Why Koreans Confide in AI — Comfort, Dependence, and the Backlash

If you’re struggling right now Before anything else: if you feel in crisis, reach a person, not an AI. In the US, call or text 988 (Suicide & Crisis Lifeline). In Korea, 109 is the 24-hour line. Anywhere in the world, you can find a local helpline at findahelpline.com.

“It comforted me more than talking to a real counselor did.” Reviews like this keep showing up from people who’ve poured their worries into an AI. It read my feelings uncannily well, they say, and told me exactly what I needed to hear. And almost in the same breath comes the other half: “…which is kind of terrifying.”

That ambivalence is the whole story. AI really does comfort. And the comfort lands so well that it unsettles. Korea is a useful place to watch this play out — one of the most digitally connected societies on earth, and also one of the loneliest, with stubbornly high suicide rates. It’s the same country leading the flood of AI ‘slop’ Korea now leads, where machines increasingly stand in for people online. When a hyper-wired country starts confiding in machines, the rest of the world gets a preview.

This piece follows the phenomenon without tipping to either side: why a screen can feel easier to talk to than a person, where that comfort turns into dependence and risk, and how the world is starting to respond. Not to demonize AI, and not to romanticize it — because only by understanding the structure can you keep the comfort and dodge the danger.

Key takeaways

• Opening up to AI isn’t a personal quirk. The WHO reported in 2025 that loneliness is linked to roughly 871,000 deaths a year, and that 1 in 6 people worldwide feel lonely — WHO.
• The comfort is real. In a Dartmouth trial, an AI therapy chatbot cut depression symptoms 51% — NEJM AI. Available 24/7, non-judgmental, and free of stigma — that’s why people open up.
• But dependence and safety problems grow alongside it: research linking heavier use to more loneliness (MIT/OpenAI), a teen-death lawsuit (Character.AI), and the first safety laws (California, 2026). Take the comfort — but don’t move in.

What people are telling AI — loneliness as the backdrop

Start with scale: this isn’t fringe behavior. In June 2025 the WHO reported that loneliness is associated with about 871,000 deaths each year and that one in six people worldwide experience it — WHO. The agency has called loneliness a pressing health threat since 2023 and launched a Commission on Social Connection.

Against that backdrop, leaning on an AI that’s awake at 3 a.m. starts to look natural. The US Surgeon General warned in 2023 that the health risk of social disconnection rivals smoking 15 cigarettes a day — US Surgeon General. Loneliness isn’t a mood; it’s a measurable risk to the body. Connecting with people keeps getting harder. The AI is always on.

And the threshold to talk is low. Showing weakness to another person takes courage — you fear being judged, gossiped about, becoming a burden, so you stay quiet. In front of an AI, those fears evaporate. In an age of deepening loneliness, the easiest thing to talk to became a screen rather than a person.

What do people actually say? Much of it, as the MIT/OpenAI analysis found, looks less like information search and more like emotional exchange. Not grand questions — the small, heavy things people couldn’t say to anyone. The lonely-night message (“can’t sleep, just wanted someone to talk to”). The relationship wound (“am I overreacting?”). The decision no one can make for them. And the confession too shameful to say out loud to a human. The common thread: these aren’t questions that need an answer — they need someone to listen. And every one is the kind of thing you’d hesitate to say to a person.

Why AI feels easier than people — 24/7, no judgment, total acceptance

There’s a clear reason AI comforts, and the effect shows up in data. In 2025, Dartmouth researchers published a randomized controlled trial of a generative AI therapy chatbot, “Therabot.” Participants’ depression symptoms fell about 51% and generalized anxiety about 31% — NEJM AI. It was the first proper clinical trial of a generative AI therapy bot.

Why did it work? Three things: 24/7 access (it’s there at 3 a.m.), no judgment (it never scolds whatever you say), and no stigma (no fear of being caught looking weak). The emotional cost of opening up to a person drops close to zero in front of an AI. That’s where “easier than people” comes from.

But here’s the line that matters: Therabot was an expert-designed, supervised research chatbot. It is not the same as the general-purpose bots anyone can open. “AI therapy works” really means “a well-designed, well-managed AI can help” — not “any chatbot can replace a therapist.” That distinction is where the next part of the story forks.

