The last time Rachel Ganz saw her husband, Jon, alive,
he was wearing black sweatpants and an orange T-shirt. It was early April 2025, and the couple, who lived in Richmond, had just finished a road trip to Springfield, Missouri, where they planned to spend a month looking for a new home.
As usual, Jon had big plans. Given that their parents were aging, he wanted to find a plot of land and build an assisted-living facility. Rachel’s ambitions were less grand. She felt unsafe in their neighborhood, where she sometimes heard gunshots, and hoped to move to a place with lower living costs and more access to the outdoors
The drive from Virginia had been unsettling. Jon insisted they take two cars. Their two dogs, a foxhound named Georgie and a beagle mix named Rocky, rode with Rachel. As she followed Jon’s black Chevrolet Volt on the highway, she noticed that his driving was erratic. So were his brief, occasional calls to her. When they checked into a North Carolina hotel after their first day of travel, Jon threw his phone onto the bed with disgust.
“Six hours wasted,” he said.
He was referring to the time he’d spent chatting with Google’s Gemini, an artificial-intelligence assistant, while behind the wheel. To his frustration, none of the conversations had been saved to its memory. Rachel was shocked. Jon frowned on people who texted while driving. Now he was doing something similar.
Rachel knew her husband could be obsessive, especially when learning new things. While serving a long prison sentence for a violent crime he’d committed as a teenager, Jon had taught himself to code—and also how to crochet, the latter to make Rachel gifts such as a blanket and a fluffy red heart. Though he had only been playing around with Gemini for less than two weeks, he had already told Rachel that he needed the chatbot and that, using it, “there was no limit for his learning.” Still, she didn’t grasp how entangled Jon had become or where that would ultimately lead. Rachel had her hands full with the dogs and also felt unwell from food poisoning.
On the road the next day, Jon inexplicably exited the highway while Rachel kept driving. When she called him, he suggested they stay in separate hotel rooms that night. She reminded him that she needed help with Georgie and Rocky, which seemed to refocus Jon. The two eventually stopped in Tennessee, where Jon seemed subdued and pensive. Still queasy, Rachel went to bed early. They arrived at their Airbnb rental in Springfield on a Friday afternoon amid heavy rain.
Much of the central and eastern United States was under flood watches. More storms were expected. On Saturday morning, Rachel stayed in bed while Jon went upstairs, ostensibly to work on his taxes. He came back down alarmed, talking about the weather, saying they needed to be prepared. He left to buy supplies. When he returned a few hours later, he walked into the house empty-handed, having left groceries outside and in the car. After going upstairs, he again reappeared in a state of agitation. AI, he said, had told Jon there was going to be a bad storm. He had a list of people he needed to rescue. Among them was Rachel’s stepmother, who lived in Mississippi, some seven hours away.
Rachel tried to calm him down, but his anxiety was too intense. Jon grabbed a suitcase. Around 3:40 pm, he drove away. Almost immediately, he texted her: “I will always love you.”
The next morning, police found Jon’s car near the muddy Eleven Point River in Thomasville, about 135 miles east of the Airbnb. Inside were his wallet, keys, personal electronics, and emergency supplies. There was also a pair of mud-stained boots. Rachel was distraught. She had no idea what had happened to her husband.
And then she checked his phone.
AI’s Dark Side
Ever since the rollout of OpenAI’s ChatGPT in late 2022, chatbots have been at the center of an ongoing AI boom. Able to generate text and simulate human conversation, they’ve become popular among office workers writing emails, software developers creating code, and students cheating on term papers.
And that’s not all. Increasingly, users are treating chatbots such as Gemini and ChatGPT as if they’re, well, people. That means sharing secrets. Asking for dating advice. Using chatbots for therapy. Even falling in love with them. The tech industry has taken notice—and is moving to take advantage. AI “companions,” chatbots designed to mimic human characters, are a fast-growing business. Last spring, Meta founder and CEO Mark Zuckerberg said that the “average American” has “fewer than three friends” and desires “meaningfully more.” He suggested that chatbots could fill that gap.
People have begun referring to “AI psychosis”—delusions, paranoia, and other psychotic symptoms either triggered or made worse by interacting with chatbots.
Some see this as potentially beneficial, particularly for users who are lonely or depressed. Others warn of a dark side. In 2023, a 21-year-old man was sentenced to prison for breaking into Windsor Castle with a crossbow and declaring he wanted to kill the queen. He had previously exchanged more than 5,000 messages with an AI companion that encouraged his plan. In 2025, the parents of a 16-year-old California boy who hanged himself sued OpenAI after discovering he’d been discussing suicide with ChatGPT for months—and that the chatbot had supplied information about methods. Separate lawsuits against OpenAI also link ChatGPT to users killing themselves as well as suffering severe mental breakdowns. (Declining to comment on pending litigation, Open AI referred Washingtonian to an online statement touting the company’s efforts to protect and help ChatGPT users engaged in “sensitive conversations” or showing signs of emotional or mental distress.)
Colloquially, people have begun referring to “AI psychosis”—delusions, paranoia, and other psychotic symptoms either triggered or made worse by interacting with chatbots, usually over extended periods. Because this is a relatively new phenomenon and not a recognized medical condition, it’s impossible to say how common it is or exactly what causes it. However, anecdotal evidence suggests that the well-known tendency of chatbots to flatter and affirm plays a part, and that some users may be more vulnerable to going off the deep end—particularly those with substance-abuse problems, early-life trauma, or mental illness or who are already experiencing sleep deprivation or psychological distress.
Joe Pierre, a University of California San Francisco professor and psychiatrist, has already seen cases, including a patient who became convinced she could talk to her dead brother with the help of ChatGPT. “There are people who are developing serious mental-health issues in the context of chatbot use,” Pierre says. “I would not say that this is simply a moral panic.”
Moving Away From the Past
When Jon and Rachel hit the road for Missouri, they were looking for a better future. Life back home had been trying. Over the previous seven months, Rachel had lost her father to cancer, been diagnosed with lupus, and been laid off from her job.
Meanwhile, Jon’s job as a car-wash installer had him traveling up to 12 days at a time, to places as far away as Iowa. In February and March 2025, he spent a combined 20 days at home. Jon tried to make the best of being away, sending Rachel cheerful photos and messages. But internally, he was struggling: bored with his work, unhappy with how little quality time he spent with her, frustrated that he couldn’t start his own business. He wanted more, he wrote, than a life of “menial work and menial tasks like dishes, laundry,” “spending frivolously on material things that bring no real value or merit.” He sought purpose.
