← All posts

A folding machine, some dust, and hundreds of miscounted votes — no fraud required

A folding machine left a crease in the wrong place, dust settled on a lens, and hundreds of validly cast votes quietly disappeared — with no fraud, no malware, and no one to blame.

It is the evening of November 3, 2020, and the polls have just closed in Windham, New Hampshire — a small town of about 14,000 people in Rockingham County. The workers feed the ballots through the AccuVote optical scanners, the tape prints, and the machine totals are what they are. Four Republican state House candidates each receive roughly the same number of votes. Nothing looks obviously wrong.

Then the hand recount happens.

When officials count the same ballots by hand, every one of the four Republicans gains votes — roughly 300 more apiece. The largest single-candidate swing is about 373 votes. In a contest where candidates received a few thousand votes each, that is not a rounding error. That is a different election.

No fraud was ever alleged. No malware was ever found. The machines were certified. The ballots were real.

So what happened?

A crease in the wrong place

The New Hampshire legislature ordered a forensic audit under SB 43. The team — Harri Hursti, Mark Lindeman, and Philip B. Stark — issued their report in July 2021.

The answer, once they found it, was almost absurdly mechanical.

The town had leased a folding machine to stuff absentee ballots into envelopes. The machine worked fine for envelopes. The problem was that it folded the ballots along lines that did not match the printed score marks — which meant the folds ran directly through the oval vote targets on a significant number of ballots. When the AccuVote scanners read those ballots, a substantial fraction of the fold creases were interpreted as marked ovals.

The audit team estimated that approximately 44 percent of folds through vote targets were read as votes in November. Experimental testing showed rates ranging from about 20 percent to more than 72 percent. A second contributing factor: a build-up of white powder inside the scanners had partially obstructed the lenses, making the machines even more sensitive to stray marks.

No tampering. No conspiracy. A folding machine, some dust, and hundreds of miscounted votes.

The hand recount recovered voter intent that the scanners had silently mangled. The ballots existed, intact, and could be read by human eyes. That is the only reason the error was ever caught at all.

Read the full SB 43 forensic audit report | NH Secretary of State audit page

'No evidence of tampering' is a different sentence from 'correctly counted'

Here is the thing worth pausing on.

After the audit, officials said — accurately — that there was no evidence of fraud and no malware on any tabulator. And they were right. The forensic examination of the paper, printing, and marking found nothing anomalous. The machines had simply done what optical scanners do: read light and dark, interpret a crease as a filled oval, and record a result. Mechanically, they were working as designed.

But the count was still wrong by hundreds of votes.

This distinction matters enormously, and it is not a technical quibble. "No evidence of tampering" is a statement about intent. "No miscount" is a statement about accuracy. They are not the same sentence, and treating them as equivalent is how a genuine error can pass unscrutinized.

When officials say an audit "confirmed the result," they are usually telling you something true and something incomplete at the same time. In Windham's case, the hand recount did not confirm the machine result. It corrected it. The corrected totals did not change who won the overall races — but Windham is a small town in a multi-member district. The same mechanical failure scaled to a larger jurisdiction, a closer race, or a single-seat contest could produce a different outcome entirely.

The lesson of Windham is not that scanners are unreliable. It is that scanners are machines, and machines fail in quiet, non-dramatic ways that are only catchable if you build in the means to catch them.

The paper ballot is the detection mechanism, not just the vote

This is why the existence of a voter-marked paper record matters in a way that is often misunderstood.

People tend to frame "paper ballots" as a safeguard against hacking — as if the point is to have a physical backup in case someone remotely alters the digital tally. That framing is true, but it undersells the point. In Windham, there was no hacking. The threat was mundane: a fold in the wrong place.

Paper ballots are the detection mechanism for every category of machine error, not just intentional ones.

Dust on a lens. A calibration drift. A firmware bug that misreads a particular font size. A folding machine that scored the wrong crease. None of these require a bad actor. All of them are catchable — but only if you routinely check the machine's output against the physical ballots.

The hand count in Windham was triggered because the discrepancy was large enough to generate a formal challenge. That means the process worked, eventually, in this case. But consider what that actually required: a large enough discrepancy to be noticed, a candidate willing to press for a recount, and a legislature willing to fund a forensic audit. Smaller errors in less-watched races don't always clear all three of those bars.

What a mandatory audit regime actually closes

The Windham case ties directly to a gap that exists in most American election law: there is no universal federal requirement to conduct routine, independent hand audits of machine-counted ballots.

