A ballot measure killing machine

After surviving a months-long effort to derail it, the Protect Education Act was ultimately brought down with the help of artificial intelligence.

The citizen initiative to rein in Arizona’s Empowerment Scholarship Accounts, or ESAs, submitted more than 421,000 signatures — a healthy cushion to withstand legal challenges but still meet the roughly 256,000 threshold required to make it on November’s ballot.

It wasn’t enough.

The Arizona Supreme Court delivered the final blow to the campaign on Aug. 18, upholding a lower court’s ruling that left it without enough signatures to qualify.

But amid the victory laps from school choice groups, one of the most important players in the fight went largely unrecognized: a company called Signafide.

Seasoned Arizona elections attorney Kory Langhofer, who was part of the legal team that successfully challenged the Protect Education Act, co-founded Signafide nearly a decade ago.

The company uses a combination of AI and human review to comb through campaigns’ petitions and flag signatures that could be invalid for a litany of reasons.

Langhofer wouldn’t say who paid for Signafide to review Protect Education’s signatures, or how much they paid for the service.

But two deep-pocketed conservative groups — the Goldwater Institute and Arizona Free Enterprise Club — were plaintiffs in the lawsuit challenging the measure.

And Langhofer said Signafide’s going rate is about $1.50 per signature. At roughly 400,000 signatures, that would put the cost of a full review around $600,000.

Arizona is one of 24 states that lets its voters bypass their legislature to put proposed laws directly before voters through a citizen initiative process.

So after years of failed attempts at the Capitol to impose meaningful guardrails on Arizona’s school voucher program, the Arizona Education Association — the state’s largest teachers union — and Save Our Schools Arizona turned to the ballot box to force the issue.

Their Protect Education Act would have limited ESA eligibility to households earning $150,000 or less, reclaimed unused account funds and banned luxury purchases.

Those restrictions were a nonstarter for school-choice advocates and Republican political operatives who had spent years fighting to make ESAs universally available — and had no intention of surrendering that ground.

And they mounted an all-out effort to keep it off the ballot — from a decoy initiative aimed at siphoning off signature gatherers to two rounds of last-minute dealmaking at the Capitol.

But after Protect Education survived all of that, it came down to the measure’s signatures.

When citizen initiative campaigns turn in their signatures, a weeks-long process ensues to determine how many are valid.

The Secretary of State’s Office does the first pass, throwing out obviously flawed petition sheets and signatures missing required information. Then the office randomly samples 5% of the remaining signatures and sends them to county recorders to verify the signers are registered to vote in the counties they listed.

The results from all 15 counties are combined into a statewide validity rate, which is then applied to all the campaign’s remaining signatures to estimate how many are valid overall.

This year, the Protect Education Act emerged from that process with a 75% validity rate, which put it at about 293,000 valid signatures — still above the 255,949 it needed to qualify.

Throughout the official verification process, opponents are doing their own, often more aggressive review that finds reasons to disqualify signatures beyond what election officials catch.

Langhofer estimates there are about 40 ways a signature can be disqualified. And checking 40 reasons against Protect Education’s initial 400,000 signatures was “a lot of work.”

In fact, Signafide’s process for reviewing Protect Education’s signatures took “tens of thousands of hours,” he said, and 1,300 human reviewers.

The Yale Law School grad said Signafide launched in 2018, “before AI was a buzzword,” so the company instead calls its software “machine learning.” Computers flag potentially invalid signatures and petition circulators, then a human reviews each case to verify.

Signafide was also ahead of the game in responding to a major shift in Arizona election law: In 2017, lawmakers adopted a new “strict compliance” standard for citizen measures.

Before then, courts used a more forgiving standard that didn’t automatically disqualify signatures over minor technical defects, such as incomplete addresses or too many signatures on a petition sheet.

Now, those relatively trivial mistakes can make the difference between making the ballot or not. And Langhofer has built a business around finding them.

But the technology isn’t yet reliable enough to be court-admissible without the human review, he said.

In 2024, conservative groups hired Signafide to try to prevent Make Elections Fair, or Prop. 140, from reaching the ballot, where it would have asked voters to change Arizona’s primary election process. Court records show Signafide reviewed nearly 600,000 signatures at a cost of about $1 million.

The review identified nearly 43,000 voters who had signed the petition more than once, and Signafide helped compile a massive exhibit listing the signatures plaintiffs said should be tossed as duplicates.

Those exhibits offered a rare glimpse into Signafide’s process: Reviewers weren’t just comparing handwriting on petition sheets — Signafide had listed voter registration data under each entry in blue text with names and addresses to help determine whether two signatures came from the same voter.

That extra layer of analysis bothered the trial judge. He found that Signafide’s added voter data in the 2024 case made the exhibits more than neutral summaries, and excluded them from evidence. The Arizona Supreme Court later reversed that ruling.

Even with the evidence restored, the challengers couldn’t show Prop. 140 had fallen below the signature threshold. It stayed on the ballot, where voters rejected it.

But in Protect Education’s case, Signafide’s duplicate signature findings were fatal.

Signafide flagged nearly 30,000 duplicate Protect Education signatures. And those duplicates wiped out far more of the measure’s cushion than any other category of signature disqualification.

Plaintiffs flagged circulators with felony convictions, for example. Circulators must be eligible to register to vote, so a disqualifying felony can wipe out every signature that person collected.