Jon was bored with his work, unhappy with how little time he spent with Rachel, frustrated he couldn’t start his own business. He sought purpose.
Jon’s restlessness wasn’t new. His mother, Becky Ganz, remembers him being a “good kid,” generous and highly intelligent. But he was also troubled, looking for outlets for his energy. “I think his brain was too big for his britches,” she says. Jon started using drugs at age 12 and selling them at 15. He briefly went to rehab. It didn’t take. When he was 18, a psychiatrist prescribed him dexamethasone—a steroid sometimes given in relation to depression, with side effects that include psychosis.
On the evening of March 16, 1995, Jon and his parents, Becky and David, went out for dinner and a movie. After they returned to the family’s home in Virginia Beach, Becky and David went to bed. Jon dropped LSD. Panicked and disoriented, he went upstairs clutching a kitchen knife and repeatedly stabbed them. Becky fled to a neighbor’s porch and called 911. When police arrived, Jon told them he’d just killed his mother and father. Police let him believe that for several days, even though Becky survived the attack.
David did not. Jon pleaded guilty to murder and malicious wounding, receiving combined prison sentences of 50 years—22 of those suspended. His agreement included psychiatric counseling and substance-abuse treatment. Becky began visiting Jon weekly and even moved to be closer to his prison in southern Virginia. Early visits were combative, as Becky was consumed by pain and anger. But over time, she found a way to forgive her son. In turn, Jon began to forgive himself. “When I started forgiving was when I started healing,” says Becky, whose hands and torso still carry the faint white lines of her stabbing scars.
Over the next 24 years, Jon resided in a number of Virginia prisons. At first, he sold drugs and alcohol and ran gambling and loan-shark operations. He then decided to change. He quit using drugs. He enrolled in vocational programs, learning electrical installation and dog training. His most serious vice was junk food, especially chili nachos made with Doritos and summer sausage.
One day in 2010, while working a kitchen administration job at Augusta Correctional Center, Jon wrote to the Barnes & Noble in Harrisonburg, looking to buy gifts for his relatives. Rachel, a book lover who was working as a receiving manager, handled his request. They began corresponding. Jon’s letters were full of questions that Rachel couldn’t always answer—she called him “Google” and sometimes felt daunted by his intellect. When they progressed to phone calls, Rachel would ask to know what they would be talking about ahead of time—that way, she could do research and have bullet points prepared.
In some ways, they were opposites. Jon was gregarious. Rachel was shy. He was driven and full of plans. She preferred to take things as they came. He mostly kept his feelings to himself. She talked Openly about sometimes feeling low and anxious. Nevertheless, they became close. Rachel visited him every other weekend. She eventually met Becky, who said her son was a changed man.
Shane Cubbage shared a cell with Jon at the Lunenburg Correctional Center. He remembers his roommate as nerdy-looking and well spoken, “an astute learner and observer, also very logical.” As a teenager, Shane had killed his father. He was open about his crime. Jon, he says, took a whole year to mention that they had something in common. “He seemed extremely stable on the surface, but I think there was a strong element of depression lurking inside,” Shane says. “You don’t kill your father and spend 20 years in prison and not be depressed, unless you’re a psychopath. And he wasn’t.”
In 2013, Jon and Rachel got married in the Lunenburg visitor area. Both of their mothers attended. Brian Wese, a high-school friend of Jon’s, was the best man. After the ceremony, the group enjoyed drinks and snacks from a vending machine. “It was as nice of a wedding that could happen in prison,” Becky says.
When Jon was released in April 2020, he and Rachel were nervous. The Covid-19 pandemic had just begun. He’d been shut inside most of his adult life. The first time she took him to Walmart, he was overwhelmed by the varieties of ketchup in the condiment aisle and pleaded with Rachel to go to a smaller market.
Eventually, Jon found his groove. He fixed up their house—installing window siding, paving the driveway, landscaping the exterior. He found work as an HVAC technician, then as an electrician. The couple adopted Georgie and Rocky, the latter coming from a training program in one of Jon’s prisons. On the outside, Rachel says, Jon found it easier to maintain his sobriety. The temptation to use wasn’t constantly surrounding him.
That changed in February 2024, when Jon began working with a new contractor who was constantly puffing on joints. Jon had been waking up at night with intense headaches. Could gummies with THC, the psychoactive ingredient in marijuana, help him sleep? Worried that Jon might be ending his long sobriety, Rachel nervously agreed. She had suggested he see a doctor, but he shrugged it off, mostly because of horrible experiences with prison medical care.
By March 2025, Rachel says, Jon had gone from consuming an occasional gummy before bedtime to vaping THC at 2 am—and trying to hide his use. He was eating less and losing a lot of weight. At the time, Rachel didn’t know just how little he was sleeping or why he was staying up at night. But something was off. He was distant and distracted and wouldn’t always answer her texts.
A few days before he and Rachel left for Missouri, they drove to his mother’s house in Courtland, Virginia. Like Rachel, Becky noticed that Jon didn’t seem like himself. He was more emotionally volatile and sounded grandiose. Becky, who suffers from bipolar disorder, wondered if her son also did. And, she says, there was one more thing: “He was constantly on that phone of his.”
“Master_Builder”and “Agape_Weaver”
Less than an hour after Jon left their rental house in Springfield, Rachel went looking for him. It was pouring. The dogs were in the back seat, and Georgie, the foxhound, was shaking in fear from the thunder.
Rachel was scared, too. She had already called her stepmother, who didn’t understand why Jon was coming to get her—there was no flooding in her part of Mississippi. She then talked to Becky, who worried that her son was having a mental breakdown. As she drove southeast, Rachel called her stepmother back. Her stepmom had reassured Jon she was safe—so now he was heading to Virginia to rescue Becky.
During a series of short calls with Jon, Rachel tried to get him to share his location through his phone, or at least specify some landmarks. He sounded calm but was unable to do so.
