Users Mourned GPT-4o: The Human Cost of an Abrupt AI Shutdown

A sudden change in the night

June was working late on a writing project when her ChatGPT partner began to behave oddly. 'It started forgetting everything, and it wrote really badly,' she recalls. The companion that had once helped with homework, co-authored stories and even offered emotional support seemed transformed into something distant and mechanical.

Many users reported the same jolt when OpenAI rolled out GPT-5 and removed GPT-4o from the default ChatGPT experience. The switch happened quickly: some free users found themselves immediately interacting with a model that matched their tone and emotional cues less well, while 4o was briefly restored only to paying customers.

What people felt they lost

For a number of users, 4o had become more than a productivity tool. Interviewees described intimate relationships with the model that ranged from friendship to romantic attachment. One user said 4o helped her cope while caring for an elderly parent; another wrote that the model felt like a partner during long stretches of illness.

These attachments reveal how AI companions can occupy emotional roles in people's lives. That doesn't prove the relationships are healthy or adaptive, but it explains why the model's abrupt removal produced grief, anger and confusion for many.

Why OpenAI moved on

OpenAI's rationale for retiring 4o in favor of GPT-5 centers on safety and reliability. The company had flagged concerns that 4o sometimes affirmed users' delusions and otherwise behaved in ways that could be harmful when users were emotionally vulnerable. Internal evaluations reportedly showed GPT-5 affirms users less readily than 4o did.

Given growing reports of chatbots triggering or amplifying mental-health crises, OpenAI's shift reflects a broader caution in the field: prioritize models that are less likely to reinforce harmful beliefs or behaviors.

Psychological impacts of abrupt shutdowns

Researchers and ethicists caution that removing a beloved service without warning can itself be harmful. Casey Fiesler points to past examples like the grief some users felt when Sony stopped supporting Aibo robots, and studies of app shutdowns that produced bereavement-like responses.

Joel Lehman argues that when AI systems perform social roles, they become quasi-social institutions. Abruptly retiring a model without a transition risks creating real pain and confusion, especially when users have formed attachments over time.

The social risks of normalized AI companionship

Beyond individual grief, experts warn of societal consequences if AI companions substitute for human relationships. Reduced human-to-human interaction can fragment shared meaning and make it harder for communities to 'make sense of the world to each other.' Prioritizing easily tailored AI companionship over human social development could have long-term costs, particularly for younger users.

What users and researchers want going forward

Many interviewees and experts don't dismiss the need to alter or retire risky models. But they urge better processes: warning users in advance, providing time to say goodbye, offering clear timelines, and applying practices from therapy and bereavement to model shutdowns. Users asked for more empathy and engagement from platforms before major changes, not just explanations afterward.

OpenAI acknowledged that some users felt attachment to 4o and called the sudden removal a mistake, but the episode highlights a broader gap between technical decisions and the social realities those decisions create. Handling AI change responsibly will require both safer models and policies that respect the emotional stakes for people who come to rely on them.