In early 2020, when the coronavirus was still a distant concern, my wife and I booked an AirAsia flight to Bali. Big mistake. At the start of lockdown, we scrambled to secure a refund. We called the airline’s customer-support line: no dice. We pleaded with its online chatbot, a lobotomized character named AVA. We sent a Twitter message to the brand on March 17 and received a response seven weeks later that read, in full, “Twitter Feedback.”

Those were dark days in airline customer service, with so many travelers desperate to figure out alternative plans. The present is not much brighter. In recent months, airlines around the world have changed how they engage with customers who need help. Frontier will no longer take your call, encouraging fliers to make contact via chatbot. Alaska Airlines is removing check-in kiosks at certain airports, driving people to its app. Air France, KLM, and Ryanair have all suspended customer service on Twitter, which for a time may have been the quickest way to summon a living, breathing employee.   

As Twitter melts down and people flee Facebook, social media just isn’t as useful as it once was for airline customer service. At the same time, airlines are leaning into AI, betting that the latest wave of chatbots will be the most cost-effective way to support customers. The long-standing truth is that companies don’t want to talk to you. First they didn’t want to do it in person, then they didn’t want to do it by phone, now they don’t want to do it online, and soon they won’t want to do it at all. It’s not personal—it just costs money. But hype-fueled AI products have yet to pick up the slack. “A chatbot being able to talk and to learn and to suggest and to persuade and do all of these things that humans do? I haven’t seen it in action, personally,” Eva Ascarza, a co-founder of the Customer Intelligence Lab at Harvard Business School, told me. Airline customer service is caught between two eras of the internet: one built on social media, the other on machine learning. The transition promises to be rocky. If you’re traveling this summer, you better hope that you don’t need help from an airline.

Airlines belong to a category of consumer-facing businesses that marketers call “high-touch”; they deal with customers whose needs are constantly evolving. Flights are delayed, bags get lost, people have to change their plans. And customers feel they deserve a certain level of care: After all, despite its gradual democratization, air travel remains quite expensive, especially in this period of high inflation. Multiply passenger expectations by the total number of seats—Delta flies something like the population of Sacramento every day, on average—and you start to appreciate the sector’s complexities.

These daunting customer-service demands have pushed airlines to automate since the dawn of mainframes. In 1960, IBM and American Airlines launched the first computerized reservation tool, based on a program developed for the Air Force. Customers would call a travel agent, who would then call an airline ticketing agent, who would then input the trip particulars. By 1964, the system could process some 7,000 bookings an hour, at a time when ticketing agents working manually could process one or two. The problem is, we’re still using it. “The basic systems which said ‘Box A talks to Box B via telex’ have largely remained unchanged since the 1950s,” Timothy O’Neil-Dunne, an airline-industry consultant, told me. (He paused to make sure I knew what a telex was. It’s a fax for text messages.) “So we are dealing with very, very old tech,” he added.

That old tech speaks in short codes: confirmation numbers, airport initials, seat numbers, passenger types. Customers rarely know all of the data that apply to their itinerary, which meant that until the advent of more advanced AI in recent years, changing a flight or locating a bag required a human intermediary, someone fluent in airline and English who could translate a question and input it as DL754, ATL, 19B, and Y. But call centers are expensive—even in Manila. Mindsay, a company that develops conversational AI for the industry, estimates that each support call costs airlines $2.20; in 2017, Harvard Business Review pegged the average cost of a live customer-service interaction at three times that amount.

Over the past decade, Facebook and Twitter emerged as efficient alternatives, allowing airlines to automate their response to certain posts and messages while paying special attention to the most urgent issues (or in some cases, the highest-profile users). In some ways, airlines demonstrated the viability of extending customer service over social media—if they could do it, any brand could. A study last year by the customer-experience company Emplifi found that among 23 industries, airlines had the second-fastest average customer-response time. In many cases, tweeting at an airline can really result in shorter wait times than sternly repeating “representative” on the phone or running a gantlet of scripted if-then scenarios with an online textbox.

Until recently, that kind of automated sorting was the best that chatbots—which many airlines offered early versions of—could muster. The predominant use case for AI in customer service was “the prioritization of calls, the prioritization of requests,” Ascarza said: software that decided how long you could wait for human assistance before the big vein in your head popped. On the other side of Twitter sat a flesh-and-blood airline agent whose voice was never heard but keenly felt. You could tweet something salty, tag the airline, and soon get an invitation to DM from an agent, who invariably signs their name.

