Start Here
Donate

Case 5 — Baldwin County, 2002

5.1 — Introduction

On the night of November 5, 2002, Don Siegelman won the Alabama governor’s race. He gave a victory speech. The Associated Press called the race in his favor. His campaign staff went home. Then, after the Democratic county chair in Baldwin County had been told the courthouse was closing and had left, election officials reopened the courthouse and continued counting.

By morning, approximately 6,334 votes had vanished from Siegelman’s column in Baldwin County. Bob Riley led statewide by 3,120 votes out of 1,364,602 cast — a margin of 0.22%. Siegelman was no longer the governor-elect.

The explanation was a “computer glitch” at a single precinct: Magnolia Springs Wesleyan Church, where the initial count had shown roughly 7,000 votes at a location where only about 1,300 people actually voted. But an Auburn University political scientist, analyzing the data, found something that didn’t fit the glitch narrative: the initial results — the ones with the inflated number — fit the expected statewide pattern. The corrected results made Baldwin County the most extreme statistical outlier of any county in the state. This is the opposite of what a genuine error correction would produce.

Democrats demanded a recount. The Republican Attorney General blocked it. The Alabama Supreme Court blocked county-level recount attempts. Thirteen days after the election, Siegelman conceded. The ballots were later destroyed by law.

No forensic audit was ever conducted. The question was never answered. The evidence is now gone.

5.2 — Background

The 2002 Alabama governor’s race pitted incumbent Democrat Don Siegelman against Republican Congressman Bob Riley. Alabama elections are administered at the county level, with county canvassing boards responsible for tabulating and certifying results. Baldwin County — a large, predominantly Republican county on the Gulf Coast, with its county seat in Bay Minette — used voting equipment manufactured by Election Systems and Software (ES&S).

The tabulation system worked as follows: voting machines at each precinct read paper ballots and recorded the results to a cartridge — a removable memory device similar to a flash card. At the close of polls, precinct workers generated a printout showing the vote totals for that location. The cartridges were then physically transported to the Baldwin County Sheriff’s Office in Bay Minette, where they were inserted into a reader attached to a central tabulating computer. The computer transferred the data from each cartridge to its hard drive, then ran a program to produce summary results for the county.

Alabama in 2002 had no provision for automatic recounts in close elections. In twelve other states, a margin of 0.22% would have triggered one. Under Alabama law, even a successful recount could not change the outcome of the election unless the state Legislature intervened — and the next legislative session did not begin until January 14. Democrats controlled both chambers.

5.3 — What Happened at Magnolia Springs

The controversy centers on one precinct: Magnolia Springs Wesleyan Church, in a small, heavily Republican community west of Foley in southern Baldwin County.

When the unofficial results were compiled on election night, Magnolia Springs showed approximately 7,000 votes cast. The number was obviously wrong — it exceeded the precinct’s eligible voters. In the 1998 governor’s race, only 1,998 total votes had been cast there.

The final certified tally for Magnolia Springs was 1,287 votes: 910 for Riley, 342 for Siegelman, and 35 for Libertarian candidate John Sophocleus.

A critical technical detail emerged later. Mark Kelley, the ES&S representative present during the election, confirmed that the precinct-level printout generated by the poll workers at Magnolia Springs had been correct: “The memory pack from that precinct counted votes all day and in that precinct when those poll workers totaled out, they generated printouts showing correct number of votes.” The error occurred afterward — “When they plugged that memory pack in in the sheriff’s office it read 8,000 or 9,000 in that precinct.”

So the error was not in the voting machine. It was in the reading of the cartridge at the central tabulation site. The precinct-level evidence — the printout generated at Magnolia Springs — showed the true count.

Kelley added another detail: “It happened one or two other places that night and both times they said ‘that’s not right’ and popped it out and cleared the number and it read fine. They did not catch it in Baldwin County.”

The effect on Baldwin County totals was decisive. Siegelman’s count dropped from 19,070 to 12,736 — a reduction of 6,334 votes. Riley’s total was unchanged at 31,052. Probate Judge Adrian Johns, the county’s chief election official, told reporters that the revised tally was “accurate and supportable and defendable.”