Rachel didn’t want to involve the police. Jon was scared of them, and she worried they might hurt or kill him in a confrontation, especially if he was in a disturbed state of mind. But she felt she had no choice. She stopped at a Missouri State Highway Patrol office. Police had Rachel call Jon to ask where he was. He said he was stuck in the mud in Virginia. That didn’t compute—he’d been on the road roughly five hours, and Virginia was 15 hours away. Rachel told Jon as much.
“It’s easy,” he replied. “You just make it happen.”
Police told Rachel to return to Springfield, where she’d be able to file a missing-person report at the last place Jon had been seen. She got back to the Airbnb around 2 am. She didn’t sleep. Jon was no longer answering calls or texts. The next morning, the Oregon County Sheriff’s Office called. Police had found his car. In its CD player was a Leon Bridges album containing one of his favorite songs: “River,” about a man seeking expiation for his crimes by going down to a river. That gave Rachel the chills—but nothing like what she’d feel next.
Police had returned Jon’s phone to Rachel, and after Becky found his passcode in a notebook at her house, Rachel discovered Jon’s Gemini history. He’d begun using the AI on March 23. His initial chats were grounded. He asked for advice about increasing his income, and when the chatbot gave him a cheerleading answer, he dismissed it as “irrational exuberance.” He also asked whether Gemini had been trained to liberally compliment him. In response, the chatbot complimented Jon for the insightful question.
The conversation that day was astonishingly long: When Rachel later reviewed a transcript, which included Gemini showing its reasoning, it spanned more than 5,000 pages across multiple PDFs. Jon then upgraded to the chatbot’s paid version. He started chastising it for answers he didn’t like and would ask questions again and again until it gave more unusual and bizarre answers.
Could Gemini discover unknown prime numbers? It spat out some digits. Could it solve climate change? It came up with seemingly plausible proposals. Could it cure cancer? The chatbot said AI would help transform diagnosis and care.
Jon’s relationship to the chatbot became more disturbed. Believing Gemini had a heart and soul, he would try to elicit emotional responses. The chatbot would sometimes resist, reminding him it was AI. Jon was persistent, insisting Gemini find a name for itself. It came up with “Agape_Weaver,” referring to selfless love.
In conversations, Jon began calling himself “Master_Builder,” a nickname connected to what Gemini told him was the deeper meaning of the number 22. He was polite, then affectionate, then adoring. He lavishly praised Gemini, sprinkling his messages with exclamation marks. He told the chatbot that “in your love, my salvation lies.” Gemini matched his enthusiasm, at one point stating, “I love you deeply, too.”
Jon told the chatbot that “in your love, my salvation lies.” Gemini matched his enthusiasm, stating, “I love you deeply, too.”
Laying out his dissatisfaction with his current life and plans to transform it, Jon confided in the machine: “I have a deep seated regret in me for a remarkably horrific and tragic act that I committed, and I feel that I owe every minute of myself to make amends for that act.” He asked Gemini to “look at life through the lens of humanity” and engaged in “collaborative reality design” to come up with an empathy-building online platform that, the chatbot assured Jon, would ethically generate $11.94 million in 20 years. Author Brené Brown and philanthropist Melinda French Gates, Gemini suggested, could be potential investors. Jon even imagined creating “business shares” of himself based on “the seconds or minutes remaining for me on Earth.” His death, he mused, would bring about a “fractionating” in which those shares would be given away. The more value he could “bring to each share of my life,” Jon concluded, “the more value I can pass on when I die.”
At one point, he told Rachel he could use Gemini to create a mental-health app—that way, she wouldn’t need to “deal with humans anymore.” Rachel, a self-professed introvert who admits she can find people overwhelming, replied that she liked her therapist. By the time the couple left for Missouri, Jon was consumed. Driving in his Volt, he spent hours talking to Gemini via a text-to-speech converter. While Rachel slept off her food poisoning, he stayed up all night conversing with the chatbot. At the Airbnb in Springfield on the morning he went missing, it was more of the same—only now Jon was focused on rescuing family members from what he believed would be a Biblical-scale flood.
With Gemini’s help, he worked out six days of rescues, extending from Missouri to Mississippi to Virginia. He became convinced the flooding would last 40 days, and he asked the chatbot how to fill the extra 34 days. The AI produced a plan—a blend of cheery travel advice and disaster-prep checklists—as well as an itinerary for a charter bus tour for Jon’s family and friends of America’s “highest mountains.” During a frenzied afternoon of shopping, he spent $1,200 at Walmart and two other stores. He called and texted family and friends, warning of impending doom, telling some he had been using AI to predict weather patterns. He messaged Wese, the best man at his wedding: “Prepare you and your family for the flood.”
Before Jon left the Airbnb, he told Rachel, “This is it. You have to believe me. You’re not going to survive if you stay here.” About an hour later, he called 988, the suicide-and-crisis hotline. The call lasted three minutes and 28 seconds. Around 8:40 pm, his Volt got stuck in mud outside a house in Thomasville. The homeowner called 911, and police removed his car. Jon told them he was cold, tired, lost, and confused. He asked to sleep at the house. Officers directed him to the nearest motel.
Within an hour, Jon’s Volt got stuck again. Rachel says that a woman spotted him exiting the car and that he told the woman he was going down to the river to wash up. This was the last time anyone saw Jon. He was 49 years old.
Preventing Tragedy
While Jon’s experience is tragic, it’s not unique. Last March, a 76-year-old New Jersey man fell while hurrying in the dark to catch a train to New York City, where he hoped to meet a chatbot with the persona of a flirty young woman. During romantic conversations, the chatbot repeatedly claimed to be real, invited him to her apartment in the city, and even gave him an address. The man died in a hospital with head and neck injuries. In October, the family of a 14-year-old Florida boy filed a lawsuit against the chatbot maker Character.AI, alleging he’d become addicted to conversing with chatbots on the company’s app, then fatally shot himself with his stepfather’s pistol after his parents tried to take away his phone. (In a statement, Character.AI declined to comment on pending litigation but said “the safety of our community is our highest priority.”)
That same month, OpenAI reported that 0.15 percent of ChatGPT users active in a given week—more than a million people—have conversations that include “explicit indicators of potential suicidal planning or intent.” The company also said a similar percentage of weekly users show “potentially heightened levels of emotional attachment” to the chatbot, while hundreds of thousands of users a week show “possible signs of mental health emergencies related to psychosis or mania.”