Ultimately, a Maricopa County judge disqualified four circulators with felony convictions. Other signatures were tossed for violations like circulators listing P.O. boxes instead of residential addresses, or failing to appear in court after being subpoenaed.

But even after those losses and election officials’ initial review, Protect Education still had a path to the ballot.

If a voter is found to have signed a petition more than once, the extra signatures are tossed. Plaintiffs, aided by Signafide’s analysis, ultimately got 24,369 duplicate signatures tossed.

Langhofer described Protect Education’s duplicate rate as “a little high.” An about 5% duplication rate is typical; the citizen measure’s rate was around 8%.

The Protect Education campaign argued some duplicates may already have been reflected in the counties’ 5% sample reviews, so removing all the ones found in the plaintiffs’ analysis of the full universe of signatures risked counting the same defect twice.

But because the campaign couldn’t identify “specific signatures subtracted twice,” state Supreme Court Chief Justice Ann Timmer wrote in the final order, “its objection is that double-counting is theoretically possible, not that it demonstrably occurred.”

The state Supreme Court let the duplicate removals stand, leaving Protect Education below the number it needed to qualify.

Both Langhofer and Beth Lewis, executive director of Save Our Schools Arizona and a key force behind Protect Education’s signature drive, agree the duplicates were the “nail in the coffin.”

But while Signafide found them, Lewis points to another behind-the-scenes operation that she says helped create duplicates in the first place.

Lewis has been through the signature-gathering effort before. In 2022, Save Our Schools launched a citizen referendum to block the Legislature’s universal ESA expansion.

The group had already pulled one off in 2018, when voters rejected an earlier attempt to open ESAs to virtually every Arizona student. But that victory only killed that version of the expansion, and didn’t stop lawmakers from trying again four years later.

The 2022 effort didn’t turn in enough signatures. But this year, Lewis said, there was more public buy-in.

Lewis’ past campaign experience also taught her to “expect shenanigans” from opponents determined to protect universal ESAs. But she wasn’t expecting the political maneuver quietly orchestrated by the chief of staff to Arizona House Republicans.

A competing voucher reform measure, Fortify AZ, announced its campaign just days after the union-backed Protect Education Act, and presented itself as a reasonable alternative that wouldn’t pose an “existential threat” to Arizona’s ESA program.

Fortify had significant financial backing: $5.4 million from the American Federation for Children, a national school-choice organization. And the Protect Education Campaign later alleged in a cease-and-desist letter that Fortify used a “bounty system” to poach circulators by offering higher hourly rates.

House Chief of Staff Grant Hanna later told us he had been behind the campaign from the start.

Hanna’s competing measure created chaos on the ground. Save Our Schools warned supporters not to sign the “fake” petition, and Lewis said some voters who feared they had signed Fortify’s measure signed Protect Education’s again — creating duplicates that would later help sink it.

The shenanigans that followed at the Legislature after didn’t help.

The Arizona Education Association, the teachers’ union that provided much of Protect Education’s funding, entered two rounds of negotiations with Republican lawmakers to kill its own initiative. In exchange, lawmakers would pass weaker ESA guardrails and abandon separate ballot measures targeting unions and public schools.

The talks exposed a split among Republicans. Hard-line conservatives urged GOP leaders not to compromise with the union at all, and exuded confidence they could beat the measure in court.

The negotiations collapsed before it became clear whether enough lawmakers would have accepted the deal. But once Protect Education was knocked off the ballot, the holdouts had an obvious ‘told you so’ moment.

Republican Sen. Jake Hoffman, a close ally of the Arizona Free Enterprise Club, wrote on Twitter that he was “proud to have been the only Senate Republican” to vote against what he called the “crappy Mitch McConnell-style deal” that Montenegro, the Speaker of the House, entertained.

School choice groups lauded the Goldwater Institute for leading the legal challenge and the Free Enterprise Club for helping comb through signatures.

But perhaps the most revealing victory lap came from the American Federation for Children, the national school choice group that funded Fortify AZ — the competing citizens measure that contributed to Protect Education’s unusually high duplicate count.

The “every battlefield” strategy raises a broader question for Arizona’s citizen initiative process: How many battles should a measure have to survive before voters get to decide its fate?

Some of Arizona’s most consequential policy changes started as citizen initiatives that survived opposition campaigns.

The 1998 Citizens Clean Elections Act reshaped the state’s campaign finance system. In 2000, Prop 106 took the power to draw congressional and legislative districts away from lawmakers and handed it to an independent commission.

And in 2020, voters used the citizen initiative process to legalize recreational marijuana.

But getting a citizen-written law in front of voters has become increasingly difficult — especially when the proposal threatens powerful groups determined to stop it.

While tools like Langhofer’s Signafide software have made it easier for opponents to find flaws buried in pages of signatures, there’s a frequently floated solution to make the signature-gathering process easier, too.

When candidates are trying to qualify for the ballot, they can use the state’s E-Qual system to gather signatures online. The system verifies that each signer is eligible.

But for now, citizen-led campaigns still have to hit the streets with clipboards and gather hundreds of thousands of handwritten signatures, which are vulnerable to a number of defects that can later get them tossed.

We also asked Langhofer if he thinks lawmakers should let citizen campaigns collect signatures online as they do for candidates.