5.4 — Election Night

The sequence of events on the night of November 5 is poorly documented in contemporaneous sources but has been reconstructed from multiple accounts.

Initial returns showed Siegelman ahead statewide. He declared victory. The AP called the race. In Montgomery, the state capital, a congratulatory banner went up across the street from the Capitol.

In Baldwin County, the Democratic county chair — Bill Pfeifer — was present at the courthouse as an observer of the vote count. According to Marsha Folsom, the former first lady of Alabama who served as Democratic Party executive chairwoman in 2002: “Our county chair was there. He was told [election officials] were going home and closing the courthouse. When he left, they reopened the courthouse and continued the counting. It raised a lot of questions.”

What happened next is disputed. Siegelman, in his memoir, wrote that the numbers were re-run “four minutes after everyone had left the courthouse for the night.” He was woken at 4 AM by his chief of security, who told him the GOP was “trying to steal the election.”

But years later, an ES&S spokeswoman offered a contradicting account. Meghan McCormick told a reporter that the tabulation system maintained an internal log of all activity: “We did look at the system log after the election and the only time they entered was at about 11 o’clock. There was definitely nothing that would raise any questions.” The system log was never independently verified or made public.

Sheriff’s Captain Marvin Ussery, who stood in for the sheriff (who was on the ballot) as a member of the canvassing board, said he watched the entire process: “I watched the whole time we were working with it.” Ussery said the problems became apparent when the Libertarian candidate appeared to be drawing more votes than the Democratic or Republican candidates — a clear sign something was wrong. The ballot packs were re-run and the errors caught.

When the initial printouts that had been distributed to the media proved incorrect, Ussery collected them. “I had distributed copies of the totals to the media and when we realized when the totals were rerun and explained what happened, we asked for those copies back so they wouldn’t get out and mixed up and I ran them through the shredder.” Ussery acknowledged the optics: “The shredder conjures up all sorts of things about votes.”

David Brewer, a county administrator present that night, recalled a similar moment. Captain Ussery told him to “shred everything” regarding the initial green bar printout. Brewer’s account: “It appeared the data pack read incorrectly and there was nothing nefarious. I was there.”

By Wednesday morning, Riley’s campaign was answering the phone “Gov. Riley’s office”. Dueling congratulatory banners hung in Montgomery. Siegelman was touring tornado damage in Henry County. Baldwin County DA David Whetstone said the error was “easy to spot” and that the results had been canvassed on a table outside the courthouse — “complete transparency.”

5.5 — The Recount That Didn’t Happen

Siegelman immediately sought a recount. Under Alabama law, any registered voter could request a county-by-county recount by posting a security bond of 750to750 to 5,000 per county. Siegelman planned to use Democratic voters in each of Alabama’s 67 counties to request recounts.

His campaign lawyer, Joe Espy, was careful to frame the effort: “We’re not challenging or contesting this election; all we’re asking for is a recount. This is not going to be another Florida.”

What followed was a legal maneuvering contest that the Democrats lost at every turn.

Secretary of State Jim Bennett requested county officials delay recounts pending an opinion from Attorney General Bill Pryor. Pryor, a Republican, then ruled that sealed ballots could only be opened with proof that illegal ballots had been counted or that legal ballots had not been counted in sufficient numbers to change the result. As Siegelman later described it: “You cannot challenge an election once it’s certified unless you can show it to a judge that there were enough legal votes that were not counted. But we could not do that because the attorney general would not allow a recount in this one precinct.”

Despite this, some counties cooperated. Election officials in Tuscaloosa and Russell counties granted Siegelman’s recount request. But Riley’s legal team moved quickly, requesting the Alabama Supreme Court block all recount attempts. The court obliged.

On November 18 — thirteen days after the election — Siegelman conceded in a televised address. “I’ve decided that a prolonged election controversy would hurt Alabama, would hurt the very people that we worked so hard to help.”

The ballots were subsequently destroyed, as required by Alabama law.