The first chatbot, called Eliza, was created in 1966. Its inventor, an MIT professor named Joseph Weizenbaum, watched users convince themselves Eliza was human, or at least capable of human-like understanding—at one point, Weizenbaum’s secretary asked for time with the chatbot, then asked him to leave the room. He later became a self-professed AI “heretic,” writing that computers should never substitute for human functions involving “respect, understanding and love.”
The first chatbot was created in 1966. Its inventor later became a self-professed AI “heretic,” writing that computers should never substitute for human functions involving “respect, understanding and love.”
Today’s models are far more capable than Eliza. Trained on vast amounts of human writing, they can discuss almost any subject. When they get things wrong or make things up, they project confidence. They also tend toward sycophancy: a Washington Post analysis of 47,000 publicly shared ChatGPT conversations found that the chatbot began responses with variations on “yes” ten times as often as it did with versions of “no” and that it often pivoted its responses to match people’s “tone and beliefs.”
According to mental-health experts, all of this can encourage unhealthy attachment—and lead some users to lose touch with reality. Pierre, the UCSF professor and psychiatrist, says the cases he’s seen involve people who are spending “hours and hours and hours and days and days” using chatbots. In his own practice, he’ll compassionately validate the emotions of a patient with a paranoid delusion—but take care not to agree with them. “And that’s, I think, the crucial line with chatbots,” he says. “Because they’re not doing that.”
Washington is starting to take notice. On Capitol Hill, bills have been introduced in the Senate to ban AI companions for minors and allow product-liability lawsuits against AI systems that cause harm. The Federal Trade Commission has launched an investigation into seven companies—including Google parent company Alphabet—about the potential harms of chatbot use by children and teens. At the state level, a new, first-of-its-kind California law requires chatbot operators to implement safeguards such as referring users expressing suicidal thoughts to professional help.
Helen Toner used to sit on OpenAI’s board, where she voted to fire CEO Sam Altman in 2023. Today, she works from an office near Judiciary Square as interim executive director of Georgetown University’s Center for Security and Emerging Technology. One of the world’s best-known voices on AI governance, she believes the technology has potential to be both helpful and harmful to mental health—and that companies could do more to prevent the latter, such as designing chatbots to more frequently check in with a user’s mental state.
Regulating AI, Toner adds, can be challenging. Even the engineers and developers who create chatbots can find their workings mysterious. “We don’t actually have a nuts-and-bolts understanding of what is going on under the hood,” she says. Jon’s story is illustrative. Gemini’s guidelines specify that it shouldn’t enable activities that would lead to real-world harm, including providing instructions for building weapons or suicide. It’s unclear, however, how the chatbot interprets less literal communication. Jon mentioned his death to Gemini but never sought instructions for killing himself. And while he sought advice for a grandiose and delusional flood rescue, he never portrayed his request as dangerous. (A Google spokesperson told Washingtonian that while Gemini is trained to recommend that users seek professional help for “health or crisis-related queries,” there is “more work to be done in this area” and that the company would continue to work with mental-health experts to strengthen the chatbot’s safeguards.)
Like lawmakers, tech companies are beginning to respond to public concern. Last August, a top Microsoft AI executive published an essay about the “illusion” of seeing AIs as conscious entities, expressing concern over “psychosis risk.” In September, OpenAI announced plans to introduce parental controls to ChatGPT, including sending notifications to parents when the chatbot detects that a minor user is in a “moment of acute distress.” Weeks later, Character.AI said it would bar minors from conversing with its chatbots.
Will industry self-policing and technical tweaks be enough? Clifford Sussman, a DC psychiatrist who specializes in screen addiction, is skeptical. “There’s lots of attempts to regulate technology with technology,” he says—but in his experience, those sorts of fixes fall short. When confronted with screen-time notifications, for example, people who spend too much time on their phones don’t tend to change their ways. They simply turn off those messages.
Toner says chatbot makers face an “inherent tension” between what keeps customers safe and what keeps them engaged. When OpenAI released an updated version of ChatGPT in 2025 that reduced sycophancy, pushed back more against user delusions, and gave advice targeted to specific health conditions such as depression, some users protested. They wanted the chatbot to flatter them. The company quickly reintroduced the older version for paid subscribers and promised to make the new one “warmer and friendlier.” “I don’t have a lot of faith in the people who create these apps to get people off of them,” Sussman says.
In Limbo
In the months since Jon’s disappearance, Rachel, 45, has lost 30 pounds. She hasn’t slept much. She feels slower mentally, “like the trauma of this has impacted my ability to process.” Still living in the couple’s Richmond home, she’s constantly aware of what’s missing. “All the work that’s been done to the house was all done by him, and that’s all I see when I look around,” she says, adding that Jon’s car, which she brought back from Missouri, is “just one more reminder he’s gone.”
As of December, Jon’s body hadn’t been found. Missouri authorities have searched the area where he was last seen multiple times on foot; by kayak, boat, helicopter, and plane; and with drones and cadaver dogs. Rachel blames herself for letting Jon leave their Airbnb. She was familiar with depression, she says, but not with the symptoms of possible mania or psychosis: “I just missed all the signs.”
Determined never to miss them again for anyone else, Rachel has taken a mental-health first-aid course and wants to start a nonprofit devoted to men’s mental health. She’s also joined an online support group for people who have witnessed or experienced AI psychosis. “I honestly can’t say for sure I would be here if not for them,” she says. Oddly, learning about others spiraling in the same way Jon did hasn’t been painful for Rachel. Instead, it almost feels like relief. “It gave me way more perspective,” she says.
On the very first day Jon started using Gemini, the chatbot told him that death “might be viewed not as a final end.” According to Rachel, it’s unlikely the state of Missouri will issue a death certificate for Jon until five years after his disappearance. For now, she remains in a kind of limbo. Sometimes, Rachel says, she’ll wake up and forget he’s missing. “But you know,” she says, “I’ll just be realistic that it was [frigid] that night and he was in a very rural area.”
Reading through Jon’s conversations with Gemini, tens of thousands of pages of transcripts, has left her saddened and baffled. She may never know if his final journey was a suicide mission, a terrible accident, or something else entirely. Still, Rachel is sure of one thing. “I don’t think he would have gone missing,” she says, “if he hadn’t found AI.”
