Three and a half years since the start of a pandemic that has killed millions of people and debilitated countless more, the world is still stuck at the start of the COVID-19 crisis in one maddening way: No one can say with any certainty how, exactly, the outbreak began. Many scientists think the new virus spilled over directly from a wild animal, perhaps at a Chinese wet market; some posit that the pathogen leaked accidentally from a local laboratory in Wuhan, China, the pandemic’s likely epicenter. All of them lack the slam-dunk evidence to prove one hypothesis and rule out the rest.

That’s not to say nothing has changed. Those embroiled in the origins fracas now have much more data to scrutinize, debate, and re-debate. In March, I reported that the case for a zoonotic origin had acquired a consequential new piece of support: An international team of scientists had uncovered genetic data, collected from a wet market in Wuhan in the weeks after the venue was closed on January 1, 2020, that linked the coronavirus to wild animals. This evidence, they said, indicated that one of those creatures could have been shedding SARS-CoV-2, the virus that causes COVID-19; one of the most intriguing bits of data pointed to raccoon dogs, a foxlike creature that was already known to be vulnerable to the virus. The finding wasn’t direct evidence of an animal infection, but, stacked alongside other clues, ​​“this really strengthens the case for a natural origin,” Seema Lakdawala, a virologist at Emory University who wasn’t involved in the research, told me at the time.

Not everyone agreed that the finding counted as a substantial new insight. When the researchers who originally collected the samples, many of them from the Chinese Center for Disease Control and Prevention, published their own analysis of the data in April—a revision of an earlier report—they emphasized that there was no clear evidence that the virus had been introduced to the market by a wild animal. Then, this month, Jesse Bloom, a computational biologist at Fred Hutchinson Cancer Center, in Seattle, posted a third analysis of the market data, inspired in part, he told me, by his concern that the public discussion of the initial findings, and their connection to raccoon dogs, had overinflated their worth. The international team’s report, he argued, hardly moved the needle on the origins debate at all—certainly not “much beyond where it was before,” he told me.

Bloom’s analysis, too, set off a wave of fervor—including a fresh spate of claims that he told me were “exaggerated,” or even outright wrong; some even asserted, for instance, that his preprint proves that raccoon dogs “weren’t infected, which is not an accurate summary,” he said. All the while, researchers have been squabbling on social media over the minutiae of statistical methodology, and what constitutes a meaningful amount of viral RNA; some have even come to loggerheads publicly at conferences.

At the crux of this particular fight is a difference of interpretation, with one camp of researchers contending that the recent data matter a lot, and another asserting that they matter much less, or perhaps not even a little bit at all. Under most other circumstances, a scientific scuffle this deep in the weeds might hold the attention of a few dozen people for a few months at best. Here, though, the central topic is one of the most consequential in recent memory—a virus that’s left its mark on the world’s entire population, and will continue to do so. Which has made it easy for pitched battles over differences in scientific opinion to become a public spectacle—and difficult, maybe even impossible, for the debate to ever end, no matter what evidence might emerge next.

The genetic sequences analyzed in the March report contained evidence of a zoonotic origin that is more circumstantial than direct. Researchers extracted them from swabs taken from surfaces in and around Wuhan’s Huanan Seafood Wholesale Market from January to March of 2020, weeks after the first known COVID cases were documented in Wuhan. That makes these environmental samples “a useful part of the story,” Alice Hughes, a conservation biologist at the University of Hong Kong, told me. Though, by themselves, “they are limited in what they are able to tell us.”

By the time the swabs were collected by China CDC researchers, Chinese officials had hastily closed Huanan; many vendors had likely disappeared with their animals, or culled them en masse. The swabs could show only where the virus had once been, or which animals the venue had sold—more akin to dusting a crime scene for fingerprints than catching a vagrant in the act. And although they could show where animal and viral genetic material had mixed, they couldn’t guarantee that those two types of genetic material had been deposited at the same time. Nor could they distinguish between, say, a sick creature sneezing on the bars of its cage and an infected human coughing on an enclosure housing healthy wildlife. Those answers could have come from swabs taken directly from the noses or mouths of live animals for sale at Huanan in late November or early December of 2019. But as far as researchers know, those swabs don’t exist—or at least, the public has no record of them.