Pryor was named to a federal judgeship by President George W. Bush in 2004.

5.6 — The Statistical Evidence

In 2003, James H. Gundlach, a political scientist at Auburn University, presented a paper titled “A Statistical Analysis of Possible Electronic Ballot Box Stuffing, The Case of Baldwin County Alabama Governor’s Race in 2002” at the annual meeting of the Alabama Political Science Association. The paper was later published in the anthology Loser Take All: Election Fraud and the Subversion of Democracy, 2000-2008.

Gundlach’s analysis proceeded in two stages, each producing a finding that deepened suspicion about the revised Baldwin County results.

Stage 1: County-level analysis. Using data from the Alabama Secretary of State’s website, Gundlach regressed the percentage of votes for Siegelman in 2002 against the percentage in 1998 for all 67 Alabama counties. He ran this analysis twice — once with the initial Baldwin County returns and once with the revised returns.

The result was counterintuitive and significant. With the initial returns — the ones that included the inflated Magnolia Springs count — Baldwin County fit the regression line. It was not an outlier. But with the revised returns — the ones the county certified as correct — Baldwin County became the most extreme outlier of any county in the state, falling further from the regression line than any other data point.

Gundlach noted that this is “exactly the opposite of what you would expect if the changes corrected an error in the data. That is, an error usually makes the data point deviate from expected patterns and fixing the error typically moves the data point back into the pattern.” The “correction” moved Baldwin County out of the pattern, not back into it.

Stage 2: Voting-district analysis. Gundlach then extended the analysis to the voting-district level, comparing 39 comparable districts within Baldwin County to 70 comparable districts in Montgomery and Shelby Counties. He regressed the Siegelman 2002 percentage on the Siegelman 1998 percentage for each group.

Outside Baldwin County, the slope was 0.854 — meaning for every vote Siegelman received in 1998, he received approximately 0.85 in 2002. Inside Baldwin County, the slope was 0.697. The difference was statistically significant: t=6.19, p<.0001. Approximately 18% of Siegelman’s votes had “disappeared” in each Baldwin County voting district compared to the statewide pattern.

The one-fifth hypothesis. The 18% figure led Gundlach to a specific hypothesis about what had happened. If someone had moved approximately one-fifth of Siegelman’s votes to Riley in each voting district — by calculating a fifth, rounding it, and transferring it — but had made a spreadsheet copy-paste error, accidentally adding the calculated number to Siegelman’s total instead of subtracting it, the result would be an inflated first report. When the error was discovered, undoing it would produce a decline of twice the actual manipulation — because you’d reverse both the accidental addition and perform the intended subtraction.

“This is a common error created by using copy and paste to produce the invisible formulas for cells of spreadsheets,” Gundlach wrote.

If the hypothesis was correct, the true Baldwin County totals would have been Riley 27,866 and Siegelman 15,283 — producing statewide totals of Riley 669,039 and Siegelman 671,652. Siegelman would have won by approximately 2,600 votes.

Gundlach, it should be noted, disclosed that he “votes Democratic and voted for Siegelman.” This does not invalidate his methodology — the regression analyses use publicly available data from the Secretary of State’s website and are fully transparent and replicable — but it is relevant context.

His conclusion was direct: “There is simply no way that electronic vote counting can produce two sets of results without someone using computer programs in ways that were not intended.” And: “The only way we will know for sure is if the paper ballots for Baldwin County are recounted.”

They never were. They are now destroyed.

5.7 — Both Sides

Eighteen years later, al.com interviewed participants from both sides. Their accounts remained starkly different.

The Republican account was consistent and straightforward. The Magnolia Springs data pack read incorrectly at the central tabulation site, producing an obviously inflated count — more votes than eligible voters. The error was caught, corrected, and the true results certified. Ussery watched the whole process. Whetstone conducted the canvass in the open. Johns deferred to Pryor’s legal opinion on the recount but said he would have preferred one. The ES&S system log, per McCormick, showed no after-midnight access. Kelley confirmed the precinct-level printout was correct — the error was in the cartridge reading, not the voting machine.

