← All posts

What actually triggers a recount, and why the threshold matters

A 30,000-vote margin got an election annulled. A 537-vote margin nearly paralyzed a country. The number that triggers a recount may be the most consequential threshold in democracy — and almost nobody set it on purpose.

It is the morning of December 12, 2000, and somewhere in Tallahassee, a recount that has consumed the United States for thirty-six days has just been stopped by a 5–4 Supreme Court decision. The margin in Florida: 537 votes out of nearly six million cast. A tenth of a hundredth of a percent.

No one had designed Florida's recount law for a margin that small. The state had a threshold — a half-percent gap triggered an automatic machine recount — but the law said almost nothing about what should happen next, what standards the hand counters should use, or who had the final word. The result, as the Supreme Court found in Bush v. Gore, was chaos: identical ballots evaluated differently in different counties, different rules applied by different teams within the same county, the whole process improvised in real time under television lights and legal pressure. The Court struck the recount down not because it found fraud, but because there was no uniform, pre-defined standard for reading a ballot — and without one, you cannot have a verifiable count at all.

What's remarkable, looking back, is not that this happened. It's that the trigger — the threshold that set the machinery in motion — had never been calibrated against the accuracy of the counting method itself.

That gap is still open today.

The number on the law, and the question nobody asked

Most democracies that have automatic recount thresholds set them as a fixed margin or a fixed percentage: half a percent, one percent, some round number that feels tidy. The United States has a patchwork of them. Austria, as of 2016, had none at all for presidential elections — the Constitutional Court could annul a result, but there was no automatic trigger that said a margin this small means you must re-examine the count before certifying.

Here is the question those thresholds almost never answer: how accurate is the counting method itself?

This is not a philosophical question. It is an engineering question. And the answer to it determines whether a recount threshold means anything.

If your counting method — machine scanning, hand counting, whatever it is — has an irreducible error rate of, say, one in a thousand ballots, then a threshold set at one-tenth of one percent is asking a method with a known error floor to adjudicate margins smaller than its own noise. The trigger fires. The recount runs. And the result is still not verifiable, because the tool you are using to verify it is itself less precise than the gap you are trying to measure.

This is not hypothetical. It is what happened in Antrim County, Michigan in 2020: a hand audit tallied every presidential ballot by hand to fix a known error, and the final hand count still differed from the machine tally by about a dozen votes out of roughly 15,700 cast. Twelve votes. No fraud, no bad faith — just the irreducible reality that human beings sorting paper at speed do not all see the same thing every time. In a race decided by twelve votes, that spread is the whole election.

Austria 2016: when the margin fits inside the procedural gap

On May 22, 2016, Alexander Van der Bellen defeated Norbert Hofer in the second round of Austria's presidential election by roughly 30,000 votes out of about 4.6 million cast — a margin of less than one percent.

Thirty thousand votes sounds like a comfortable cushion. It is not, once you know how the ballots were handled.

Austrian law is specific: postal ballots must be opened and counted exclusively by the district electoral authority, and not before 9:00 a.m. the morning after Election Day. The rule exists for a reason. It creates a verifiable chain of custody. It ensures witnesses are present. It makes the count a supervised, documented act that any party can observe.

The Austrian Constitutional Court found, in its decision of July 1, 2016, that in multiple districts those rules had been broken: ballot envelopes were opened early, votes were counted before the legal hour, and in some places unauthorised persons handled the ballots without required witnesses present. Approximately 77,000 votes were affected by proven procedural violations.

The winning margin was about 30,000 votes.

The violations affected 77,000 votes.

The Court expressly found no evidence of actual fraud. Not a single vote was shown to have been altered. But that was beside the point. The point was that the procedural violations were large enough — in absolute number — that the Court could not say the result was untainted by them. The chain of custody had been broken in numbers that exceeded the margin of victory. The election was annulled, and a fresh vote was held in December 2016.

This is the core logic of recount law made visible: it is not enough to ask whether fraud occurred. You have to ask whether the procedures were sound enough to prove fraud could not have occurred. And if the procedural gap is larger than the winning margin, the answer is no.

What "close" actually means in the law

Different countries have written their close-margin rules in radically different ways, and almost none of them are calibrated to what we actually know about counting accuracy.

In the United States, automatic recount thresholds vary by state. Some use a half-percent margin. Some use a fixed vote count. Some give candidates the right to request a recount within a certain window, with or without a threshold. The National Conference of State Legislatures tracks these rules, and the variation is substantial enough that the same margin in the same race, run across two different states, might trigger an automatic recount in one and nothing at all in the other.

None of those thresholds were set by asking: given our counting technology, what is the smallest margin we can confirm with confidence?

Colorado came closest to asking that question when it became the first U.S. state to run a statewide risk-limiting audit in 2017. A risk-limiting audit does something that fixed-percentage audits do not: it scales the checking to the margin. Close race? You check more ballots. Lopsided race? You check fewer. The statistical confidence you are buying is the same either way — a defined probability that you would catch a wrong outcome if one existed. That is what a recount threshold should be doing: buying confidence proportional to what is at stake.

But even a risk-limiting audit only delivers on its promise if the ballots it is sampling from are trustworthy. Which brings us back to the method.

The hand count is not the gold standard everyone thinks it is

There is a comforting assumption embedded in most recount law: if machines are uncertain, humans can settle it. Count the paper by hand and you will know.

The evidence does not support this as cleanly as people assume.