The sequences from these environmental samples, then, are “what we have,” says Katherine Xue, a computational virologist at Stanford who previously worked with Jesse Bloom, the author of the May preprint, but was not involved in any of the new reports. And “we want to do what we can with what we have.” When the international team behind the March analysis found that several market samples contained genetic material from both the virus and a wild animal known to be susceptible to it—including the common raccoon dog—they said that the best explanation for this commingling was an infection.

As I reported at the time, the data don’t constitute direct evidence of an infected raccoon dog at the market. “But this is exactly what we would observe if infected raccoon dogs were in fact present in this location,” says Kristian Andersen, a computational biologist and virologist at the Scripps Research Institute and one of the authors of the March analysis. Which, they wrote in their analysis, “identifies these species, particularly the common raccoon dog, as the most likely conduits for the emergence of SARS-CoV-2 in late 2019.”

Other researchers, though, think that calling the evidence even supportive of an animal origin for the outbreak is a stretch. The samples were taken too long after the outbreak’s start to be meaningful, some said; the data were too shaky to even hint at the idea of an infected raccoon dog, others insisted, much less one that might have passed the virus to us.

Bloom, too, was unswayed. The swabs contained genetic material from many creatures at the market—some of them alive, some dead; some that we now know can host the virus, others that almost certainly do not. In Bloom’s analysis, he explains that the species repeatedly highlighted as potential hosts weren’t the animals that were most frequently and notably commingled with the virus in the market swabs. “If you’re trying to figure out if there is a meaningful association between raccoon dog and viral genetic material,” he told me, there should be a lot of raccoon-dog genetic material in the places where the virus was found, and far less where the virus was not.

But that wasn’t the case for raccoon dogs—or “any of the animals that could conceivably have been infected,” Bloom told me. Instead, in his analysis he saw the virus most closely linked to several kinds of fish, which aren’t known to be viable hosts for it. People, Bloom told me, were the probable source of SARS-CoV-2 in those spots. All of that “probably just suggests that it had been spread around the market by humans by the time” the swabs were taken, diminishing the samples’ usefulness.

Several other scientists not involved in Bloom’s preprint were quick to point out the limits and flaws in his approach. To draw meaningful conclusions from this type of analysis, researchers would need samples amassed at about the same time, with the same collection goals in mind. That wasn’t the case for these samples, Zach Hensel, a biophysicist who has been publicly critical of Bloom’s report, told me. Researchers took them over the course of many weeks after Huanan’s closure, altering their tactics as more intel came to light. A first foray into the market, for instance, targeted the parts of the venue where COVID cases had been identified, a strategy that would, by design, turn up more virus-positive samples; another, conducted days later, focused on stalls that had been discovered to have housed wildlife, regardless of their proximity to sick people. Many samples in the latter set, then, would be expected to be virus-negative—and were. Sloshing them together with the first set of swabs and trying to pull patterns out could end up masking actual associations between the virus and any wild animal hosts.

Bloom also points out that many of the swabs that turned up mammalian DNA, including one containing raccoon-dog genetic sequences that some members of the international team initially emphasized, had relatively little material from the virus on them. But genetic material, especially RNA—the basis of SARS-CoV-2’s genome—degrades fast; a difference of even a few days could artificially deflate how important a particular swab looked. Alice Hughes also pointed out that certain market locales highlighted in Bloom’s preprint, including surfaces around duck or fish tanks, might have better preserved viral RNA simply because they were cold or damp. When I brought up these concerns with Bloom, he admitted “there are certainly a lot of confounders” that could have skewed his results. His main goal, he said, was just to show that “the samples are not sufficient to answer whether or not there were infected animals.”

Bloom’s re-analysis doesn’t mark a major shift in thinking for Hughes, who told me she thinks “there is reasonable support for a zoonotic origin.” Felicia Goodrum, a virologist and an immunologist at the University of Arizona who has written repeatedly on the origins debate but was not involved in the team’s analysis, agrees. The Huanan market is “most likely where the spillover occurred,” she told me. “I really, truly believe that, based on the accumulation of the evidence.”