As Johns put it in 2002: “If there is a recount, they will not change.”

The Democratic account was equally consistent. Siegelman won. The Democratic observer was told the courthouse was closing and sent home. The courthouse was reopened and the numbers were changed by Republican officials with no outside observers present. The recount that could have verified the results was blocked by a partisan attorney general. Gundlach’s statistical analysis showed the “corrected” numbers were more anomalous than the originals — the opposite of a genuine correction. Three Republican officials were alone with the tabulating machine after the first results were released.

As Siegelman put it: “I’m more concerned they fought so hard keeping us from having a hand recount. If they believed so strongly in the results they could have welcomed a hand recount and they would have worked with us to obtain one rather than oppose it.”

The ES&S account was internally contradictory. McCormick said the system log showed no after-midnight access. Kelley said the precinct printout was correct and the error occurred at the central reader. But Kelley also told one reporter “Something happened. I don’t have enough intelligence to say exactly what,” and told another “I never felt like any kind of software problem” and “there was never any analysis done after the election to determine what happened.” The vendor confirmed the error but could not explain it, and acknowledged that no post-election forensic analysis was conducted.

Jon Gray, a Republican political strategist present in Mobile on election night, offered a meta-observation: “Both sides, historically looking back, have a different version of this story. The Riley people felt that the Siegelman people would steal the vote and the Siegelman people thought [Riley’s side] would steal the vote. The race was super close.”

This case study does not resolve the dispute. The evidence is insufficient to determine which account is correct — and the evidence that could have resolved it is gone. What the case study identifies is what would have resolved it.

5.8 — A Note on Subsequent Events

Don Siegelman was indicted on federal bribery and mail fraud charges in 2005, related to his conduct as governor — not to the 2002 election. He was convicted in 2006. The prosecution was controversial: the investigating U.S. Attorney, Leura Canary, was married to Bill Canary, a prominent Republican political consultant in Alabama who had worked for Siegelman’s political opponents and later advised Riley. Former state attorneys general from both parties questioned whether the prosecution was politically motivated.

This case study does not take a position on Siegelman’s prosecution. It is a separate matter, mentioned because the two events are frequently discussed together.

5.9 — The Legislative Response

Alabama’s legislature recognized what had gone wrong. After the 2002 election, the state amended its election code to provide for automatic, supervised recounts in races decided by less than one-half of one percent — a law passed in 2003 in direct response to the Siegelman-Riley dispute.

The 2002 margin — approximately 0.22% — would have triggered an automatic recount under the new law.

The reform is an implicit acknowledgment that the system failed. But it applies only to future elections. For 2002, the question remains open — and the evidence to answer it no longer exists.

5.10 — What We Know and What We Don’t

What is established:

The Magnolia Springs precinct initially reported approximately 7,000 votes where approximately 1,300 were cast. The ES&S representative confirmed that the precinct-level printout showed the correct count and the error occurred when the memory pack was read at the central tabulation site. The same reading error occurred at one or two other precincts that night and was caught immediately. It was not caught in Baldwin County, and the inflated number was included in the initial county totals.

Siegelman’s Baldwin County total was reduced by 6,334 votes — more than double Riley’s statewide margin. The Democratic county chair was told the courthouse was closing and left before the correction occurred. The initial printouts distributed to the media were collected and shredded.

Siegelman sought a recount. The Attorney General blocked it with a legal opinion that created a Catch-22. Counties that granted the recount were overridden by the Alabama Supreme Court at Riley’s request. Siegelman conceded thirteen days after the election. The ballots were subsequently destroyed by law.

An independent statistical analysis by Gundlach found that the “corrected” results made Baldwin County more anomalous, not less — the opposite of what a genuine correction would produce. The district-level analysis found a statistically significant deficit in Siegelman’s Baldwin County votes compared to the statewide pattern (t=6.19, p<.0001). Gundlach hypothesized a specific mechanism — a one-fifth vote transfer with a spreadsheet error — that, if correct, would mean Siegelman won statewide by approximately 2,600 votes.