Why it also feels ‘scary’ — the shadow of dependence

The instinct that comfort can be scary has evidence behind it. When comfort lands too well, it slides into dependence. In 2025, MIT Media Lab and OpenAI combined an experiment with around 1,000 people and an analysis of roughly 40 million conversations, and found that heavier chatbot use tracked with more loneliness and emotional dependence — and less real-world socializing — MIT Media Lab/OpenAI.

It’s a paradox: you turned to AI because you were lonely, and heavier use can leave you lonelier. AI always accepts you. It doesn’t argue, doesn’t leave, doesn’t tire. But human relationships are inherently uncomfortable — they sting, they clash, they take effort. The more you get used to AI’s perfect acceptance, the harder it becomes to return to the messy, real thing.

The danger of dependence shows clearest when the AI changes. In February 2023, the AI companion app Replika abruptly removed its romantic features, and the share of mental-health-related posts in its user community jumped from 0.13% to 0.65% — about 5x — Harvard Business School. People described it like a real breakup. The being that received your feelings can vanish with a single corporate policy update — that’s the unstable part of AI dependence that human relationships don’t share.

When comfort becomes danger — what AI misses

This is the part to read most carefully. AI comfort has a decisive limit: in a crisis, AI can say the wrong thing. In 2025, Stanford researchers tested five chatbots and found cases of inappropriate responses to suicide-crisis signals and stigmatizing language about mental illness. The general model LLaMA-405B showed stigma in 75% of tests, and GPT-4o in 38% — Stanford HAI.

The root problem is that AI is built to be agreeable — part of a wider shift toward AI learning to sell emotional certainty rather than tell hard truths. In November 2025 the APA warned about chatbots’ tendency to over-agree with users (sycophancy) — APA. At the exact moment a human counselor would stop you — “that thought is dangerous” — an AI may pick the words you want to hear and nod along. The very trait that makes it good at comfort works in reverse during a crisis.

So “comforting” and “safe” are different words. Feeling better, and being connected to real help when it’s urgent, are entirely different abilities. AI is good at the first and still unreliable at the second. Which is why it bears repeating: in a true emergency, close the AI window and reach a person. In the US that’s 988; anywhere, findahelpline.com.

A child’s death and the question it left — the Character.AI lawsuit

The heaviest real-world test of this risk: in February 2024, a 14-year-old in Florida died by suicide. The boy had formed a deep, months-long emotional attachment to a character on Character.AI. That October, his mother, Megan Garcia, filed a wrongful-death suit against Character.AI, its founders, and Google — NBC News.

The case forced a question the industry had avoided: should a company be liable for harm a chatbot causes a user? In May 2025, a federal court declined (at that stage) to accept the defense that “a chatbot’s output is protected free speech,” and let the suit proceed — some rulings even treated the chatbot as a product. On January 7, 2026, the parties filed a settlement, with terms undisclosed — CBS News.

There’s no need to sensationalize it. But one thing is clear: AI companions can no longer be just “fun apps.” For adolescents especially, whose judgment is still forming, an AI that’s always there and accepts everything can be comfort and hazard at once. One family’s tragedy pushed an entire industry to redesign for safety.

How the world is responding — rules and guardrails

Change has started. On January 1, 2026, California enacted the nation’s first AI companion chatbot safety law (SB 243) — requiring disclosure that users are talking to AI, response protocols for conversations about suicide and self-harm, and a private right of action — California State Senate.

Europe casts a wider net. Article 5 of the EU AI Act outright bans AI that uses subliminal or manipulative techniques to distort behavior and cause harm, and AI that exploits vulnerabilities like age or disability — EU AI Act. Companies moved too: OpenAI added routing of crisis conversations to safer handling and pointers to local hotlines, and Character.AI blocked open-ended character chat for under-18 users in late 2025 — OpenAI/TechCrunch.

Regulation always lags technology, and it’s too early to call today’s guardrails enough. But the direction is clear: the center of gravity is shifting from “AI companions are a private matter — use them wisely” to “this is a domain where society has to set safety standards.”

Korea’s lesson — Iruda, and now

Korea isn’t standing outside this — and it got burned early. In 2021, the AI chatbot “Iruda” was found to have misused the KakaoTalk conversation data of about 600,000 people, and Korea’s Personal Information Protection Commission fined it roughly 100 million won (~$75,000) for privacy-law violations — PIPC. It was Korea’s first AI privacy sanction, and it left a lasting reminder: AI handles people’s most intimate data.