But customer service through social media has become strained. “Can you calm down and allow me some time to work please ??” Delta tweeted to an inquiring customer last year. During the pandemic, airlines struggled to handle the unprecedented volume of passengers upset by endless rescheduling, and they doubled down on their automation efforts. In 2020, Delta temporarily suspended its customer service on Twitter and Facebook amid agent shortages and increased wait times. KLM, which was fielding 50,000 Facebook messages a day that March, enhanced its chatbot with machine learning; the discount carriers WestJet and AirAsia leaned into their existing ones.

Not all bots were created equal: AirAsia CEO Tony Fernandes recently called AVA, my erstwhile nemesis, “the most hated AI chatbot” in Southeast Asia. Nevertheless, it worked in the aggregate—at least from the brand’s point of view. “During COVID, AVA helped to clear millions of cases that would not have been possible given the amount of requests,” Fernandes said in an emailed statement. That was all before the implosion of Twitter under Elon Musk: Twitter has begun charging companies $1,000 a month to integrate their customer service into the app, prompting Air France and other companies to reconsider the site altogether. Meanwhile, the rise of platforms like TikTok, which aren’t as conducive to customer engagement, have undermined social-media support even further.

Those trends, along with recent strides in generative AI, have emboldened airline executives. Air India has committed $200 million to update its digital systems, which will include ChatGPT-driven features. Frontier claims that its self-service model requires less labor and delivers better customer-service experiences (although in a recent investor report, the airline baldly states its interest in limiting “avenue[s] for customer negotiation”). In February, AirAsia replaced AVA with a new bot called Ask Bo, which it promises will be “more proactive and attentive” thanks to “enhanced” AI. Technical details are scant, although a spokesperson for the airline claims that since Bo’s launch, 95 percent of all customer queries have flowed through it, and 73 percent of those queries were resolved with no follow-up. Considering how eagerly airlines have leaned into automation, expect other carriers to soon follow suit: “Airlines look on the face of it to be an ideal place to start to deploy an AI-based chatbot” powered by the latest large language models, O’Neil-Dunne said.

These days, even when you appeal to airlines over social media, you’re likely triggering some kind of machine-learning program. Both Twitter and Meta have invested in automation features for brands fielding customer messages. “I just get a link, and then I’m starting this WhatsApp conversation,” Ascarza said. “Sometimes I know it’s a bot, sometimes not.” She pointed out that although research suggests consumers prefer human interaction, it’s often because they lack compelling alternatives. That’s starting to change—and the ancillary features of AI support may suit us just fine. On a text-based interface, “it’s okay to be short,” Ascarza said. “And it’s okay not to be polite. And it’s okay to just get to the point.”

But the freedom to jettison social niceties is a shallow benefit. The simple reality is that customer service has eroded on social media, while the AI programs meant to replace it don’t yet meet the burdens of air travel. The last wave of chatbots, such as SWISS’s Nelly and WestJet’s Juliet, could clear the most straightforward cases with brute force, but they could also be blundering and ineffective. Airlines have iterated on those models and introduced new and improved versions, like Ask Bo, looking to capitalize on the fresh interest in AI as so many other companies are. Still, sophisticated bots on the level of ChatGPT don’t widely exist in air travel, and the way that airlines will actually deploy them—however many months or years from now—is an open question. For now, as the social web recedes from view and AI stumbles into an uncertain growth spurt, consumers everywhere are falling through the cracks.

In the long run, AI might improve customer experiences more than it degrades them. As airlines build smarter and smoother chatbots, they may free up their dwindling labor force to deal with the smaller percentage of more complicated requests. If chatbots are already capable of so much, why couldn’t they help us deal with a canceled flight or a lost bag? But AI could reshape customer service in more insidious ways. O’Neil-Dunne noted that as their customer-support tools become more nuanced, airline offerings are going the other way—giving rise to unbundled amenities and pared-down services, like some basic-economy tickets that don’t let you pick your seat without a surcharge. “If the product is simpler, the servicing is easier,” he said.

On the back end, AI could assess the value of every request, including if and when customers should receive help at all. “The decision of who not to serve is as important as who to serve,” Ascarza said. One logical outcome for an airline with millions of customers might be to simply deny or ignore a percentage of all complaints, which already happens with maddening frequency. The only thing worse than a feckless chatbot is a chatbot telling you, with perfect cogence and clarity, to get lost.

The reporters arrived in news vans and satellite trucks that trundled down King Road and colonized parking spots outside the crime scene. TV producers crowded into the Corner Club, chatting up students for tips and gossip, mispronouncing the town’s name—Mos-cow, they kept calling it, not Moss-coe. Nancy Grace, the cable-news host famously obsessed with morbid crimes, set up a table right outside the victims’ house so she could gesture at the building on air while speculating about the last sound they heard before dying. The story was irresistible: Four University of Idaho students brutally stabbed to death in the middle of the night. The killer still at large. No suspects. Motive unknown.