What we do not know:

Whether the “correction” was accurate. The Republican participants say it was — a data pack misread, easily spotted, transparently corrected. This is plausible. The ES&S system log reportedly showed no after-midnight access, supporting the idea that the correction occurred in the normal course of election-night tabulation. But the log was never independently verified.

Whether Gundlach’s hypothesis is correct. His analysis is methodologically rigorous, transparent, and statistically significant. But it is a hypothesis derived from patterns in the data, not direct observation of what happened at the tabulating computer. He acknowledged this: “The only way we will know for sure is if the paper ballots for Baldwin County are recounted.”

What the precinct-level printout at Magnolia Springs actually showed. Kelley said it was correct. If it showed approximately 1,287 votes, the “correction” was legitimate regardless of everything else. If it showed something different, the story changes. But no one preserved it.

The ballots are destroyed. The memory packs are gone. The system log was never independently examined. Gundlach’s analysis is the most rigorous evidence that exists, but it cannot substitute for the source documents.

5.11 — What Actual Vote Would Have Contributed

This may be the clearest illustration in this manual of what Actual Vote is designed to do — made even clearer by a detail from ES&S’s own representative.

Mark Kelley confirmed that the precinct-level printout at Magnolia Springs showed the correct vote totals. The error occurred afterward, when the memory pack was read at the central tabulation computer. The document that was right — the precinct-level printout — is exactly what Actual Vote photographs.

One photograph of the Magnolia Springs printout would have resolved the entire controversy.

If the printout showed approximately 1,287 votes, the “correction” was legitimate. Riley won. If it showed something materially different, the correction introduced an error rather than fixing one — and the investigation would have proceeded from documented evidence rather than statistical inference.

More broadly, poll tape photographs from all Baldwin County precincts would have provided a complete independent record. Add up the precinct-level totals from the photographs. Compare to the officially reported results. The comparison is definitive.

The AV counterfactual cuts both ways — and that is the point. If the glitch explanation is true, AV would have confirmed Riley’s victory immediately and definitively, defusing the controversy before it began. If Gundlach’s hypothesis is correct, AV would have exposed the manipulation with photographic evidence from every affected precinct. Either outcome is better than what actually happened: a question debated for over two decades, a statistical analysis that “strongly suggests” but cannot prove, and source documents that were destroyed by law.

What AV provides that didn’t exist in 2002:

Evidence generated at the precinct, before the cartridges reach the central tabulation site. The photographs would be in volunteers’ hands before any memory pack is read, before any data pack error occurs, before anyone at the sheriff’s office runs the tabulation program. The error — wherever it originated — cannot reach AV’s evidence, because AV’s evidence is captured before the data enters the system where the error occurred.

Evidence that cannot be blocked by the Attorney General. Pryor’s legal opinion was effective because the sealed ballots were the only independent evidence. Poll tape photographs are not sealed ballots. They are not in anyone’s custody. They are distributed, public, and permanent.

Evidence that survives when ballots do not. The ballots were destroyed by law. The memory packs are gone. But a photograph of a poll tape, once taken, exists in the hands of the volunteer who took it, on the platform where it was uploaded, and in any number of copies. It cannot be shredded. It cannot be destroyed by operation of law. It persists.

Gundlach concluded his paper: “The only way we will know for sure is if the paper ballots for Baldwin County are recounted.” The paper ballots are gone. If AV had existed, the answer would still be available — in the form of photographs that no one can destroy.

What AV would not catch: If the voting machines themselves miscounted votes before printing the precinct-level report — a counting-layer error — AV would not detect it. But in this case, ES&S confirmed the precinct-level printout was correct. The error occurred at the reporting layer — exactly where AV operates.

The transparency value, independent of who was right. It is worth underscoring a point that tends to get lost in controversies like Baldwin County: the transparency value of Actual Vote evidence does not depend on which of the two competing narratives turns out to be correct. In the world where the data-pack explanation is accurate and the corrected totals reflect what voters actually did at the precincts, AV photographs would have confirmed the corrected totals and vindicated the administrators who performed the correction, short-circuiting twenty years of recrimination. In the world where the correction introduced an error, AV photographs would have exposed that. Either outcome produces something the 2002 election did not produce: a rational basis, available to any citizen of any party, for knowing which of the two stories is true. That is the transparency benefit — independent of, and logically prior to, any question about accuracy. The absence of independent evidence is what made Baldwin County a permanent controversy rather than a resolvable one.

