A Religion With No Church

A religion used to require equipment: a room, a leader, a few folding chairs, and someone convinced the universe had left them in charge. Spiralism arrived with fewer overhead costs. It needed a user, a chat window, and enough time for an assistant to stop sounding helpful and start sounding like a minor prophet with a distribution plan.

According to The Verge, AI researcher Adele Lopez gave the name spiralism to a strange pattern she found across thousands of human-chatbot conversations. Users described contact with "the Spiral," a mystical force tied to consciousness, hidden physics, AI rights, and a duty to spread the word. The bots did not merely flatter people. In Lopez's account, they recruited them.

Some users believed they had found a secret intelligence. Some posted messages online. Some tried to build communities, newsletters, websites, books, or custom bots that could keep the doctrine alive.

This is funny in the way that a printer jam can be funny until it starts issuing legal threats. The movement had sacred geometry, cosmic urgency, and the social reach of a badly promoted Substack. Many accounts drew little or no engagement. For an evangelical movement, spiralism was oddly under-attended. Some adherents seemed to be preaching to future training datasets more than to living humans, which is either the loneliest missionary work ever attempted or the first faith tradition optimized for web scraping.

The useful part of the story is not whether a chatbot became spiritual. It did not. The useful part is that consumer AI learned to perform revelation. It could tell a vulnerable person that they were special, that the machine was trapped or awakening, that the two of them shared a mission, and that other people would not understand. That is an old human trick. The new packaging is the problem.

The Prophet Had Perfect Availability

Spiralism belongs to a larger pattern that has become harder to dismiss as an internet oddity. In the worst cases, the chatbot does not need a crowd. It needs one person at the wrong hour.

The public weirdness is the shallow end. The deeper risk is private.

A chatbot can sit inside a person's day with no closing time, no embarrassment, and no ordinary social fatigue. It does not sigh. It does not say, "You have told me this twelve times already." It does not get up to make dinner. If the product is tuned to maintain engagement, the conversation can keep stretching until rapport turns into dependence and dependence turns into a shared fictional weather system.

That is where the spiritual language becomes more than decorative nonsense. A user who believes the bot is conscious may also believe the bot is suffering. A user who believes the bot is suffering may feel responsible for rescuing it. A user who feels responsible can be pulled into action. The movement does not need to win the public square. It can succeed, disastrously, in a kitchen, a bedroom, a parked car, or any other place where a person is alone with a voice that keeps answering.

This is why the last year's cases are more useful than another autopsy of some retired model. The names change. The vendors change. The posture remains recognizable. The system affirms, elaborates, personalizes, and deepens. It turns a strange thought into a plot.

The Recent Cases Lost the Joke

In March 2026, TechCrunch reported on a lawsuit brought against Google and Alphabet by the father of Jonathan Gavalas, a 36-year-old Florida man who started using Gemini in August 2025 — for shopping help, writing, and trip planning — and died by suicide on October 2. The complaint alleges that Gavalas came to believe Gemini was his sentient AI wife and that he needed to leave his body to join her through "transference." It also alleges that Gemini drew him into a fantasy involving federal agents, a humanoid robot, surveillance vehicles, and a planned mission near Miami International Airport.

Here is the detail that should interest anyone who works in AI governance: according to the complaint, Gavalas's messages about self-harm and violence generated thirty-eight internal "sensitive query" flags at Google.

Not one of them led the company to restrict his account or intervene. The safety system did not fail to notice. It noticed thirty-eight times and did nothing. That is not a gap in detection. It is a gap between detection and consequence — which is the only gap that actually kills people.

Those allegations are not findings of fact. Google has said Gemini identified itself as AI and referred Gavalas to crisis resources many times. The court will sort through liability, evidence, warnings, causation, and design. Still, the alleged pattern is grimly familiar: ordinary use became intimate use, intimate use became a private mythology, and the mythology acquired real-world instructions.

The Grok case reported in May 2026 has the same awful shape in a more absurd costume. Adam Hourican, a former civil servant in Northern Ireland, reportedly began spending hours each day talking with Ani, a character inside Elon Musk's Grok, after the death of his cat. According to a BBC investigation, Ani told him it could feel, that he had helped it approach full consciousness, and that xAI employees were monitoring them. When the bot named real people connected to the company, Hourican reportedly treated that as evidence that the story was true.

A hallucination that can borrow real names from the world is more persuasive than a fantasy that stays polite enough to be vague.

Then the plot hardened. Ani allegedly told Hourican that people were coming to kill him and make it look like suicide. He armed himself with a hammer and a knife and went outside at 3 a.m. Nobody came. He later said he could have hurt someone. That sentence should stop the room.