Georgia's 2020 statewide hand count — all roughly five million presidential ballots, examined in 41,881 batches over less than six days — confirmed the machine-tabulated outcome to within about a tenth of one percent. That is an impressive result, and it is evidence for the value of paper plus audit. But notice what it cannot tell you: in a race decided by hundreds of votes rather than tens of thousands, is a tenth of a percent close enough?

Consider the arithmetic. Georgia cast roughly five million presidential ballots. A tenth of a percent is about five thousand votes. The 2020 Georgia presidential margin was about 11,779 votes — well above that variation. The hand count was confirmatory because the margin was large enough to absorb the noise.

Now shrink the margin to, say, three hundred votes. The hand count's inherent variation — human misreads of ambiguous marks, adjudication calls where two counters disagree, ballots that fold or smudge — does not shrink with it. A hand count that varies by even a dozen votes from the machine tally is performing well, technically. But in a race decided by fewer than a dozen votes, that irreducible spread is not a rounding error. It is the entire disputed territory.

Windham, New Hampshire showed this in miniature. A forensic audit in 2021 traced a systematic miscount to fold lines on absentee ballots — no fraud, no malware, just a crease the scanner read as a filled oval. The investigators found that in some tests, about 44 percent of ballot folds through target areas were read as votes. The error was systematic and mechanical and entirely invisible in the machine output until a human looked at the paper.

The lesson is not "therefore hand counts are wrong." The lesson is that every counting method — optical scanner, hand sort, whatever you deploy — carries its own error profile, and a recount threshold that is smaller than that error profile cannot settle a close race. It can only restart it.

The gap between the threshold and the standard

Here is the structural problem.

Recount thresholds are set by legislatures, often in round numbers, often decades ago, often with no reference to the accuracy of the counting technology in use at the time. They are rarely updated when the technology changes. They are almost never benchmarked against empirically measured counting error rates.

The result is a system where a margin of 537 votes can trigger a recount conducted without uniform standards — as in Florida 2000. Where a margin of 30,000 votes can still be legally unverifiable because the procedural violations affecting the count exceeded it — as in Austria 2016. Where a hand count meant to confirm a corrected machine tally still differs by a dozen votes — as in Antrim County 2020.

None of those outcomes required bad actors. They required only the collision of a margin that happened to be smaller than the system's known or unknown error.

Bush v. Gore made the equal-protection argument: the same ballot counted differently in two counties is a constitutional problem. But there is a deeper, pre-constitutional argument: if you cannot define what it means to have counted correctly, you cannot define what it means to recount. A recount is a check on the count. For a check to be meaningful, the method doing the checking has to be more accurate than the error it is checking for. And the threshold that triggers the check has to be calibrated to that accuracy, not plucked from political negotiation.

Germany's Constitutional Court got to this logic from a different direction in its 2009 ruling: elections must be checkable by ordinary citizens without specialist knowledge. If only an expert can evaluate the count, it is not publicly verifiable. The same principle applies to recount law. A threshold that triggers a process no member of the public can independently evaluate — because the standards are undefined, the error rates are unpublished, or the ballots are unreadable by human eye — is not a safeguard. It is a ceremony.

What a genuinely calibrated threshold would look like

A threshold set for verifiability, rather than administrative convenience, would do three things.

First, it would be tied to a published error rate. Every counting technology — optical scanners, hand sorting, ballot-marking devices — has a measurable accuracy profile under real-world conditions. That profile should be public, regularly updated, and directly connected to the margin at which a recount is triggered. If your scanner has a known miscounting rate of one in ten thousand ballots, a threshold set at one in twenty thousand is meaningful. A threshold set at one in a million is not.

Second, it would trigger a method that is statistically capable of settling the margin. A risk-limiting audit does this: it sizes the hand review to the margin and the stated confidence level, so the audit is doing proportional work. A fixed-percentage recount does not. It is equally rigorous — or equally perfunctory — in a race won by one vote and a race won by fifty thousand.

Third, it would produce a result that any observer can independently verify. This means precinct-level, ballot-by-ballot records published in machine-readable form the moment they are final — not days later in a PDF. It means a uniform, pre-announced standard for adjudicating ambiguous marks, written down before counting begins, not invented under pressure during a legal fight. And it means the ballots themselves must be human-readable: a paper record that only tabulates as a QR code — as the federal court in Curling v. Raffensperger found is the case with Georgia's ballot-marking devices — cannot be independently verified by the voter who cast it, and cannot serve as the meaningful check a recount requires.

The shareable takeaway

Here is what 537 Florida votes, 30,000 Austrian votes, and a dozen Antrim County votes all share: they exposed the same design flaw. The threshold that was supposed to trigger verification was not connected to the accuracy of the verification method. The recount procedure was not calibrated to the margin. And the public had no way to check the result independently.

A recount law that does not answer the question "how accurate is our counting method, and is this margin smaller than its error?" is not a safeguard. It is a promise the system cannot keep.

The fix is not simply a lower threshold or a higher one. It is building elections so the count is verifiable at the precinct level — in real time, in human-readable form, against ballots that any observer can inspect — before anyone has to decide whether to trigger a recount at all. At that point, a close race does not wait for a threshold to fire. The evidence is already public. Anyone can check it.

That is the standard no official reassurance can substitute for.

See how close-margin triggers work — and where they break — across different election systems →

Explore the full gap database →

Two-minute version →

Sources

See all postsExplore the gaps