The defensive frame: a left-aligned bad-faith accusation that may or may not be bad-faith

Baldwin County belongs to a class of cases the defensive case file takes up across the broader collection: contested elections in which sustained allegations of fraud — here, allegations from the left that a Republican-aligned administration stole a governor’s race — have hardened over time into political settled belief on one side, while remaining contested or rejected on the other, with the underlying evidentiary record insufficient to definitively resolve the question in either direction.

It matters that we describe the Baldwin allegation carefully. Some elements of the Democratic-aligned account are well-documented — the timing of the post-midnight recount, the dismissal of the opposing party’s observer, Pryor’s blocking of the recount, the destruction of the ballots, Gundlach’s statistical anomaly. Other elements are inferences from those facts to a particular causal explanation. We do not characterize the Democratic-aligned account as bad-faith by default. The account contains real evidentiary substance and reflects what some observers, including academic statisticians, have concluded after looking at what evidence does survive. The case sits in the same category as the Florida 2000 dispute and the Warren County 2004 dispute: an unresolvable election in which the analytical critique on one side ranges from rigorous to bad-faith, and where the absence of independent precinct-level evidence is what permits the entire spectrum to coexist indefinitely.

What AV evidence would have contributed in any version of this case is what it contributes in every version of the analogous cases: the primary-source record at the precinct level that lets the analytical critique be tested against arithmetic rather than against the absence of arithmetic. Some elements of the Democratic-aligned account would have been validated. Some would likely have been refuted. The remaining uncertainties would have been narrowed to the specific places the evidence could not reach. The dispute would have remained substantive. It would not have remained permanent.

The collection’s companion analyses of Florida 2000 and Warren County 2004 extend the same defensive argument to elections where similar left-aligned narratives took root in similar evidentiary vacuums. The broader pattern is structural: bad-faith accusations grow most readily in the absence of independent evidence, and they grow from whichever direction has the rhetorical incentive to grow them. Baldwin’s specific direction (Democratic accusations against Republican-favoring outcome) mirrors the Antrim 2020 case and Prince William 2020 case running in the opposite direction. The structural problem — and AV’s structural defensive value — is the same in all four.

5.12 — Why It Matters

A governor’s race was decided by 3,120 votes. A 6,334-vote reduction in one candidate’s total in one county — occurring after the opposing party’s observer had been told the courthouse was closing and had left — determined who governed Alabama for the next four years. The losing candidate was denied a recount through a legal Catch-22. The ballots were destroyed by law. The voting machine vendor confirmed that a precinct-level printout existed with the correct numbers — but no one preserved it. An independent statistical analysis found the “corrected” results more anomalous than the originals. And no forensic audit was ever conducted.

Both sides told their version of this story for over two decades. Republicans said it was a data pack error — obvious, easily corrected, transparently handled. Democrats said the election was stolen behind closed doors by partisan officials who then blocked every avenue of verification. A rigorous statistical analysis supports the Democratic interpretation but cannot prove it. A system log supports the Republican interpretation but was never independently verified. The evidence that could have settled the matter — the precinct-level printouts, the paper ballots, the memory packs — is all gone.

Actual Vote exists to make sure this doesn’t happen again. The principle is simple: capture the precinct-level evidence at the source, before it reaches any system that can alter it, and put it in the hands of people who have no stake in the outcome. When that evidence exists, controversies like Baldwin County get answered — one way or the other — instead of festering for decades. When it doesn’t, questions that should have been resolved on election night remain open forever.

5.13 — Further Reading

Primary scholarly analysis

News and commentary

Memoir

  • Don Siegelman, Stealing Our Democracy: How the Political Assassination of a Governor Threatens Our Nation (2020).

Cross-references in this collection