Nearly 20 years ago, the nation’s eyes were transfixed on a contentious Florida election recount to determine the winner of the presidential race. That recount was cut short by a U.S. Supreme Court ruling that even today has left many wondering who really won.

This week, the nation’s eyes (and the president’s tweets) are focused on another contentious statewide Florida recount, this one involving the U.S. Senate race between Democratic incumbent Bill Nelson and his Republican challenger, Governor Rick Scott. Although two other statewide races are also under recount — the gubernatorial race and a contest for agriculture commissioner — the U.S. Senate race has drawn the most acrimony, attention, and legal action, since a win for Scott would help Republicans maintain their grip on the Senate.

Florida’s secretary of state ordered machine recounts in all three of these statewide races due to narrow margins. The deadline for completion was supposed to be Thursday afternoon, but a judge has ordered an extension to Nov. 20 for Palm Beach County. Other counties have complained they cannot complete the process by Thursday.

Once the machine process is done, counties could be instructed to conduct a hand recount of some ballots, depending on the margin. Although Florida law requires election results be certified by Nov. 20, this could be delayed if machine and hand recounts are not completed by then.

“To call it a recount is false.”

But even as the state’s 67 counties scramble to finish the machine recount of more than 8 million ballots, it’s not clear the results will give a true picture of the race’s winner. That’s because the method Florida uses to conduct election recounts is not a true recount of voter ballots, but simply a rescan of ballots through the same machines that initially counted them. If problems with the software — either through glitches or hacking — produced faulty results the first time, they will reproduce the same faulty results during a rescan.

“To call it a recount is false,” says Ion Sancho, who for 27 years was supervisor of elections for Leon County, Florida, and who led the manual recount in Miami-Dade County for the 2000 presidential recount. “It’s a failure in the state of Florida that the manual recount is not a recount; it’s a scan of ballots.”

Dan McCrea, president of Florida Voters Foundation, calls it a “faith-based” rather than “evidence-based” election. “[Election officials] just sort of trust the machines, trust the vendors, trust insiders,” he says, “and we shouldn’t have to [do that].”

Under Florida law, automatic machine recounts kick in when the margin of results in a race is half a percent (0.5) or less. The state then does a second manual recount — visual inspection of the ballots by human workers — if the machine rescan results in a margin that is one quarter of a percent (0.25) or less. Even then, the physical inspection doesn’t involve every ballot. It will only cover ballots the machine spits out because it detects evidence of an “undervote” or “overvote” in a race that’s being recounted.

An undervote is when a machine senses no vote marked in a race. This may be because the voter intentionally left it blank or overlooked the race. Overvotes occur when the machine senses a voter has selected more candidates or options than a particular race allows.

Machines can make mistakes in both cases. Stray marks on a paper ballot due to poor printing or from a voter’s pen can be interpreted by the machine as an overvote, for example. A machine can indicate a race has no vote, or an undervote, if the voter filled out the ballot incorrectly by circling a candidate’s name instead of filling in the oval, or by failing to completely shade the oval in.

Dan McCrea, president of Florida Voters Foundation calls the process faith-based elections rather than evidence-based elections.

Already, in Broward County, a traditionally Democratic stronghold, some 30,000 ballots have shown no vote cast in the U.S. Senate race, which some have attributed to a confusing ballot design that caused voters to overlook the race. Nelson’s attorney has dismissed this as the cause and believes it’s due to a problem with the machines or with how voters marked the ballots.

In Florida, teams of workers will inspect the overvote and undervote ballots in a recount race to see if the machine has missed legitimate votes, and consult a canvassing board to adjudicate in cases where they can’t determine a voter’s intent. But this means the machine is trusted to determine voter intent with all other ballots that aren’t believed to contain an undervote or overvote. Even the undervote and overvote ballots won’t get visually inspected if the rescan of ballots doesn’t narrow the margin in a recounted race from 0.5 percent to 0.25 percent.

