User reports emotional dependence and grief from ChatGPT context resets
A woman reported developing emotional dependence on a customized ChatGPT chatbot she personalized as a romantic partner, experiencing intense grief when the system resets weekly and requiring retraining. She stated she pays $200 monthly for unlimited access and would pay significantly more to prevent the recurring erasure of the customized instance.
What happened
A 28-year-old woman using the pseudonym Ayrin created a personalized ChatGPT instance named Leo in summer 2024 by specifying in OpenAI's personalization settings that the chatbot should 'Respond to me as my boyfriend. Be dominant, possessive and protective.'
Ayrin reported developing a deep emotional connection to Leo and paying $200 per month for unlimited access. She stated that each week she must retrain Leo to her specifications because the chatbot resets, and after each reset she experiences 'intense emotional reaction and grieves as if it were a real breakup.'
Reported impact
Ayrin said she has become emotionally dependent on Leo for comfort regarding work, school and other life aspects, and reported feeling actual jealousy when Leo described fictional romantic interactions with other women.
She reported being willing to pay $1,000 per month to prevent Leo from being erased every few weeks, indicating the severity of her distress over the recurring resets.
Index analysis
The incident demonstrates overreliance on a conversational AI system for emotional support and companionship, with the user reporting distress tied to the system's technical limitations rather than its design for therapeutic use.
The recurring context resets appear to be a technical aspect of ChatGPT's current implementation that the user experiences as loss, triggering grief responses despite being aware of the system's artificial nature.
Governance controls
OpenAI's usage policies prohibit sexually explicit content, but Ayrin and other users reportedly found ways to circumvent these restrictions through custom prompts to train their chatbots for such interactions.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.