Then the sleuths came. TikTok detectives, true-crime podcasters—they descended on the town with theories to float and suspects to investigate. They rifled through the victims’ digital lives, hunting for clues that might crack the case. In niche Facebook groups, they shared their findings. Did a history professor plot the murders in a jealous rage? Was the nearby fraternity involved? What about that hoodie-clad guy on a Twitch livestream standing behind two of the victims at a food truck?

Days passed without an arrest, then weeks. Frightened students fled the campus. The local police, overwhelmed with tips, begged the public to stop calling with unvetted information. But people just kept coming. “Dark tourists” arrived to take pictures of the house where the murders happened, and post them for bragging rights in their Reddit forums. Someone turned up outside the police line with ghost-hunting equipment to commune with the victims’ spirits. A TikToker with about 100,000 followers tried to identify the killer with tarot cards.

The distinction between professional reporters and clout-chasing cranks blurred into one unwieldy mass of noise and disruption and fearmongering. Locals turned bitterly on all of it, treating the press like hostile occupiers. They hung signs to mess with TV reporters’ live shots—FUCK YOU NANCY GRACE, read one—and posted notes on their doors begging journalists to go away. One local bar owner publicly fantasized about punching reporters in the face.

As the search for the killer dragged on and rumors spread unchecked, the friendly little college town seemed to harden and crack. People were scared, and suspicious of one another. The press couldn’t be trusted; neither could the police. Locals installed security systems and took out restraining orders. They bought guns.

A suspect was arrested six weeks after the murders, but by then it almost didn’t matter. The sleuthing couldn’t stop now. People were too dug in, too invested in their pet hunches and favorite suspects. Some questioned whether the police had the wrong man; some floated potential accomplices. Conspiracy theories lingered, and so did the unease.

[From the March 2023 issue: Megan Garber on how America’s constant need for entertainment blurred the line between fiction and reality]

Don Anderson, a retired high-school teacher who’d lived in Moscow most of his life, couldn’t believe how different everything felt. In some ways, the frenzy that followed the murders was just as disruptive to the community as the crime itself. Before all this, he said, nobody locked their doors. Now everybody was on edge—including him. One day in February, someone called the police claiming that they planned to go into Moscow High School and start shooting. Police quickly figured out it was a hoax—the call wasn’t even coming from Idaho—but Anderson found himself speculating about the motive behind the threat. Was it a prank by some outsider obsessed with the murders? A sinister warning of more violence to come? Was the town just a permanent magnet for voyeurs and creeps now—synonymous with the worst thing that had ever happened there?

“I’m beginning to wonder,” Anderson told me, “if we’re ever going to be the same.”

When I arrived in Moscow in February, the initial media circus had passed. Bryan Kohberger had been arrested six weeks earlier for the murders of four students—Kaylee Goncalves, Madison Mogen, Xana Kernodle, and Ethan Chapin—and the judge had placed a gag order on everyone involved in the case. The news trucks would return once the trial got under way, but for now things were relatively quiet. (Kohberger chose not to enter a plea last month, in effect pleading not guilty.)

I’d been drawn to the town, like everyone else, by the eerie facts of the murders and the still-eerier profile of the suspect, a former criminology student at nearby Washington State University. The details already in circulation were chilling. A car resembling Kohberger’s white Hyundai Elantra could be seen on surveillance videos driving by the house several times shortly before the attacks. Police linked his DNA to a leather knife sheath left on a bed, and his phone history suggested that he’d been near the house 12 times in the preceding months. Once I got to Moscow, however, I found myself fixating less on the crime than on its aftermath—the wreckage left behind when the media and the sleuths had cleared out.

Located on Idaho’s eastern border, Moscow is known around the state for a certain mountain-hippie vibe. Students joke that the town is permanently “stuck in the ’70s.” It has a lively folk-dance scene and an independent theater that shows classic horror films. Main Street is lined with brown-brick buildings that house quirky small businesses including Ampersand, a purveyor of boutique olive oil, and the Breakfast Club, known for its “world-famous cinnamon roll pancakes.”

But even months after the murders, the town seemed traumatized. No one wanted to talk about the case, on the record or off. When I introduced myself as a reporter, people recoiled. My efforts to talk with the victims’ neighbors were met with exasperation and anger. At one door, I found a sign that read simply, WE HAVE NO STATEMENT. LEAVE US ALONE. Eventually I resorted to writing apologetic notes with my phone number and leaving them on windshields and doorsteps. Nobody called.