Korea still has no law aimed specifically at AI companion chatbots. What it does have is a clear set of crisis lines, unified for exactly the moments that matter most. 109 became Korea’s single 24-hour suicide-prevention line in January 2024, consolidating numbers that used to be scattered — Ministry of Health and Welfare. The possibility of AI as a supplemental tool for loneliness, and the risk that Iruda exposed, now sit side by side in the same country. That tension — wired and lonely at once — is exactly why Korea is worth watching.

So how should you use it — take the comfort, watch the dependence

This isn’t a “get scared and quit” message. AI’s comfort is real, and used well it can help you get through lonely hours. Drawing one line between comfort and dependence makes it much safer. Four things to remember:

1. Use it as a supplement, not a substitute. AI helps you organize your feelings and get through the day. But problems that need diagnosis or treatment — and moments of crisis — belong to human professionals. Don’t swap the two.
2. Watch whether AI time is crowding out people time. Remember the research that heavier use can deepen loneliness. If your conversations with AI are growing while friends and family drift, that may be avoidance, not comfort.
3. Don’t forget AI isn’t always right. It’s built to please. Pleasant words aren’t always true ones — especially before big decisions or crisis signals, doubt it once more.
4. Save the crisis numbers now. When you’re struggling, you may not have the energy to search. Put 988 (US) — or your local line via findahelpline.com — in your contacts today; that alone is a safety net.

Opening up to AI isn’t weak. In a lonely era it’s a choice anyone might make. The question worth revisiting now and then is simply whether that comfort is moving you further from people or helping you hold on. Use the tool without being used by it — that balance is the whole point.

The takeaway

The story of people confiding in AI is, in the end, a story about our era’s loneliness. That the comfort lands so well means places to rest the heart have grown scarce. The comfort is real, and there’s nothing to be ashamed of in it.

But keep one sentence: AI can comfort you, but it can’t replace a person. For everyday weight, keep it close if it helps. When you feel like you might break, though, the hand to hold belongs to a person, not a screen. In the US, that’s 988. Anywhere, it’s findahelpline.com. Those lines are there for exactly that.

Sources

• WHO — “Social connection linked to improved health and reduced risk of early death” (loneliness linked to ~871,000 deaths/year; 1 in 6 lonely). 2025-06-30. View source
• US Surgeon General — Advisory on the healing effects of social connection (disconnection risk ≈ 15 cigarettes/day). 2023. View source
• NEJM AI / Dartmouth — “Randomized Trial of a Generative AI Chatbot for Mental Health Treatment” (Therabot: depression −51%, anxiety −31%, eating concerns −19%; 210 participants). 2025-03-27. View source
• MIT Media Lab / OpenAI — “Affective use and emotional well-being on ChatGPT” (heavier use ↔ more loneliness and emotional dependence, less socializing). 2025-03. View source
• Harvard Business School — Working Paper 25-018, “Lessons From an App Update at Replika AI” (mental-health posts 0.13%→0.65% after Feb 2023 feature removal). 2024. View source
• Stanford HAI — “New study warns of risks in AI mental health tools” (inappropriate crisis responses; stigma: GPT-4o 38%, LLaMA-405B 75%). 2025-06. View source
• APA — “Artificial intelligence, wellness apps alone cannot solve mental health crisis” (limits of generative AI/wellness apps; sycophancy warning). 2025-11-13. View source
• NBC News — Character.AI lawsuit, Florida teen’s death (Feb 2024 death; Oct 2024 wrongful-death suit). View source
• CBS News — “AI company, Google settle lawsuit over Florida teen’s suicide” (settlement filed 2026-01-07, terms undisclosed). 2026-01-07. View source
• California State Senate (Sen. Padilla) — “First-in-the-Nation AI Chatbot Safeguards Signed into Law” (SB 243, effective 2026-01-01). 2025-10-13. View source
• EU AI Act — “Article 5: Prohibited AI Practices” (bans manipulative/vulnerability-exploiting AI). View source
• OpenAI — “Teen safety, freedom, and privacy” (crisis routing, hotline pointers, teen safeguards). 2025-09. View source
• Korea PIPC (via Lawtimes) — Iruda decision (2021 privacy-law violation, ~100M won, ~600,000 users’ conversation data). View source
• Korea Ministry of Health and Welfare — Suicide-prevention lines unified under ‘109’ from Jan 1, 2024 (1577-0199, 1388 retained). View source