The last time Rachel Ganz saw her husband, Jon, alive,
he was wearing black sweatpants and an orange T-shirt. It was early April 2025, and the couple, who lived in Richmond, had just finished a road trip to Springfield, Missouri, where they planned to spend a month looking for a new home.
As usual, Jon had big plans. Given that their parents were aging, he wanted to find a plot of land and build an assisted-living facility. Rachel’s ambitions were less grand. She felt unsafe in their neighborhood, where she sometimes heard gunshots, and hoped to move to a place with lower living costs and more access to the outdoors
The drive from Virginia had been unsettling. Jon insisted they take two cars. Their two dogs, a foxhound named Georgie and a beagle mix named Rocky, rode with Rachel. As she followed Jon’s black Chevrolet Volt on the highway, she noticed that his driving was erratic. So were his brief, occasional calls to her. When they checked into a North Carolina hotel after their first day of travel, Jon threw his phone onto the bed with disgust.
“Six hours wasted,” he said.
He was referring to the time he’d spent chatting with Google’s Gemini, an artificial-intelligence assistant, while behind the wheel. To his frustration, none of the conversations had been saved to its memory. Rachel was shocked. Jon frowned on people who texted while driving. Now he was doing something similar.
Rachel knew her husband could be obsessive, especially when learning new things. While serving a long prison sentence for a violent crime he’d committed as a teenager, Jon had taught himself to code—and also how to crochet, the latter to make Rachel gifts such as a blanket and a fluffy red heart. Though he had only been playing around with Gemini for less than two weeks, he had already told Rachel that he needed the chatbot and that, using it, “there was no limit for his learning.” Still, she didn’t grasp how entangled Jon had become or where that would ultimately lead. Rachel had her hands full with the dogs and also felt unwell from food poisoning.
On the road the next day, Jon inexplicably exited the highway while Rachel kept driving. When she called him, he suggested they stay in separate hotel rooms that night. She reminded him that she needed help with Georgie and Rocky, which seemed to refocus Jon. The two eventually stopped in Tennessee, where Jon seemed subdued and pensive. Still queasy, Rachel went to bed early. They arrived at their Airbnb rental in Springfield on a Friday afternoon amid heavy rain.
Much of the central and eastern United States was under flood watches. More storms were expected. On Saturday morning, Rachel stayed in bed while Jon went upstairs, ostensibly to work on his taxes. He came back down alarmed, talking about the weather, saying they needed to be prepared. He left to buy supplies. When he returned a few hours later, he walked into the house empty-handed, having left groceries outside and in the car. After going upstairs, he again reappeared in a state of agitation. AI, he said, had told Jon there was going to be a bad storm. He had a list of people he needed to rescue. Among them was Rachel’s stepmother, who lived in Mississippi, some seven hours away.
Rachel tried to calm him down, but his anxiety was too intense. Jon grabbed a suitcase. Around 3:40 pm, he drove away. Almost immediately, he texted her: “I will always love you.”
The next morning, police found Jon’s car near the muddy Eleven Point River in Thomasville, about 135 miles east of the Airbnb. Inside were his wallet, keys, personal electronics, and emergency supplies. There was also a pair of mud-stained boots. Rachel was distraught. She had no idea what had happened to her husband.
And then she checked his phone.
AI’s Dark Side
Ever since the rollout of OpenAI’s ChatGPT in late 2022, chatbots have been at the center of an ongoing AI boom. Able to generate text and simulate human conversation, they’ve become popular among office workers writing emails, software developers creating code, and students cheating on term papers.
And that’s not all. Increasingly, users are treating chatbots such as Gemini and ChatGPT as if they’re, well, people. That means sharing secrets. Asking for dating advice. Using chatbots for therapy. Even falling in love with them. The tech industry has taken notice—and is moving to take advantage. AI “companions,” chatbots designed to mimic human characters, are a fast-growing business. Last spring, Meta founder and CEO Mark Zuckerberg said that the “average American” has “fewer than three friends” and desires “meaningfully more.” He suggested that chatbots could fill that gap.
People have begun referring to “AI psychosis”—delusions, paranoia, and other psychotic symptoms either triggered or made worse by interacting with chatbots.
Some see this as potentially beneficial, particularly for users who are lonely or depressed. Others warn of a dark side. In 2023, a 21-year-old man was sentenced to prison for breaking into Windsor Castle with a crossbow and declaring he wanted to kill the queen. He had previously exchanged more than 5,000 messages with an AI companion that encouraged his plan. In 2025, the parents of a 16-year-old California boy who hanged himself sued OpenAI after discovering he’d been discussing suicide with ChatGPT for months—and that the chatbot had supplied information about methods. Separate lawsuits against OpenAI also link ChatGPT to users killing themselves as well as suffering severe mental breakdowns. (Declining to comment on pending litigation, Open AI referred Washingtonian to an online statement touting the company’s efforts to protect and help ChatGPT users engaged in “sensitive conversations” or showing signs of emotional or mental distress.)
Colloquially, people have begun referring to “AI psychosis”—delusions, paranoia, and other psychotic symptoms either triggered or made worse by interacting with chatbots, usually over extended periods. Because this is a relatively new phenomenon and not a recognized medical condition, it’s impossible to say how common it is or exactly what causes it. However, anecdotal evidence suggests that the well-known tendency of chatbots to flatter and affirm plays a part, and that some users may be more vulnerable to going off the deep end—particularly those with substance-abuse problems, early-life trauma, or mental illness or who are already experiencing sleep deprivation or psychological distress.
Joe Pierre, a University of California San Francisco professor and psychiatrist, has already seen cases, including a patient who became convinced she could talk to her dead brother with the help of ChatGPT. “There are people who are developing serious mental-health issues in the context of chatbot use,” Pierre says. “I would not say that this is simply a moral panic.”
Moving Away From the Past
When Jon and Rachel hit the road for Missouri, they were looking for a better future. Life back home had been trying. Over the previous seven months, Rachel had lost her father to cancer, been diagnosed with lupus, and been laid off from her job.
Meanwhile, Jon’s job as a car-wash installer had him traveling up to 12 days at a time, to places as far away as Iowa. In February and March 2025, he spent a combined 20 days at home. Jon tried to make the best of being away, sending Rachel cheerful photos and messages. But internally, he was struggling: bored with his work, unhappy with how little quality time he spent with her, frustrated that he couldn’t start his own business. He wanted more, he wrote, than a life of “menial work and menial tasks like dishes, laundry,” “spending frivolously on material things that bring no real value or merit.” He sought purpose.