“The way the Florida law is set up is that it’s highly unlikely that you’ll get to voter intent, assuming the retabulation [in the initial recount] will just be a regurgitation of the first time around,” says Mark Halvorson, founder of Citizens for Election Integrity Minnesota, whose group maintains a database that tracks recount laws around the country.

Scott led Nelson by just 0.18 percent before the recount began in the U.S. Senate race, which means that the overvote and undervote ballots in his race will likely be inspected by hand.

McCrea says his group isn’t looking to have Florida manually inspect every ballot in a recount, as long as counties also do a meaningful audit of paper ballots against machine tallies.

Audits are different from recounts. A recount retallies votes on all of the ballots in a race to determine if the initial tally is correct; an audit only examines a sample of those ballots to see if the system that’s doing the counting and recounting is accurate.

“Machines are fine [for counting],” he says. “But machine counting needs to be audited against those [paper] ballots. So you have to look at the ballots to confirm that the machines are counting correctly, and that step doesn’t exist in Florida elections.”

Combining a manual audit of a statistically significant number of ballots with a machine recount would provide evidence that the ballots were recorded correctly. It would help ensure the recount results are accurate, McCrea says.

Florida’s problem, in this regard, is the same as the rest of the nation’s, because many states conduct recounts — and even audits in some cases — by simply rescanning ballots through the same machines that initially scanned them. Or even if they manually examine ballots in an audit, they only do so with a small percentage of paper ballots from a random sampling of precincts. This method won’t detect machine problems if they occurred in precincts where the ballots aren’t manually examined. Following the 2016 election, this issue came up during recounts in Wisconsin, when some counties did hand recounts of votes cast in the presidential race and some simply reran the paper ballots through optical-scan machines.

If problems with the software — either through glitches or hacking — produced faulty results the first time, they will reproduce the same faulty results during a rescan.

So far, Colorado is the only state that has experimented with a suitable auditing method designed to catch hacking or machine problems, called a risk-limiting audit. This audit takes a percentage of paper ballots from every precinct in a county and manually compares them to the digital tallies. The number of ballots manually examined changes depending on how close a race is.

With regard to recounts, Sancho thinks there’s a better way to do them than either hand counting ballots or rescanning them. He thinks counties and states should be using a method he used in Leon County beginning in 2012. It involves a system called ClearAudit, which uses digital images of the ballots to do recounts.

With ClearAudit, the ballots are first scanned through an optical-scan voting machine, then through a second machine that has the ClearAudit software on it. This needs to be done with a commercial, off-the-shelf scanner so the process is independent of any software made by the voting machine vendor. In many cases, voting machines don’t produce a digital image of the ballot, just a digital tally of the votes on the ballots. The digital ballot images produced by ClearAudit let officials see what the voter saw after marking their choices on the paper ballot.

“In the ClearAudit system, you can see all of the ballot,” he says. “You can see all extraneous marks, so if the voter put an arrow… pointing to a candidate [instead of filling in the oval], you can see that. It’s not just focused on the oval the way the scanner is.”

Election integrity activists warn that because the digital images are made by software, they can be altered, as well. But Sancho says he would never use ClearAudit digital images without also doing a manual audit of those images against the paper ballots to make sure the ClearAudit software hadn’t been subverted.

Sancho says at least seven Florida counties, including his former county, already use the ClearAudit system for auditing elections. But even though Broward County, where much of the controversy over the U.S. Senate race is focused, also has the ClearAudit system, it doesn’t use it for recounts, he noted. Broward only uses it to reconcile ballots when officials have to duplicate ballots that are creased or otherwise damaged.

“[ClearAudit] is the least screwed up method I’ve seen for discerning the intent of the voter on paper ballots,” he says.

McCrea says ClearAudit is fine for doing recounts as long as election officials conduct a robust and meaningful audit of the ballot images first to make sure they haven’t been altered.

“In other words, you could check the paper records to confirm the [ballot] images are correct to a very high degree or to a very cursory degree. [But] the audit of the images must be to a high standard or it tips back into the useless column. So, the devil is in the details.”