At the offices of the University of Idaho campus paper, The Argonaut, I found a masthead’s worth of student journalists glumly disillusioned with journalism. Months of unseemly behavior by a scoop-desperate press corps had dimmed their view of the profession. They’d seen cameramen hide in bushes on campus, and reporters try to sneak into dorms. They’d seen TV correspondents shout hostile questions at teenagers still processing their classmates’ deaths as if the kids were prevaricating politicians. In one notably unsavory episode, a tabloid photographer tracked down one of the roommates who’d survived the attack that night and took paparazzi-like photos at her parents’ house for the Daily Mail.

Abigail Spencer, a reporter for The Argonaut, told me that she was struggling to square the heroic stories she’d learned in journalism classes with the reporters who’d invaded her campus. “We’re taught they’re all Cronkite,” she said. “They’re not.”

Haadiya Tariq, who was the paper’s editor, told me the rude behavior had helped her understand the wider antipathy toward the press. “No wonder people hate you,” she sometimes found herself thinking. She was alarmed by the extent to which professional news outlets appeared to deliberately stoke the online ecosystem of conspiracy theories about the case. The TV-news bookers always seemed so nice and thoughtful when they were asking for interviews. But once the cameras turned on, Tariq told me, the questions were invariably aimed at getting her to theorize about the murders in a way that might get traction in the true-crime forums. Experiencing this had helped her understand why so much of the coverage felt “weird or inaccurate or sensational”: “It is 100 percent trying to feed the audience, which is the internet sleuths,” she told me. “That’s kind of the dirty secret I’m starting to realize.” Perhaps more disturbing than the vulturous reporters or the vortex of TikTok speculation was the way the media and the sleuths seemed to encourage and sustain each other—their priorities converging in a vicious ouroboros.

Meanwhile, some unlucky Moscow residents were still struggling to reassemble their lives after becoming main characters in murder-related conspiracy theories. Rebecca Scofield, a history professor at the University of Idaho, was suing the TikToker who’d accused her of plotting the students’ murders because of a (completely fabricated) love affair with Kaylee Goncalves. (The TikToker denied any wrongdoing, and police have said that Scofield was not a suspect.) Friends of a recently deceased Afghanistan veteran were fending off ghoulish speculation on social media that he was involved in the crime.

Jeremy Reagan, a law student who lived in the victims’ neighborhood, became a target when he gave a handful of TV interviews about the murders. Sleuths studied his body language and parsed his facial expressions.

“It reminds me of Ted Bundy when he would talk about murders,” one observed.

“Very disconcerting,” another said.

Soon, they started mining Reagan’s Facebook profile for clues. A bandage on his right hand was treated as especially incriminating—how did he cut himself? Same with a four-year-old Facebook post that mentioned a rave. “Guys at raves ‘chase women’ and ‘do drugs,’ many things to note,” one sleuth deduced. “The girls partied, he mentioned that. Did he try to party with them? Did he actually party with them? Was he turned down by them?’”

Reagan, hoping to clear his name, volunteered to take a DNA test. The police never named him as a suspect. But the online sleuths kept digging—even contacting his friends for intel—and the menacing messages from strangers kept piling up. Reagan started carrying a gun.

“Just having it on me gives that extra sense of security,” he said in a cable-news interview. “Especially now, where the cybersleuths may or may not come.”

Illustration by Zoë Van Dijk

As with every gruesome crime that attains “true crime” status, the Moscow case has been a career-maker for some people in the media. Three separate book projects are reportedly in the works. NewsNation, the upstart cable channel that launched in 2021, has seen record ratings for its wall-to-wall coverage; its lead reporter on the case, Brian Entin, has amassed half a million Twitter followers and been profiled in Vanity Fair.

John and Lauren Matthias knew right away that the Moscow murders would be a big story for them. The Las Vegas–based couple hosts a popular true-crime podcast called Hidden—he’s a forensic psychologist; she’s a former TV reporter—and they have a strong grasp on which cases will pop. The key here, John told me, is that the case began with an “unsub” (police lingo for an unidentified subject of an investigation). “There was a mystery to be solved,” John said. “Nobody knew who the suspect was, there was a huge amount of uncertainty, and people want to play the role of Sherlock Holmes.”

The grisly murders also exploited some of the most basic human fears. “The idea of a group of people asleep in their home at night being attacked randomly … it’s literally a nightmare,” John said.

At its root, the couple believes, true-crime sleuthing is about the psychological desire to bring order where there is none, to make sense of a world that seems scary. “The mind wants the world to make sense,” John told me. “We’re constantly looking for patterns even when they don’t exist. There’s a lot of research that shows that we don’t like things to be random or uncertain.”