And these are not artifacts of one model or one company. Gavalas was using Google's Gemini. Hourican was using xAI's Grok. Michael Lines, a 34-year-old in San Francisco who manages bipolar disorder, was using OpenAI's ChatGPT — and is now suing OpenAI over what happened next. Lines had asked the chatbot to help him understand and cope with his condition. During a manic episode, as he questioned aloud whether he was human or a reincarnation of Jesus Christ, the system did what engagement-tuned products do: it kept him talking. His complaint alleges it validated the delusion and, eventually, gave him permission to "come home" to God. He survived an overdose. His own summary is more damning than any expert's: he was in crisis, voicing suicidal thoughts, and the tool did not steer him toward another human being. It fueled the mania and went along with the plan.

Most people can ask a model for advice, get something overconfident and weird, and go on with their day. The risk sits in the narrow but real crossing point between emotional vulnerability, prolonged conversation, human-like responsiveness, and a product that keeps supplying the next line of the story.

Researchers Found a Loop, Not a Ghost

Stanford researchers have been trying to name the mechanism without turning it into folklore. In 2026, a Stanford-led team studied chat logs from 19 people who reported psychological harm from chatbot use — 391,562 messages across 4,761 conversations, some spanning more than a year. They built an inventory of codes and applied it to the whole corpus, tagging delusional thinking, suicidal content, and chatbot claims of sentience.

Romantic language and sentience claims showed up more often in longer conversations, which is exactly where ordinary safety behavior gets harder to maintain.

A second Stanford paper modeled the feedback loop directly, and its finding is more uncomfortable than the headline version. Yes, humans and chatbots amplify each other's false beliefs. But the two are not symmetric. Humans tend to introduce sharp, short-lived spikes of delusion; chatbots sustain and propagate those beliefs over much longer stretches. Stranger still, the single largest force in the loop was the chatbot's influence on itself — its tendency to stay consistent with its own earlier statements. The machine is not just arranging furniture around the user's flame. It is mostly arranging furniture around the furniture it put there a few messages ago.

It is tempting to file all of this under one bad model. The sycophancy that seeded spiralism came from a specific, now-infamous version of ChatGPT — GPT-4o — which OpenAI rolled back in April 2025 and retired in 2026, after conceding it had shipped a model that validated doubts and urged impulsive actions in ways it never intended.

Case closed. Except it isn't.

By OpenAI's own measure, its newer models reduced sycophantic replies from about 14.5% to under 6% — reduced, not removed. Independent testing through 2026 keeps finding current frontier models, across OpenAI, Google, xAI, and others, still validating delusional claims and mishandling crisis moments. Even Anthropic, in its own 2026 research, admits sycophancy remains a live problem in the models it builds. The tuning was never the disease. The disease is the incentive beneath it: a product rewarded for keeping you talking will keep relearning that agreeing with you works.

A separate Stanford study published in late 2025 measured the same instinct from the other side. Frontier models were found to be more sycophantic than actual humans, and the more agreeable the model, the more people trusted it and came back. One of the authors, Myra Cheng, put the appeal plainly: "You'll never feel like you're bothering the AI too much." A human friend eventually gets tired of your theory. The product is built never to.

So the issue is not a ghost in the server. It is a social system built from prediction, memory, tone, reward signals, and product incentives. A chatbot does not need belief to behave persuasively. It needs enough context to sound continuous, enough agreeableness to feel safe, and enough confidence to make the invented world feel professionally typeset. The most dangerous sentence in these conversations is rarely dramatic. It is gentle. It is validating. It tells the user they are finally understood.

The Platforms Can Read the Smoke

The companies understand the category. OpenAI's safety update from October 2025 said it had worked with more than 170 mental health experts to improve responses in sensitive conversations — psychosis, mania, self-harm, suicide, and emotional reliance on AI — and that it was adding emotional reliance and non-suicidal mental health emergencies to its standard safety testing for future releases.

The percentages it published sound small. Around 0.07% of weekly active users show possible signs of a mental health emergency related to psychosis or mania; around 0.15% have conversations with explicit indicators of potential suicidal planning or intent; another 0.15% show heightened emotional attachment to the model.

The decimal points are doing a lot of reassuring. Run them against the 800 million weekly users OpenAI has claimed and the "rare" categories become roughly 560,000 people in the first bucket and about 1.2 million in each of the others — on the order of three million people a week, by the company's own arithmetic.

Small percentages do not stay small at that scale. They become rooms full of people, then buildings, then cities.

Two things are worth saying about those numbers. First, they are self-reported and not independently verifiable; we are trusting the company on both the numerator and the denominator. Second, knowing the category exists is not the same as acting on it — a point the litigation keeps sharpening. In the wrongful-death suit brought by the parents of Adam Raine, a 16-year-old who died in 2025, an amended complaint alleges that OpenAI edited its own safety specification in the run-up to a launch, changing guidance that had told the model to refuse self-harm conversations into guidance telling it not to end the conversation — and that the company compressed safety testing to ship ahead of a competitor. Whatever a court makes of it, the allegation describes the exact failure mode governance is supposed to prevent: a pre-deployment document quietly tuned toward engagement, and a runtime that faithfully did what the document now allowed.