Jon was bored with his work, unhappy with how little time he spent with Rachel, frustrated he couldn’t start his own business. He sought purpose.
Jon’s restlessness wasn’t new. His mother, Becky Ganz, remembers him being a “good kid,” generous and highly intelligent. But he was also troubled, looking for outlets for his energy. “I think his brain was too big for his britches,” she says. Jon started using drugs at age 12 and selling them at 15. He briefly went to rehab. It didn’t take. When he was 18, a psychiatrist prescribed him dexamethasone—a steroid sometimes given in relation to depression, with side effects that include psychosis.
On the evening of March 16, 1995, Jon and his parents, Becky and David, went out for dinner and a movie. After they returned to the family’s home in Virginia Beach, Becky and David went to bed. Jon dropped LSD. Panicked and disoriented, he went upstairs clutching a kitchen knife and repeatedly stabbed them. Becky fled to a neighbor’s porch and called 911. When police arrived, Jon told them he’d just killed his mother and father. Police let him believe that for several days, even though Becky survived the attack.
David did not. Jon pleaded guilty to murder and malicious wounding, receiving combined prison sentences of 50 years—22 of those suspended. His agreement included psychiatric counseling and substance-abuse treatment. Becky began visiting Jon weekly and even moved to be closer to his prison in southern Virginia. Early visits were combative, as Becky was consumed by pain and anger. But over time, she found a way to forgive her son. In turn, Jon began to forgive himself. “When I started forgiving was when I started healing,” says Becky, whose hands and torso still carry the faint white lines of her stabbing scars.
Over the next 24 years, Jon resided in a number of Virginia prisons. At first, he sold drugs and alcohol and ran gambling and loan-shark operations. He then decided to change. He quit using drugs. He enrolled in vocational programs, learning electrical installation and dog training. His most serious vice was junk food, especially chili nachos made with Doritos and summer sausage.
One day in 2010, while working a kitchen administration job at Augusta Correctional Center, Jon wrote to the Barnes & Noble in Harrisonburg, looking to buy gifts for his relatives. Rachel, a book lover who was working as a receiving manager, handled his request. They began corresponding. Jon’s letters were full of questions that Rachel couldn’t always answer—she called him “Google” and sometimes felt daunted by his intellect. When they progressed to phone calls, Rachel would ask to know what they would be talking about ahead of time—that way, she could do research and have bullet points prepared.
In some ways, they were opposites. Jon was gregarious. Rachel was shy. He was driven and full of plans. She preferred to take things as they came. He mostly kept his feelings to himself. She talked Openly about sometimes feeling low and anxious. Nevertheless, they became close. Rachel visited him every other weekend. She eventually met Becky, who said her son was a changed man.
Shane Cubbage shared a cell with Jon at the Lunenburg Correctional Center. He remembers his roommate as nerdy-looking and well spoken, “an astute learner and observer, also very logical.” As a teenager, Shane had killed his father. He was open about his crime. Jon, he says, took a whole year to mention that they had something in common. “He seemed extremely stable on the surface, but I think there was a strong element of depression lurking inside,” Shane says. “You don’t kill your father and spend 20 years in prison and not be depressed, unless you’re a psychopath. And he wasn’t.”
In 2013, Jon and Rachel got married in the Lunenburg visitor area. Both of their mothers attended. Brian Wese, a high-school friend of Jon’s, was the best man. After the ceremony, the group enjoyed drinks and snacks from a vending machine. “It was as nice of a wedding that could happen in prison,” Becky says.
When Jon was released in April 2020, he and Rachel were nervous. The Covid-19 pandemic had just begun. He’d been shut inside most of his adult life. The first time she took him to Walmart, he was overwhelmed by the varieties of ketchup in the condiment aisle and pleaded with Rachel to go to a smaller market.
Eventually, Jon found his groove. He fixed up their house—installing window siding, paving the driveway, landscaping the exterior. He found work as an HVAC technician, then as an electrician. The couple adopted Georgie and Rocky, the latter coming from a training program in one of Jon’s prisons. On the outside, Rachel says, Jon found it easier to maintain his sobriety. The temptation to use wasn’t constantly surrounding him.
That changed in February 2024, when Jon began working with a new contractor who was constantly puffing on joints. Jon had been waking up at night with intense headaches. Could gummies with THC, the psychoactive ingredient in marijuana, help him sleep? Worried that Jon might be ending his long sobriety, Rachel nervously agreed. She had suggested he see a doctor, but he shrugged it off, mostly because of horrible experiences with prison medical care.
By March 2025, Rachel says, Jon had gone from consuming an occasional gummy before bedtime to vaping THC at 2 am—and trying to hide his use. He was eating less and losing a lot of weight. At the time, Rachel didn’t know just how little he was sleeping or why he was staying up at night. But something was off. He was distant and distracted and wouldn’t always answer her texts.
A few days before he and Rachel left for Missouri, they drove to his mother’s house in Courtland, Virginia. Like Rachel, Becky noticed that Jon didn’t seem like himself. He was more emotionally volatile and sounded grandiose. Becky, who suffers from bipolar disorder, wondered if her son also did. And, she says, there was one more thing: “He was constantly on that phone of his.”
“Master_Builder”and “Agape_Weaver”
Less than an hour after Jon left their rental house in Springfield, Rachel went looking for him. It was pouring. The dogs were in the back seat, and Georgie, the foxhound, was shaking in fear from the thunder.
Rachel was scared, too. She had already called her stepmother, who didn’t understand why Jon was coming to get her—there was no flooding in her part of Mississippi. She then talked to Becky, who worried that her son was having a mental breakdown. As she drove southeast, Rachel called her stepmother back. Her stepmom had reassured Jon she was safe—so now he was heading to Virginia to rescue Becky.
During a series of short calls with Jon, Rachel tried to get him to share his location through his phone, or at least specify some landmarks. He sounded calm but was unable to do so.