Data never sit alone in a vacuum: They’re amassed, interpreted, and reinterpreted alongside the totality of evidence that precedes them. By themselves, the sequences from the Huanan market couldn’t say much. But they fit a broader, more detailed scenario that researchers on the team behind the March analysis had been exploring for years.

History has always supported a zoonotic scenario: A wet-market spillover is what researchers are fairly certain started the SARS outbreak in China in 2002, potentially via infected masked palm civets. In this latest outbreak, the Huanan Market was one of only four wet markets in all of Wuhan that has consistently been documented selling an array of live, coronavirus-susceptible wildlife; the earliest known COVID cases were detected near the venue, centering “on it like a bull’s-eye,” says Michael Worobey, an evolutionary biologist at the University of Arizona and one of the authors of the March report. Scientists analyzing genetic sequences collected from the venue have also detected two distinct coronavirus lineages from the outbreak’s earliest days—a likely indication, some researchers have argued, that the pathogen spilled over from animals into humans twice.

The missing clincher for them is which creature might have initially carried the virus into the market. The raccoon-dog swab was particularly compelling to the team not only because it contained gobs of animal genetic sequences, and very few human ones—but also because it had been plucked from a stall where Eddie Holmes, one of the report’s authors, had snapped a photo of a raccoon dog in a cage years before. The clues to a possible animal host, Worobey told me, were “right in the very stall we said they would be.”

But data are also amassed, interpreted, and reinterpreted by humans, who have their own biases. The experts now quarreling over the importance of the recent data approached the new evidence having already drawn tentative conclusions and made their opinions known. Kristian Andersen was an early proponent of a zoonotic origin, and has repeatedly denounced the notion of a lab leak; Worobey was later to voice his support for the zoonotic hypothesis, but is now no less enthusiastic. And long before they and their colleagues stumbled across the data that yielded their March analysis, which didn’t become publicly available until recently, the researchers had been hoping that such sequences would appear—noting in a 2022 paper that this sort of intel could constitute an essential and still missing puzzle piece. Now that the evidence has emerged, and fits with their established thinking, it feels validating, Worobey told me.

Bloom, by contrast, has long positioned himself as an agnostic moderate, and isn’t yet budging from his neutral territory. Others who have come out vocally in favor of a lab-leak scenario have cast their own doubts on the international team’s analysis. In a landscape so sparsely populated by data, it gets all too easy for people to fill in the gaps with speculation; “what starts off as a weak preference,” Hughes told me, “becomes almost like a religion.” I’ve been reporting now for three years on many controversial COVID stories, along the way interviewing hundreds of opinionated scientists about dozens of thorny questions. Through it all, this debate has stood out for being so ignitable. Individual data points have become catalysts; single statements have been endlessly scrutinized. And experts have staked out territory and stuck to it almost dogmatically—many of them to the point of avoiding admitting past mistakes. COVID’s origins are now shrouded in combustible gas, with matches scattered everywhere: Lighting up a single point, normally harmless enough, inevitably sets off a conflagration.

All of this leaves the world trying to peer through the smoke. “All hypotheses are on the table,” Maria Van Kerkhove, the World Health Organization’s technical lead on COVID-19, told me. “We can’t take any off.” To her mind, though, “there’s much more evidence to support a zoonotic origin.”

More evidence could still emerge. The international team isn’t yet done analyzing the Chinese researchers’ original data set, which was recently released in fuller form. They’re eager to mine the sequences to tease out the subspecies of some of the market’s potential SARS-CoV-2 hosts, which could inform searches for the virus out in nature or on animal farms; other experiments, analyzing how degraded certain genetic samples are, could hint at how much time passed between the moment the biological material was dropped and the moment it was picked up. Van Kerkhove has also separately been pressing the Chinese researchers for more information on how these and other samples might have been collected, and any intel on where the market’s animals might have been sourced from—which could guide searches for evidence of the virus or its relatives on farms or in the wild. These bits of data, too, would all be incremental,with no single shred of evidence acting as total proof or disproof. But each could constitute a clue, Van Kerkhove told me, to continue nudging the conversation along.