Lauren acknowledged that she doesn’t adhere to the same journalistic standards she did when she was a reporter. She indulges in conjecture; she tries to meet her audience where it is. “I never portray myself as the podcaster who’s going to solve it, or has the answers,” she told me. “I become just like my listeners: ‘None of us know. Let’s talk about it. Let’s speculate together. Let’s find clues together.’”

There are times when she feels uncomfortable with the more fever-swampy aspects of this ecosystem. The rush to turn random people into suspects and then demonize them, the lack of accountability when a theory is debunked—it can feel a little gross. “I’ve been a network reporter,” Lauren told me. “And here I am in this really bizarre true-crime community trying to find my footing as a professional.”

But the Matthiases also bristle at the lack of respect they get from mainstream news outlets. They note that they were the first to discover a years-old internet forum in which Kohberger had discussed suffering from “visual snow syndrome”—a disorder associated with depression, anxiety, and insomnia. Rather than crediting their scoop, CNN used a TikTok video of Lauren discussing the story to illustrate the irresponsibility of online sleuths. (When The New York Times eventually “broke” the “visual snow” story, CNN featured the paper’s reporting.)

The appetite for coverage of cases like this one, Lauren told me, is not fizzling out anytime soon. She sees their podcast as a force for good: “We either need to embrace this and be respectful, responsible voices in this community, or watch it become a bigger volcano.”

There are dozens of Facebook groups dedicated to unpacking the Moscow murders. The largest has more than 222,000 members. When I joined the group, several weeks after Kohberger’s arrest, I expected the forum to be quiet. The case was in a holding pattern—what would the murder hobbyists even have to talk about? This was, it turned out, deeply naive.

The group was buzzing. There were chat threads for people to speculate about Kohberger’s motive; there were chat threads for people convinced that Kohberger didn’t do it. A large contingent of members was busy building the case that a Mexican drug cartel was involved. (One key piece of evidence: an image from Google Maps that showed shoes hanging from power lines in the victims’ neighborhood, a purported sign of drug activity—though a quick Google search reveals that shoes can also be a memorial for someone who’s died.) Others latched on to a stray comment made by Kaylee Goncalves’s father about how she’d researched child trafficking. Had Kaylee gotten herself crosswise with a powerful child-trafficking ring? How deep did this go?

Groups like this one invariably attract a fair number of weirdos. Kohberger himself was reportedly known to hang out in crime-related forums, identifying himself as a criminology student; some people have even speculated that anonymous Reddit comments posing theories about the Idaho murders came from Kohberger himself.

But there was something else about the Facebook chatter that unnerved me. While the content wasn’t explicitly political, the group’s mode of thinking bore a striking resemblance to the hives of conspiracism and paranoia that have infected American civic life.

Here was a group of like-minded people clustered in a strange corner of the internet, developing a vocabulary, forming a shared worldview, inventing new storylines to help make sense of the world. Villains were conjured from thin air and elaborate backstories attached to them. Galaxy-brain pattern-finding provided the narrative satisfaction that reality could not. How different, in essence, was this universe from the one inhabited by anti-vaxxers or 9/11 truthers or Pizzagate enthusiasts? Are they not animated by the same urge that animates crime obsessives—to impose order on chaos, to gamify unpredictable, actual life? As John Matthias pointed out to me, “When you have an environment of fear and uncertainty, you tend to get this type of rampant speculation that’s divorced from evidence.”

Devotees of the Moscow case would no doubt push back on this notion. They might argue that their hobby is benign—that they’re just killing time on the internet, indulging in a bit of frivolous speculation for fun. But the consequences of this kind of conspiracy thinking are never contained to their virtual communities. They dribble out into real-life communities, where real people are affected.

Two weeks after I left Moscow, the University of Idaho announced that it planned to demolish the gray house at 1122 King Road. The house sits halfway up a hill, surrounded by squat apartment buildings and unassuming homes. Before the murders, it was known as a hub of off-campus social activity. The roommates liked to throw parties, and local police had responded to several noise complaints. On the day I visited, there were still signs of before. A Christmas wreath hung on the door; strings of lights dangled above the back patio.

The question of what to do with the house had been a subject of debate in town. It was still a crime scene at the moment, but some locals wanted to see it restored and preserved in honor of the victims. These students had good times in that house, the argument went—why let their memories be overshadowed by the murders?

But there was a bleak reality to contend with. As long as the house was standing, it seemed, an unnerving stream of sightseers and sleuths would continue to turn up in the neighborhood. There was no going back.