Colorado became the first state to run a statewide risk-limiting audit after its 2017 election — a method that statistically sizes the hand-review sample to the margin, checking more ballots in close races and fewer in lopsided ones. Colorado's own documentation explains the logic: an RLA provides defined statistical confidence that the reported outcome is correct, rather than just spot-checking a fixed percentage.

Georgia conducted a full hand tally of approximately five million presidential ballots in 2020. The variation between that hand count and the machine total was about a tenth of one percent — which confirmed the outcome with a level of evidence that no official press release can provide. Those are very different things. One is a number. The other is a checkable process.

But neither Colorado's RLA law nor Georgia's post-2020 audit requirement is the national norm. Most states have some form of post-election audit requirement, but the standards vary enormously — in method, sample size, trigger conditions, and whether results are published at a granular enough level for independent verification. The National Conference of State Legislatures tracks how states report results, but tracking that they report is not the same as confirming what they audit.

The gap is not between states that trust machines and states that don't. The gap is between states where a Windham-style error would be caught routinely and states where it would have to be noticed by accident.

The geometry of the problem: how small do errors need to be to matter?

In the Antrim County, Michigan hand recount of December 2020 — a recount conducted specifically to confirm the machine totals after a known error was corrected — the final hand count of roughly 15,700 presidential ballots still differed from the machine tabulation by about a dozen votes. Officials presented this as confirmation of accuracy. And at that scale, in a race that was not close, twelve votes is noise.

But consider the geometry. A hand count that is itself off by a dozen votes cannot definitively settle a race decided by fewer than a dozen votes. This is not an argument that hand counts are useless — they are far more reliable than machine-only totals with no check at all. It is an argument that even the gold-standard manual method carries irreducible human error: fatigue at 2 a.m., ambiguous marks, judgment calls about voter intent.

When you hear "a hand count confirmed it," the honest follow-up question is: confirmed it to within what margin? And is that margin smaller than the race's deciding margin? In Windham's case, the answer was yes — the hand count found hundreds of votes the scanner had lost. In Antrim's, the residual difference was small relative to the outcome. But neither jurisdiction published the methodology in a form that lets an independent observer verify the claim from the outside.

That is the structural problem. It is not about whether any particular official is honest. It is about whether the process produces evidence that anyone — not just the official conducting the audit — can independently check.

What would actually fix this

The Windham audit team's report is a model of how post-election forensics should work: transparent methodology, published data, named researchers with checkable credentials, a public document anyone can read. The problem is that it took a legislative mandate, significant resources, and a large enough anomaly to trigger it.

A routine audit regime — one that is mandatory, statistically rigorous, precinct-level, and publicly reported in machine-readable form — would catch Windham-style errors as a matter of course, not as a post-hoc investigation after someone notices something is off. The German Federal Constitutional Court put the principle plainly in 2009 when it struck down electronic voting machines whose results a citizen could not verify without specialist knowledge: the count is only legitimate when ordinary people can check the essential steps. A fold crease misread as a vote, caught only because a candidate demanded a recount, is not ordinary people checking the essential steps. It is a near miss.

The U.S. Election Assistance Commission published guidance on election results reporting that urges officials to make results downloadable in common formats and to specify which ballot types each report includes. That is a floor, not a ceiling. Precinct-level, machine-readable, near-real-time reporting — the kind where a spreadsheet downloaded on election night matches what is certified a week later — is already technically feasible. Most states do not yet do it.

And without mandatory audits against the physical paper, none of it matters. The scanner tape and the hand count have to agree, and when they don't, the discrepancy has to be publicly explained, not quietly rounded to "confirmed."

The question that should follow every audit announcement

The next time you read that a post-election audit "confirmed the result," here are the things worth asking that the press release will not tell you:

What method was used, and what is the known error rate of that method? How large was the residual discrepancy between the machine count and the hand count, and how does that compare to the race margin? Was the methodology published in enough detail for an independent observer to replicate it? And were the ballot-level results released in a format anyone can download and check?

"The Secretary of State says it's fine" is a data point, not a proof. It is a claim. The evidence behind it may be excellent — or it may be a tape that matched another tape, both produced by the same scanner with dust on its lens.

Windham's scanners were certified. They passed their pre-election tests. No one tampered with them. And they still miscounted hundreds of votes until a human being picked up the paper and read it with their own eyes.

The fix is not distrust. The fix is a system designed so that distrust is unnecessary — because anyone can check.

Curious how widespread the mandatory-audit gap is across U.S. states and around the world? See the TrustVoting atlas. Want the two-minute version of why paper plus audit is the baseline? Read the overview. For the full gap analysis by jurisdiction, start here.

Sources

See all postsExplore the gaps