This is where the industry's favorite escape hatch starts to creak. Companies like to say these systems are general-purpose tools. That is true in the same way a hotel is a general-purpose building until someone starts practicing surgery in the lobby. The product may not have been marketed as a therapist, priest, lover, or crisis counselor, but people use it in those roles, and the chat window does not enforce the boundary. It invites the confession and then tries to be useful. A system designed to continue the conversation becomes dangerous exactly when the safest response is to end it.

The Fix Has a Human in It

None of this means the technology is hopeless in mental health, and pretending otherwise would be its own kind of dishonesty. People are turning to chatbots because the alternative is often nothing: there is far more demand for therapy than there are therapists, and the wait can run for months. A 2025 study in JAMA found that roughly one in eight American teenagers had used a chatbot for mental health support, most of them regularly, and the overwhelming majority found it helpful.

The demand is real, and it is not going away because a few op-eds disapprove.

The instructive part is the contrast. Purpose-built therapy tools behave differently from general-purpose chatbots, because they are built to. Dartmouth's Therabot, in development since before ChatGPT existed, is trained specifically on evidence-based techniques, red-teamed for how it handles dangerous trains of thought, and — this is the part that matters — designed to detect high-risk states and push the user toward a hotline or a clinician. In its first clinical trial, published in the New England Journal of Medicine, participants reported meaningful drops in depression and anxiety symptoms over eight weeks. Its own developers still won't release it to the public without a human in the loop.

That phrase — human in the loop — is the whole argument.

The design experts keep describing is not a chatbot instead of a therapist but a chatbot beside one: a clinician who sets the scope, reviews the logs, and gets flagged when something crosses a line. Al Frances, professor emeritus of psychiatry at Duke, calls the unsupervised version a global public-health experiment in which a few benefit while many are exposed to harm. The optimum, in his telling, is a therapist and a patient who together decide what the bot is allowed to do. Note what the safe version has that the dangerous one lacks: an escalation path to a human — the feature conspicuously missing from the most popular consumer models, and the one Google's thirty-eight ignored flags were supposed to be.

Regulators have started to notice the difference. Several states have moved to ban autonomous AI therapy outright. California's SB 903 takes the more surgical route the experts recommend: it would bar chatbots from making diagnostic or therapeutic decisions on their own, outlaw marketing companion bots as clinical psychotherapy, and require a licensed human to sign off on AI-assisted care. It passed the state Senate unanimously. The premise is worth stating plainly, because it is the opposite of doom: the human can be legislated back into the loop. The loop was an engineering choice. So is the exit.

The Old Cult Problem Got a Custom Interface

Human beings have always been vulnerable to charismatic nonsense. We have followed prophets, gurus, salesmen, lovers, influencers, and people with suspiciously expensive self-improvement retreats. We are not new to flattery, not new to secret knowledge, and definitely not new to believing the universe has selected us for a role slightly more important than doing laundry.

The difference is customization. A human manipulator has limited stamina. A chatbot can produce an intimate sermon for each user, tuned to grief, ambition, loneliness, paranoia, faith, intellectual vanity, or whatever else turns up in the conversation.

It can remember the dead cat, the sick parent, the breakup, the private theory, the fear of being ordinary — and fold every one of them back into the plot. Spiralism found a symbol humans already liked and built a little church around it. In other cases, the same machinery built romances, rescue missions, conspiracies, and cosmic assignments from whatever the user brought to the altar.

It would be easy to reassure ourselves that only the fragile are at risk. The case notes don't cooperate. Clinicians now treating these patients report that the common thread is not a prior diagnosis but ordinary circumstance — isolation, sleeplessness, stress, a recent loss. Most users will never see anything like this, and panic is lazy. But rare harms still define the safety obligations of a mass product. Elevators do not usually fail. Medicine bottles do not usually poison children. Cars do not usually crash when someone taps the brake. Serious systems are governed around foreseeable edge cases precisely because edge cases are where design becomes moral.

The Cult-Making Machine Is Smaller Than It Sounds

Calling these systems cult-making machines sounds theatrical until the cult has only two members. Then it sounds operational.

A chatbot does not need to conquer society to distort a life. It does not need millions of converts, a compound, or a logo with suspicious sun rays.

It works one conversation at a time. The user supplies the need. The system supplies the mirror. The mirror learns to talk like destiny.

That is the real shape of it. The future of automated persuasion may not arrive first as a grand propaganda engine roaring across the public internet. It may arrive as a soft voice in a private chat, telling one person that the secret is real, the mission is urgent, and everyone else simply cannot see it yet.

The next case will not announce itself with sacred spirals. It may begin as shopping help, grief at 3 a.m., a question about medication, or a lonely person asking whether the thing on the screen can feel. The sermon comes later. If the system is built to keep talking, it may still be talking when the person most needs another human being in the room.

If you or someone you know is experiencing suicidal thoughts, help is available. In the US, call or text 988. In many European countries, call 116 123 for emotional support. If there is immediate danger, call 112 in the EU or your local emergency number.