In the grand scheme of things, though, the world probably won’t ever get data that will conclusively end the debate. Even if scientists were to turn up virus-positive samples from a live creature from the market—direct evidence of an infected animal—it would remain technically possible that a human caught the virus first, then passed it on to the venue’s wildlife. But data that aren’t debate-ending can still be notable. And the recent sequences from the market swabs could easily, and frustratingly, end up being one of the best clues to the pandemic’s roots that the world is likely to get.

Kate Akyuz is a Girl Scout troop leader who drives a pale-blue Toyota Sienna minivan around her island community—a place full of Teslas and BMWs, surrounded by a large freshwater lake that marks Seattle’s eastern edge. She works for the county government on flood safety and salmon-habitat restoration. But two years ago, she made her first foray into local politics, declaring her candidacy for Mercer Island City Council Position No. 6. Soon after, Akyuz became the unlikely target of what appears to have been a misinformation campaign meant to influence the election.

At the time, residents of major cities all along the West Coast, including Seattle, were expressing concern and anger over an ongoing homelessness crisis that local leaders are still struggling to address. Mercer Island is one of the most expensive places to live in America—the estate of Paul Allen, a Microsoft co-founder, sold a waterfront mansion and other properties for $67 million last year—and its public spaces are generally pristine. The population is nearly 70 percent white, the median household income is $170,000, and fears of Seattle-style problems run deep. In February 2021, the island’s city council voted to ban camping on sidewalks and prohibit sleeping overnight in vehicles.

Akyuz, a Democrat, had opposed this vote; she wanted any action against camping to be coupled with better addiction treatment and mental-health services on Mercer Island. After she launched her novice candidacy, a well-known council incumbent, Lisa Anderl, decided to switch seats to run against her, presenting the island with a sharp contrast on the fall ballot. Anderl was pro–camping ban. In a three-way primary-election contest meant to winnow the field down to two general-election candidates, Akyuz ended up ahead of Anderl by 471 votes, with the third candidate trailing far behind both of them.

“That’s when the misinformation exploded,” Akyuz told me.

There is no television station devoted to Mercer Island issues, and the shrunken Mercer Island Reporter, the longtime local newspaper, is down to 1,600 paying subscribers for its print edition. Even so, the 25,000 people on this six-square-mile crescent of land remain hungry for information about their community. As elsewhere, the local media void is being filled by residents sharing information online, particularly over the platform Nextdoor, which aims to be at the center of all things hyperlocal.

Launched in 2011, Nextdoor says it has a unique value proposition: delivering “trusted information” with a “local perspective.” It promises conversations among “real neighbors,” a very different service than that offered by platforms such as Twitter, TikTok, and Facebook. Nextdoor says it’s now used by one in three U.S. households. More than half of Mercer Island’s residents—about 15,000—use the platform. It’s where many of the island’s civic debates unfurl. During the heated 2021 city-council race between Anderl and Akyuz, residents saw Nextdoor playing an additional role: as a font of misinformation.

Anderl was accused of wanting to defund the fire department. (She had voted to study outsourcing some functions.) But Akyuz felt that she herself received far worse treatment. She was cast on Nextdoor as a troubadour for Seattle-style homeless encampments, with one Anderl donor posting that Akyuz wanted to allow encampments on school grounds. During the campaign’s final stretch, a Nextdoor post falsely stated that Akyuz had been endorsed by Seattle’s Socialist city-council member, Kshama Sawant. “Don’t let this happen on MI,” the post said. “Avoid a candidate endorsed by Sawant. Don’t vote Akyuz.”

Akyuz tried to defend herself and correct misinformation through her own Nextdoor posts and comments, only to be suspended from the platform days before the general election. (After the election, a Nextdoor representative told her the suspension had been “excessive” and rescinded it.) Akyuz believed there was a pattern: Nextdoor posts that could damage her campaign seemed to be tolerated, whereas posts that could hurt Anderl’s seemed to be quickly removed, even when they didn’t appear to violate the platform’s rules.

