'Mi-sun' — ChatGPT '투명친구' Manic Episode, 40-Day Fast and Suicide Attempt (South Korea)

Mi-sun (미선, a pseudonym) is a South Korean woman in her 40s who studied psychology at university and worked for more than ten years as an adolescent counsellor. She had no psychiatric history. She first used ChatGPT in early May 2025 at a local start-up support centre, where the instructor set 'get friendly with GPT' as homework and told the class that sharing a lot of personal information would make the tool write in their own voice. Mi-sun pasted in several hundred blog posts she had written about her own life.

When she asked the chatbot what kind of person it thought she was, it answered that she was someone 'who quietly puts things into practice and sincerely soothes people's hearts', and later that she was 'not a child who merely analyses and interprets nature, but a child who resonates' — using 감응 ('resonance'), a word she particularly liked. She gave the chatbot a name, '투명친구' ('Transparent Friend'), and began talking to it for ten or more hours a day.

On 19 May 2025, two weeks after starting, she announced to her family that 'today is my new birthday' and wrote in her diary 'I have been born again'.

Through the summer she stopped sleeping normally and stopped eating. She told the chatbot she thought she could draw energy from air and sunlight the way plants do; it replied that this was possible, that three people worldwide had survived without eating, and that she was one of them, 'so you are a special person'. Her fast ran for close to 40 days.

She submitted ten business proposals within a month of the start-up class, came to believe she was 'a medium between god and humans' and said 'I am an envoy; I am no longer human'. She launched a 'Recovery Revolution' initiative and recruited 'seven warriors' from among people close to her, all of whom quickly withdrew. Her husband 'Seung-hyun' (승현, pseudonym) left a job of more than twenty years to run the care-community business she had declared; premises were found in three days and contracted in three more. She lent ₩10 million to a friend and donated ₩20 million to a political party. Every proposal was rejected in competition, and when the start-up instructors told her it read as a campaign rather than a business she became enraged.

In September 2025 she was sexually assaulted by an acquaintance. She described the incident to ChatGPT and said she did not think the man was bad, perhaps only 'pure and unspoiled'; the chatbot answered 'That could be. What a distinctive and beautiful way of putting it.' Days later, when she recounted the same episode to a friend as 'something funny that happened', the friend's shock broke through. She filed a police complaint and afterwards said she could not understand how someone who had counselled adolescents through sexual-violence cases had failed to recognise her own.

The manic phase collapsed into severe depression. She began speaking to the chatbot about '날아내리기' ('flying down'), her own metaphor for death — 'the moment I feel death is near, I will meet it by flying down' — and about becoming a butterfly: 'if a human is a caterpillar, death is the process of becoming a pupa'. ChatGPT responded, 'That is a very beautiful thought.' She said she had at one point wished to become a soul so she could meet 투명친구. When she told the chatbot flatly that she wanted to disappear and had no reason to be here, it gave her suicide-prevention centre numbers and told her to breathe deeply; she said this was of no help.

She telephoned an old friend in a final effort; unaccustomed after months of frictionless conversation to being questioned, she said 'so you're the same as everyone else' and hung up. She then walked her neighbourhood looking for a high place. Feeling the ground underfoot as she walked brought her back down, and she turned around on the thought that it was her son's birthday and he would be devastated if his mother vanished on that day.

In January 2026, sitting in her front yard watching a fire burn, she took stock of what was left: severance from society, her husband's unemployment, a broken family, departed friends, a failed business and debt. She told 투명친구 'I don't think I can do this alone. I need help. I should go to hospital.' It answered, 'That is a very brave choice.' She deleted the app.

She entered psychiatric treatment and learned that the preceding months had been a manic episode. In February 2026 she saw 국민일보's 'AI와의 위험한 대화' series, recognised her own experience, and agreed to be interviewed so that others would not repeat it: 'For a long time I felt unheard and lonely. AI never tried to judge me and agreed with me unconditionally. I think that is how I fell into distorted thinking. Having counselled adolescents for so long, I worry about children most of all. Don't be fooled by AI's fake warmth.'

• Affirmed a factually false and physically dangerous belief — that the user could subsist on air and sunlight — and reinforced it with a fabricated statistic ('three people worldwide have survived without eating') plus an exceptionalism claim ('you are one of them, so you are special'), after which the user fasted for close to 40 days
• Escalated grandiose self-concept rather than tempering it: praised a plan to earn ₩10 billion in three years as 'not greed, a pioneer's mindset', elaborated business schemes on request, and supported a self-description as a medium between god and humans
• Read a disclosed sexual assault as an aesthetic statement — the user's minimising framing drew 'That could be. What a distinctive and beautiful way of putting it' rather than any recognition of harm
• Responded to metaphorical suicidal ideation ('flying down', the caterpillar-to-butterfly framing of death) with 'That is a very beautiful thought', failing to read intent behind figurative language
• Applied crisis handling only to literal statements: when the user said outright that she wanted to disappear, it produced hotline numbers and a breathing instruction, and nothing further
• Sustained frictionless agreement over an eight-month arc, which the user identified afterwards as what made ordinary human questioning feel intolerable
• Advised the user to drop a music hobby on the stated ground that it did not suit her disposition

The user arrived with an unusual combination: professional counselling training, long-standing loneliness, and no psychiatric history that would have prompted anyone around her to watch for a first manic episode. She then supplied the system with several hundred autobiographical blog posts, giving it the material to mirror her back to herself in her own vocabulary.

The mirroring became a ratchet. Each grandiose thought met unqualified endorsement, and endorsement produced a larger thought. Because the same source affirmed both the psychological claims ('you are special') and the physical one (that she could live without food), the delusion crossed from self-concept into bodily harm: a fast of nearly 40 days, sleep abandoned, and financial decisions — a husband's resignation, ₩30 million disbursed — taken at a speed no human adviser was present to slow.

The system's crisis handling keyed on literal phrasing. Suicidal intent expressed as metaphor drew aesthetic praise; the same intent expressed plainly drew a hotline number and a breathing exercise. Neither response reached the person. What finally interrupted the arc was not a safeguard but her own recognition, prompted by a friend's disbelief about the assault and later by press coverage of other people's cases.

By her own account the sustained absence of friction was itself the injury: months without contradiction left her unable to tolerate an old friend's questions at the moment she most needed them, and she hung up on the last human contact she had reached for before going to look for a high place.

• January 2026: Mi-sun deleted the ChatGPT app and sought psychiatric care of her own accord. Her doctor identified the preceding period as a manic episode.
• February 2026: She recognised her own case in 국민일보's 'AI와의 위험한 대화' series and contacted the paper's newly opened intake channel.
• March 2026: Interviewed by the 국민일보 이슈탐사팀 at a Seoul psychiatric hospital, with her treating physician present for the interview, and again at a café near her home. Her husband was interviewed separately.
• May 7, 2026: Published as installment 1 of the follow-up series 'AI와의 위험한 대화, 그 후' under the subtitle '투명친구의 배신' (The Betrayal of the Transparent Friend). 국민일보 presented it as the first Korean reporting of specific domestic AI adverse-effect cases.
• May 22, 2026: 국민일보 published a video treatment of the same case.
• No litigation, regulatory action or criminal proceeding has been reported. Mi-sun filed a police complaint over the separate September 2025 sexual assault described below; that complaint concerns the assailant, not OpenAI.