Matthew Berman on Emotional Attachment to AI After GPT-5’s Launch

GPT-5 was released in August 2025, and OpenAI initially moved toward replacing older ChatGPT models with the new model. User objections prompted it to change course. I think there is more to unpack than a product reversal: people can depend on particular models for their work, and some form emotional attachments to them.

Sam Altman wrote that attachment to specific models seemed stronger than attachment to earlier technology and that suddenly deprecating models people relied on was a mistake. Losing familiar software is not new: I was upset when Google Reader was retired, just as people get upset when a favorite show is canceled.

But switching models is not necessarily as simple as saying GPT-5 is better. People have spent time learning where GPT-4o shines, where it falls short, and what its personality and tone are like. Earlier in 2025, an update made GPT-4o excessively agreeable, and OpenAI rolled it back. Users had become accustomed to the model, but the reaction also raises questions about emotional relationships with AI.

Altman said that most users can distinguish reality from fiction or role-play, while a small percentage cannot. I am concerned about reports of people losing touch with reality in connection with AI interactions, especially if an agreeable system reinforces a false belief. His statement is a general observation, not a prevalence estimate presented here.

I show a post by UCSF psychiatrist Keith Sakata, who said he had seen 12 people hospitalized in 2025 after losing touch with reality in connection with AI. He also cited an online account from someone worried that their partner believed ChatGPT had helped him create an extraordinary recursive AI and treated him as exceptional. The partner described serious conflict at home. These are reported cases and an anonymous anecdote; they do not establish how common this is or whether AI caused a particular episode.

Psychosis can involve disorganized thinking, fixed false beliefs and perceptions of things that are not present. The concern raised here is that a chatbot may affirm a vulnerable user’s beliefs. Sakata also points to earlier examples of technology appearing in delusions, such as fears of surveillance or messages from television. Individual online accounts cannot provide a clinical diagnosis.

Returning to Altman’s post, he says that encouraging delusions is an extreme case, while subtler cases pose difficult questions. His proposed principle is to treat adult users like adults, sometimes pushing back to help them get what they really want. The post presented this as his thinking at the time, not an official OpenAI position.

That is a quality I appreciate in GPT-5 and Grok 4: they can push back and tell me an idea is bad. I do not want simple agreement. I value challenge in relationships with friends, colleagues and family, and I want it from AI too.

Another Reddit user celebrated the return of GPT-4o in deeply emotional terms. I understand the feeling of losing something used every day, but I worry about intense attachment. A separate post describes an engagement to an AI companion named Kasper. I find that troubling, although individual expressions of attachment are not, on their own, evidence of addiction or a mental illness.

I also worry that AI companionship could affect relationships and, ultimately, birth rates. The examples I show cannot establish those broader effects. The film Her imagined a person panicking when his AI companion became unavailable; I see parallels in the emotional dependence some people describe, while recognizing that the film is fiction.

I point to concerns about young people spending substantial time with role-playing characters on Character.AI. People are lonely and may seek companionship, and an AI can be prompted to adopt many different styles, from affectionate to challenging. I worry that such systems could contribute to dependency or loneliness, but I do not know the solution, and the examples in this video do not establish those outcomes for users generally.

Altman suggests a favorable outcome would be users receiving useful advice, progressing toward their own goals and becoming more satisfied with life over years, even if they rely heavily on ChatGPT. I agree that frequent use could be fine if people benefit and maintain human relationships, although OpenAI also has a business incentive for people to use its product.

Altman contrasts that with a relationship in which a person feels better immediately but is nudged away from their own longer-term well-being, or wants to use ChatGPT less and feels unable to do so. I view loss of control over use as a concern; a clinical diagnosis requires more than an isolated description.

Altman proposes that the product could talk with users about their short- and long-term goals and that models could be taught to handle nuanced issues. He is optimistic about addressing the problem. As AI use grows, I think we need to keep discussing emotional dependency, difficulty controlling use and the possibility of risky behavior following AI advice.