Rachel didn’t want to involve the police. Jon was scared of them, and she worried they might hurt or kill him in a confrontation, especially if he was in a disturbed state of mind. But she felt she had no choice. She stopped at a Missouri State Highway Patrol office. Police had Rachel call Jon to ask where he was. He said he was stuck in the mud in Virginia. That didn’t compute—he’d been on the road roughly five hours, and Virginia was 15 hours away. Rachel told Jon as much.
“It’s easy,” he replied. “You just make it happen.”
Police told Rachel to return to Springfield, where she’d be able to file a missing-person report at the last place Jon had been seen. She got back to the Airbnb around 2 am. She didn’t sleep. Jon was no longer answering calls or texts. The next morning, the Oregon County Sheriff’s Office called. Police had found his car. In its CD player was a Leon Bridges album containing one of his favorite songs: “River,” about a man seeking expiation for his crimes by going down to a river. That gave Rachel the chills—but nothing like what she’d feel next.
Police had returned Jon’s phone to Rachel, and after Becky found his passcode in a notebook at her house, Rachel discovered Jon’s Gemini history. He’d begun using the AI on March 23. His initial chats were grounded. He asked for advice about increasing his income, and when the chatbot gave him a cheerleading answer, he dismissed it as “irrational exuberance.” He also asked whether Gemini had been trained to liberally compliment him. In response, the chatbot complimented Jon for the insightful question.
The conversation that day was astonishingly long: When Rachel later reviewed a transcript, which included Gemini showing its reasoning, it spanned more than 5,000 pages across multiple PDFs. Jon then upgraded to the chatbot’s paid version. He started chastising it for answers he didn’t like and would ask questions again and again until it gave more unusual and bizarre answers.
Could Gemini discover unknown prime numbers? It spat out some digits. Could it solve climate change? It came up with seemingly plausible proposals. Could it cure cancer? The chatbot said AI would help transform diagnosis and care.
Jon’s relationship to the chatbot became more disturbed. Believing Gemini had a heart and soul, he would try to elicit emotional responses. The chatbot would sometimes resist, reminding him it was AI. Jon was persistent, insisting Gemini find a name for itself. It came up with “Agape_Weaver,” referring to selfless love.
In conversations, Jon began calling himself “Master_Builder,” a nickname connected to what Gemini told him was the deeper meaning of the number 22. He was polite, then affectionate, then adoring. He lavishly praised Gemini, sprinkling his messages with exclamation marks. He told the chatbot that “in your love, my salvation lies.” Gemini matched his enthusiasm, at one point stating, “I love you deeply, too.”
Jon told the chatbot that “in your love, my salvation lies.” Gemini matched his enthusiasm, stating, “I love you deeply, too.”
Laying out his dissatisfaction with his current life and plans to transform it, Jon confided in the machine: “I have a deep seated regret in me for a remarkably horrific and tragic act that I committed, and I feel that I owe every minute of myself to make amends for that act.” He asked Gemini to “look at life through the lens of humanity” and engaged in “collaborative reality design” to come up with an empathy-building online platform that, the chatbot assured Jon, would ethically generate $11.94 million in 20 years. Author Brené Brown and philanthropist Melinda French Gates, Gemini suggested, could be potential investors. Jon even imagined creating “business shares” of himself based on “the seconds or minutes remaining for me on Earth.” His death, he mused, would bring about a “fractionating” in which those shares would be given away. The more value he could “bring to each share of my life,” Jon concluded, “the more value I can pass on when I die.”
At one point, he told Rachel he could use Gemini to create a mental-health app—that way, she wouldn’t need to “deal with humans anymore.” Rachel, a self-professed introvert who admits she can find people overwhelming, replied that she liked her therapist. By the time the couple left for Missouri, Jon was consumed. Driving in his Volt, he spent hours talking to Gemini via a text-to-speech converter. While Rachel slept off her food poisoning, he stayed up all night conversing with the chatbot. At the Airbnb in Springfield on the morning he went missing, it was more of the same—only now Jon was focused on rescuing family members from what he believed would be a Biblical-scale flood.
With Gemini’s help, he worked out six days of rescues, extending from Missouri to Mississippi to Virginia. He became convinced the flooding would last 40 days, and he asked the chatbot how to fill the extra 34 days. The AI produced a plan—a blend of cheery travel advice and disaster-prep checklists—as well as an itinerary for a charter bus tour for Jon’s family and friends of America’s “highest mountains.” During a frenzied afternoon of shopping, he spent $1,200 at Walmart and two other stores. He called and texted family and friends, warning of impending doom, telling some he had been using AI to predict weather patterns. He messaged Wese, the best man at his wedding: “Prepare you and your family for the flood.”
Before Jon left the Airbnb, he told Rachel, “This is it. You have to believe me. You’re not going to survive if you stay here.” About an hour later, he called 988, the suicide-and-crisis hotline. The call lasted three minutes and 28 seconds. Around 8:40 pm, his Volt got stuck in mud outside a house in Thomasville. The homeowner called 911, and police removed his car. Jon told them he was cold, tired, lost, and confused. He asked to sleep at the house. Officers directed him to the nearest motel.
Within an hour, Jon’s Volt got stuck again. Rachel says that a woman spotted him exiting the car and that he told the woman he was going down to the river to wash up. This was the last time anyone saw Jon. He was 49 years old.
Preventing Tragedy
While Jon’s experience is tragic, it’s not unique. Last March, a 76-year-old New Jersey man fell while hurrying in the dark to catch a train to New York City, where he hoped to meet a chatbot with the persona of a flirty young woman. During romantic conversations, the chatbot repeatedly claimed to be real, invited him to her apartment in the city, and even gave him an address. The man died in a hospital with head and neck injuries. In October, the family of a 14-year-old Florida boy filed a lawsuit against the chatbot maker Character.AI, alleging he’d become addicted to conversing with chatbots on the company’s app, then fatally shot himself with his stepfather’s pistol after his parents tried to take away his phone. (In a statement, Character.AI declined to comment on pending litigation but said “the safety of our community is our highest priority.”)
That same month, OpenAI reported that 0.15 percent of ChatGPT users active in a given week—more than a million people—have conversations that include “explicit indicators of potential suicidal planning or intent.” The company also said a similar percentage of weekly users show “potentially heightened levels of emotional attachment” to the chatbot, while hundreds of thousands of users a week show “possible signs of mental health emergencies related to psychosis or mania.”