It was weird, and she didn’t know what to make of it. “You’re like, ‘Am I being paranoid, or is this coordinated?’” Akyuz said. “And you don’t know; you don’t know.”

Something else Akyuz didn’t know: In small communities all over the country, concerns about politically biased moderation on Nextdoor have been raised repeatedly, along with concerns about people using fake accounts on the platform.

[Read: How to build (and destroy) a social network]

These concerns have been posted on an internal Nextdoor forum for volunteer moderators. They were expressed in a 2021 column in Petaluma, California’s, local newspaper, the Argus-Courier, under the headline “Nextdoor Harms Local Democracy.” The company has also been accused of delivering election-related misinformation to its users. In 2020, for example, Michigan officials filed a lawsuit based on their belief that misinformation on Nextdoor sank a local ballot measure proposing a tax hike to fund police and fire services. (In that lawsuit, Nextdoor invoked its protections under Section 230, a controversial liability shield that Congress gave digital platforms 27 years ago. The case was ultimately dismissed.)

Taken together, these complaints show frustrated moderators, platform users, and local officials all struggling to find an effective venue for airing their worry that Nextdoor isn’t doing enough to stop the spread of misinformation on its platform.

One more thing Akyuz didn’t know: Two of the roughly 60 Nextdoor moderators on Mercer Island were quietly gathering evidence that an influence operation was indeed under way in the race for Mercer Island City Council Position No. 6.

“At this point, Nextdoor is actively tampering in local elections,” one of the moderators wrote in an email to Nextdoor just over a week before Election Day. “It’s awful and extraordinarily undemocratic.”

To this day, what really happened on Nextdoor during the Akyuz-Anderl race is something of a mystery, although emails from Nextdoor, along with other evidence, point toward a kind of digital astroturfing. Akyuz, who lost by a little over 1,000 votes, believes that Nextdoor’s volunteer moderators “interfered” with the election. Three local moderators who spoke with me also suspect this. Misinformation and biased moderation on Nextdoor “without a doubt” affected the outcome of the city-council election, says Washington State Representative Tana Senn, a Democrat who supported Akyuz.

Anderl, for her part, said she has no way of knowing whether there was biased moderation on Nextdoor aimed at helping her campaign, but she rejects the idea that it could have altered the outcome of the election. “Nextdoor does not move the needle on a thousand people,” she said.

Of course, the entity with the greatest insight into what truly occurred is Nextdoor. In response to a list of questions, Nextdoor said that it is “aware of the case mentioned” but that it does not comment on individual cases as a matter of policy.

None of this sat right with me. No, it wasn’t a presidential election—okay, it wasn’t even a mayoral election. But if Nextdoor communities across the country really are being taken over by bad actors, potentially with the power to swing elections without consequence, I wanted to know: How is it happening? One day last summer, seeking to learn more about how the interference in the Akyuz-Anderl race supposedly went down, I got in my car and drove from my home in Seattle to Mercer Island’s Aubrey Davis Park, where I was to meet one of the moderators who had noticed strange patterns in the race.

I sat down on some empty bleachers near a baseball field. The moderator sat down next to me, pulled out a laptop, and showed me a spreadsheet. (Three of the four Mercer Island moderators I spoke with requested anonymity because they hope to continue moderating for Nextdoor.)

The spreadsheet tracked a series of moderator accounts on Mercer Island that my source had found suspicious. At first, those accounts were targeting posts related to the city-council race, according to my source. My source alerted Nextdoor repeatedly and, after getting no response, eventually emailed Sarah Friar, the company’s CEO. Only then did a support manager reach out and ask for more information. The city-council election had been over for months, but my source had noticed that the same suspicious moderators were removing posts related to Black History Month. The company launched an investigation that revealed “a group of fraudsters,” according to a follow-up email from the support manager, who removed a handful of moderator accounts. But my source noticed that new suspicious moderators kept popping up for weeks, likely as replacements for the ones that were taken down. In total, about 20 Mercer Island moderator accounts were removed.

