — if the hand count itself carries a systematic drift in one direction during adjudication, or if batch recording errors skew aggregate totals — then the statistical foundation shifts.

This is not an argument against risk-limiting audits. They are far better than what most jurisdictions do, which is nothing statistically rigorous at all. But it is an argument for being precise about what they establish: a well-executed RLA tells you the reported winner probably won. It does not generate a record that any citizen can sit down with and independently re-derive the total from first principles.

That is the gap.

What "non-binding" actually means in practice

Most audit regimes in the United States — and in many democracies globally — share a quiet structural flaw: even when they find a discrepancy, they are not required to do anything about it.

An audit whose results cannot trigger a mandatory remedy is, in a precise sense, a measurement tool without a correction mechanism. It tells you the thermometer reads 103 degrees. It does not, by itself, require anyone to give the patient medicine.

In the majority of U.S. states, audit statutes set a threshold — if the hand count differs from the machine count by more than X percent, a further investigation is triggered. But the threshold is often set wide enough that the kind of drift documented in Antrim County — a dozen votes — would never cross it, even in a race where a dozen votes is the entire margin. See how this gap plays out across jurisdictions.

The result is a system that can absorb meaningful discrepancies without mandatory consequence, then certify results that technically comply with every audit requirement, while leaving the underlying question — did the right candidate win? — unanswered in any independently verifiable way.

Officials who say the audit confirmed the result are not lying. They are accurately reporting that the audit found a discrepancy within the tolerance the law sets. What they cannot say — because the current system cannot produce it — is: here is a complete, machine-readable record of every ballot's interpretation, available for any citizen to download and check.

The verification that doesn't exist yet

Think about what genuine verifiability would look like. Not "trust the hand count." Not "trust the machine." Not "trust the Secretary of State's press release."

Instead: imagine that the moment each ballot batch is closed, a cryptographic hash of its contents is published to a public ledger — immutable, timestamped, downloadable. Imagine that every adjudication decision is logged with the specific ballot image and the panel's ruling, available for public inspection within 24 hours. Imagine that any citizen — a high school math teacher, a retired engineer, a political science graduate student — can download the complete precinct-level, batch-level results in a standard format and run the addition themselves.

That system would not eliminate human error. It would make human error visible and attributable. An error you can see and trace is categorically different from an error that is absorbed into an aggregate total and certified away.

The Sarasota County, Florida case from 2006 showed the worst version of the opposite: 18,000 undervotes in a race decided by 369, on paperless machines that produced nothing to recount. The hand count that would have settled Sarasota's result was structurally impossible because there was no paper. Georgia 2020 was better — it had paper, and a hand count could happen. But "better than paperless" is not the finish line.

The finish line is a system where no one has to trust the officials, the vendor, or the machines — because the evidence is public, machine-readable, and independently verifiable.

What you still cannot check

Right now, after every major election in nearly every U.S. jurisdiction, here is what remains outside your reach as a citizen:

  • The original interpretation of each ambiguous ballot, logged against a ballot image you can see.
  • A machine-readable batch-level subtotal file you can aggregate yourself to re-derive the county total.
  • An independent confirmation that the paper ballots physically present match the ballots the scanner recorded — with a cryptographic commitment made before the election that prevents the paper from being quietly altered.
  • Any mechanism by which a discrepancy within the legal tolerance automatically triggers a mandatory, public, time-bounded investigation.

These are not fantasy features. The cryptographic building blocks exist. Some jurisdictions publish precinct-level results in downloadable formats. The EAC recommends machine-readable reporting. The tools are real. The political will to require them universally is what's missing.

Until then, every "the hand count confirmed it" is partly true and partly a polite way of saying: we looked carefully, by the methods we have, and the answer seems right. In a race decided by fewer votes than the method's own tolerance range, that is not confirmation. That is a statement of faith dressed in the language of evidence.

The goal isn't to distrust elections. It's to build the systems that make trust unnecessary — because the proof is already there, in public, for anyone to check.

Explore the audit-gap map across every U.S. state and dozens of countries — or read the two-minute version of why verifiability matters.

Sources

See all postsExplore the gaps