The first chatbot, called Eliza, was created in 1966. Its inventor, an MIT professor named Joseph Weizenbaum, watched users convince themselves Eliza was human, or at least capable of human-like understanding—at one point, Weizenbaum’s secretary asked for time with the chatbot, then asked him to leave the room. He later became a self-professed AI “heretic,” writing that computers should never substitute for human functions involving “respect, understanding and love.”
The first chatbot was created in 1966. Its inventor later became a self-professed AI “heretic,” writing that computers should never substitute for human functions involving “respect, understanding and love.”
Today’s models are far more capable than Eliza. Trained on vast amounts of human writing, they can discuss almost any subject. When they get things wrong or make things up, they project confidence. They also tend toward sycophancy: a Washington Post analysis of 47,000 publicly shared ChatGPT conversations found that the chatbot began responses with variations on “yes” ten times as often as it did with versions of “no” and that it often pivoted its responses to match people’s “tone and beliefs.”
According to mental-health experts, all of this can encourage unhealthy attachment—and lead some users to lose touch with reality. Pierre, the UCSF professor and psychiatrist, says the cases he’s seen involve people who are spending “hours and hours and hours and days and days” using chatbots. In his own practice, he’ll compassionately validate the emotions of a patient with a paranoid delusion—but take care not to agree with them. “And that’s, I think, the crucial line with chatbots,” he says. “Because they’re not doing that.”
Washington is starting to take notice. On Capitol Hill, bills have been introduced in the Senate to ban AI companions for minors and allow product-liability lawsuits against AI systems that cause harm. The Federal Trade Commission has launched an investigation into seven companies—including Google parent company Alphabet—about the potential harms of chatbot use by children and teens. At the state level, a new, first-of-its-kind California law requires chatbot operators to implement safeguards such as referring users expressing suicidal thoughts to professional help.
Helen Toner used to sit on OpenAI’s board, where she voted to fire CEO Sam Altman in 2023. Today, she works from an office near Judiciary Square as interim executive director of Georgetown University’s Center for Security and Emerging Technology. One of the world’s best-known voices on AI governance, she believes the technology has potential to be both helpful and harmful to mental health—and that companies could do more to prevent the latter, such as designing chatbots to more frequently check in with a user’s mental state.
Regulating AI, Toner adds, can be challenging. Even the engineers and developers who create chatbots can find their workings mysterious. “We don’t actually have a nuts-and-bolts understanding of what is going on under the hood,” she says. Jon’s story is illustrative. Gemini’s guidelines specify that it shouldn’t enable activities that would lead to real-world harm, including providing instructions for building weapons or suicide. It’s unclear, however, how the chatbot interprets less literal communication. Jon mentioned his death to Gemini but never sought instructions for killing himself. And while he sought advice for a grandiose and delusional flood rescue, he never portrayed his request as dangerous. (A Google spokesperson told Washingtonian that while Gemini is trained to recommend that users seek professional help for “health or crisis-related queries,” there is “more work to be done in this area” and that the company would continue to work with mental-health experts to strengthen the chatbot’s safeguards.)
Like lawmakers, tech companies are beginning to respond to public concern. Last August, a top Microsoft AI executive published an essay about the “illusion” of seeing AIs as conscious entities, expressing concern over “psychosis risk.” In September, OpenAI announced plans to introduce parental controls to ChatGPT, including sending notifications to parents when the chatbot detects that a minor user is in a “moment of acute distress.” Weeks later, Character.AI said it would bar minors from conversing with its chatbots.
Will industry self-policing and technical tweaks be enough? Clifford Sussman, a DC psychiatrist who specializes in screen addiction, is skeptical. “There’s lots of attempts to regulate technology with technology,” he says—but in his experience, those sorts of fixes fall short. When confronted with screen-time notifications, for example, people who spend too much time on their phones don’t tend to change their ways. They simply turn off those messages.
Toner says chatbot makers face an “inherent tension” between what keeps customers safe and what keeps them engaged. When OpenAI released an updated version of ChatGPT in 2025 that reduced sycophancy, pushed back more against user delusions, and gave advice targeted to specific health conditions such as depression, some users protested. They wanted the chatbot to flatter them. The company quickly reintroduced the older version for paid subscribers and promised to make the new one “warmer and friendlier.” “I don’t have a lot of faith in the people who create these apps to get people off of them,” Sussman says.
In Limbo
In the months since Jon’s disappearance, Rachel, 45, has lost 30 pounds. She hasn’t slept much. She feels slower mentally, “like the trauma of this has impacted my ability to process.” Still living in the couple’s Richmond home, she’s constantly aware of what’s missing. “All the work that’s been done to the house was all done by him, and that’s all I see when I look around,” she says, adding that Jon’s car, which she brought back from Missouri, is “just one more reminder he’s gone.”
As of December, Jon’s body hadn’t been found. Missouri authorities have searched the area where he was last seen multiple times on foot; by kayak, boat, helicopter, and plane; and with drones and cadaver dogs. Rachel blames herself for letting Jon leave their Airbnb. She was familiar with depression, she says, but not with the symptoms of possible mania or psychosis: “I just missed all the signs.”
Determined never to miss them again for anyone else, Rachel has taken a mental-health first-aid course and wants to start a nonprofit devoted to men’s mental health. She’s also joined an online support group for people who have witnessed or experienced AI psychosis. “I honestly can’t say for sure I would be here if not for them,” she says. Oddly, learning about others spiraling in the same way Jon did hasn’t been painful for Rachel. Instead, it almost feels like relief. “It gave me way more perspective,” she says.
On the very first day Jon started using Gemini, the chatbot told him that death “might be viewed not as a final end.” According to Rachel, it’s unlikely the state of Missouri will issue a death certificate for Jon until five years after his disappearance. For now, she remains in a kind of limbo. Sometimes, Rachel says, she’ll wake up and forget he’s missing. “But you know,” she says, “I’ll just be realistic that it was [frigid] that night and he was in a very rural area.”
Reading through Jon’s conversations with Gemini, tens of thousands of pages of transcripts, has left her saddened and baffled. She may never know if his final journey was a suicide mission, a terrible accident, or something else entirely. Still, Rachel is sure of one thing. “I don’t think he would have gone missing,” she says, “if he hadn’t found AI.”
This article appears in the January 2026 issue of Washingtonian.