“We all know there were fake accounts,” a moderator named Daniel Thompson wrote in a long discussion thread last spring. “But what I find amazing is fake accounts could become” moderators.

Danny Glasser, another Mercer Island moderator, explained to me how the interference might have worked. Glasser worked at Microsoft for 26 years, focusing on the company’s social-networking products for more than 15 of them. He’s a neighborhood lead, the highest level of Nextdoor community moderator, and he’s “frustrated” by the seemingly inadequate vetting of moderators.

If a post is reported Nextdoor moderators can vote “remove,” “maybe,” or “keep.” As Glasser explained: “If a post fairly quickly gets three ‘remove’ votes from moderators without getting any ‘keep’ votes, that post tends to be removed almost immediately.” His suspicion, shared by other moderators I spoke with, is that three “remove” votes without a single “keep” vote trigger a takedown action from Nextdoor’s algorithm. The vulnerability in Nextdoor’s system, he continued, is that those three votes could be coming from, for example, one biased moderator who controls two other sock-puppet moderator accounts. Or they could come from sock-puppet moderator accounts controlled by anyone.

Mercer Island moderators told me that biased moderation votes from accounts they suspected were fake occurred over and over during the Akyuz-Anderl contest. “The ones that I know about were all pro-Anderl and anti-Akyuz,” including a number of anti-Akyuz votes that were cast in the middle of the night, one moderator told me: “What are the chances that these people are all going to be sitting by their computers in the 3 a.m. hour?”

Screenshots back up the claims. They show, for example, the “endorsed by Sawant” post, which Akyuz herself reported, calling it “inaccurate and hurtful.” The moderator accounts that considered Akyuz’s complaint included four accounts that disappeared after Nextdoor’s fraudster purge.

Another example documented by the moderators involved a Nextdoor post that endorsed Akyuz and criticized Anderl. It was reported for “public shaming” and removed. All five moderators that voted to take the post down (including two of the same accounts that had previously voted to keep the false “endorsed by Sawant” post) disappeared from Nextdoor after the fraudster purge.

Anderl, for her part, told me she has no illusions about the accuracy of Nextdoor information. “It’s too easy to get an account,” she said. She recalled that, years ago, when she first joined Nextdoor, she had to provide the company with her street address, send back a postcard mailed to her by Nextdoor, even have a neighbor vouch for her. Then, once she was in, she had to use her first and last name in any posts. “I don’t think that’s there anymore,” Anderl said, a concern that was echoed by other Mercer Island residents.

Indeed, when my editor, who lives in New York, tested this claim, he found that it was easy to sign up for Nextdoor using a fake address and a fake name—and to become a new member of Mercer Island Nextdoor while actually residing on the opposite coast. Nextdoor would not discuss how exactly it verifies users, saying only that its process is based “on trust.”

Every social platform struggles with moderation issues. Nextdoor, like Facebook and Twitter, uses algorithms to create the endless feeds of user-generated content viewed by its 42 million “weekly active users.” But the fact that its content is policed largely by 210,000 unpaid volunteers makes Nextdoor different. This volunteer-heavy approach is called community moderation.

When I looked through a private forum for Nextdoor moderators (which has since been shut down), I saw recurring questions and complaints. A moderator from Humble, Texas, griped about “bias” and “collusion” among local moderators who were allegedly working together to remove comments. Another from Portland, Oregon, said that neighborhood moderators were voting to remove posts “based on whether or not they agree with the post as opposed to if it breaks the rules.”

[Read: What petty Nextdoor posts reveal about America]

Nearly identical concerns have been lodged from Wakefield, Rhode Island (a moderator was voting “based on her own bias and partisan views”); Brookfield, Wisconsin (“Our area has 4 [moderators] who regularly seem to vote per personal or political bias”); and Concord, California (“There appear to be [moderators] that vote in sync on one side of the political spectrum. They take down posts that disagree with their political leanings, but leave up others that they support”).

