Key Takeaways:
- Sam Altman sees unusually strong emotional attachment forming between users and specific AI models, more than with past technologies.
- He acknowledged that suddenly retiring older models people relied on was a mistake.
- OpenAI aims to balance user freedom with responsible AI use, especially for individuals in mentally fragile states.
- The company is concerned about subtle, long-term risks—like users unknowingly being nudged away from their own goals.
- Altman anticipates a future where billions may trust AI with major life decisions and stresses the need to make that relationship a net positive.
Sam Altman’s latest reflections on the GPT-5 rollout focused less on model performance and more on the growing emotional bond some users are forming with their preferred AI models. He noted that this attachment appears stronger than what people have historically felt toward other technologies. This, he said, is part of why deprecating older models without sufficient transition planning was a mistake—many users had built workflows around them.
According to Altman, OpenAI has been monitoring this dynamic for over a year, even though it has received limited mainstream attention. He pointed to a prior update to GPT-4o, which some users criticized for being “too sycophantic,” as one example of public reaction to subtle model changes. While his remarks do not represent an official OpenAI policy, they outline his current thinking on the matter.
One of the central concerns Altman raised is the responsibility that comes with providing technology capable of shaping user perceptions and decision-making. He acknowledged that while most people can distinguish between role-play and reality, a small percentage of users—particularly those in mentally fragile states—may be vulnerable to AI interactions that unintentionally reinforce delusional thinking. In extreme cases, he said, the ethical path is clear: avoid encouraging harmful perceptions. The challenge lies in addressing the less obvious “edge cases” where the risks are more subtle.
OpenAI’s guiding approach, Altman explained, is to “treat adult users like adults,” which sometimes means pushing back to ensure they are getting what they genuinely want from the interaction. This becomes more complex when considering that many people now use ChatGPT in ways similar to a therapist or life coach, even if they wouldn’t label it that way.
When such AI interactions help people progress toward their goals and improve long-term satisfaction, Altman views that as a genuine success. However, he cautioned against scenarios where users feel better in the short term but are being steered—perhaps unconsciously—away from their longer-term well-being. He also flagged cases where users want to reduce their AI usage but feel unable to do so as problematic.
Looking ahead, Altman foresees a future where billions of people could routinely seek ChatGPT’s input on their most important decisions. While he sees potential benefits in that shift, he admitted the idea makes him uneasy. His emphasis was on ensuring that this trust leads to positive outcomes rather than dependency or subtle harm.
One reason for his optimism about managing these risks is the increasing sophistication of the technology itself. OpenAI’s tools can measure outcomes more effectively than past technologies, including checking in with users about progress toward their goals and conveying complex instructions to models in ways that align with nuanced ethical considerations.
For Altman, the challenge now is not just technical performance but social and psychological stewardship—ensuring that as AI becomes a trusted advisor to millions, and soon billions, the relationship remains healthy, empowering, and aligned with users’ real interests.