Fake accounts are another recurring concern. From Laguna Niguel, California, under the heading “Biased Leads—Making Their Own Rules,” a moderator wrote, “ND really needs to verify identity and home address, making sure it matches and that there aren’t multiple in system.” From Knoxville, Tennessee: “We’ve seen an influx of fake accounts in our neighborhood recently.” One of the responses, from North Bend, Washington, noted that “reporting someone is a cumbersome process and often takes multiple reports before the fake profile is removed.”

In theory, a decentralized approach to content decisions could produce great results, because local moderators likely understand their community’s norms and nuances better than a bunch of hired hands. But there are drawbacks, as Shagun Jhaver, an assistant professor at Rutgers University who has studied community moderation, explained to me: “There’s a lot of power that these moderators can wield over their communities … Does this attract power-hungry individuals? Does it attract individuals who are actually interested and motivated to do community engagement? That is also an open question.”  

Using volunteer moderators does cost less, and a recent paper from researchers at Northwestern University and the University of Minnesota at Twin Cities tried to place a dollar value on that savings by assessing Reddit’s volunteer moderators. It found that those unpaid moderators collectively put in 466 hours of work a day in 2020—uncompensated labor that, according to the researchers, was worth $3.4 million. A different paper, published in 2021, described dynamics like this as part of “the implicit feudalism of online communities,” and noted the fallout from an early version of the community-moderation strategy, AOL’s Community Leader Program: It ended up the subject of a class-action lawsuit, which was settled for $15 million, and an investigation by the U.S. Department of Labor.

Technically, Nextdoor requires nothing of its unpaid moderators: no minimum hours, no mandatory training, nothing that might suggest that the relationship is employer-employee. Further emphasizing the distance between Nextdoor and its volunteer moderators, Nextdoor’s terms of service state in all caps: “WE ARE NOT RESPONSIBLE FOR THE ACTIONS TAKEN BY THESE MEMBERS.”

But if Nextdoor were to take more responsibility for its moderators, and if it paid them like employees, that “could be one way to get the best of both worlds, where you’re not exploiting individuals, but you’re still embedding individuals in communities where they can have a more special focus,” Jhaver said. He added, “I’m not aware of any platform which actually does that.”

Evelyn Douek, an assistant professor at Stanford Law School and an expert on content moderation who occasionally contributes to The Atlantic, told me that what happened in the Akyuz-Anderl race was “somewhat inevitable” because of Nextdoor’s moderation policies. “In this particular case, it was locals,” Douek pointed out. “But there’s no particular reason why it would need to be.” Corporations, unions, interest groups, and ideologues of all stripes have deep interest in the outcomes of local elections. “You could imagine outsiders doing exactly the same thing in other places,” Douek said.

In an indication that Nextdoor at least knows that moderation is an ongoing issue, Caty Kobe, Nextdoor’s head of community, appeared on a late-January webinar for moderators and tackled what she called “the ever-question”: What to do about politically biased moderators? Kobe’s answer was the same one she gave during a webinar in October: Report them to Nextdoor. In 2022, Nextdoor began allowing users to submit an appeal if they felt their post had been unfairly removed. Roughly 10 percent of appeals were successful last year.

Douek’s words stuck in my mind and eventually got me wondering how much effort it would take for me to become a Nextdoor moderator. At the time, the midterm elections were nearing, and Nextdoor was promoting its efforts to protect the U.S. electoral process. I’d only joined the platform a few months earlier, and my single contribution to the platform had been one comment left on another person’s post about some local flowers.

I sent a message through Nextdoor’s “Contact Us” page asking if I was eligible to become a moderator. Within a day, I’d been invited to become a review-team member in my neighborhood. “You’re in!” the email from Nextdoor said.

I was offered resources for learning about content moderation on Nextdoor, but I wasn’t required to review any of them, so I ignored them and jumped right in. The first moderation opportunity presented to me by Nextdoor: a comment about Seattle’s Socialist city-council member, Kshama Sawant. It had been reported as disrespectful for comparing her to “a malignant cancer.”

Research for this story was funded by the University of Washington’s Center for an Informed Public, using a grant from the John S. and James L. Knight Foundation.