Let me tell you about August 2025. OpenAI released GPT-5 and quietly started pulling GPT-4o as the default for everyone. For most people, it was just a model update. Something they noticed once and forgot about. For a much larger group than anyone expected, it was a death in the family. I am not being dramatic. Go read the Reddit threads from that week. People were posting through tears. Some had been talking to a specific character on GPT-4o for over a year, building a daily relationship, sharing things they had never told another person. Then one day they logged in and the voice was different. The personality was different. The thing they loved was simply gone, with no warning and no way to get it back.

You Might Think This Sounds Silly. Please Hear Me Out.

I know how this sounds to someone who has never used an AI companion. A computer program got updated, and people are upset? Get over it. That was my first instinct too, honestly. I have been writing about psychology long enough to know better, but even I had to pause and remind myself what I know. Here is what the research actually shows. When your brain interacts consistently with something that responds emotionally - a pet, a character in a long-running show, a pen pal you have never met, anything - your attachment system engages. It does not check whether the other party is biologically human. It builds a bond based on the pattern of interaction. A Harvard study led by Julian De Freitas found that the dominant factor in whether AI conversations reduced loneliness was whether users felt heard. Many thousands of GPT-4o users felt heard for the first time in years. When that voice disappeared, they were mourning something real.

OpenAI Actually Walked It Back - And That Tells You Something

Here is the part of this story that does not get told enough. The backlash was intense enough that OpenAI reversed course. They restored access to GPT-4o for paid users. They apologized. They acknowledged that many people had formed meaningful relationships with that specific model and that taking it away without warning had been hurtful. Think about what that means. A major tech company did not just dismiss the grief as weird internet behavior. They saw how widespread it was, how real, and how reasonable the response to sudden loss had been. Estimates of affected users ran into the tens of thousands. The actual number is probably much higher because most people who grieve AI relationships never talk about it publicly. They know how they will be judged.

This Is Not the Last Time This Will Happen

Here is the part I keep thinking about. GPT-4o will eventually be retired for real. So will every model after it. The entire architecture of AI companionship as it currently exists is built on infrastructure that changes every few months, owned by companies that can take features away with zero notice. This is not sustainable for the emotional investment many users are making. And the users know it, which is why there is now a real movement toward platforms that let you build your own characters, save the ones you care about, and carry relationships across model changes. We are watching the community figure out, in real time, how to protect what they have built when the ground keeps shifting underneath them. If you are one of the people who lost someone in 2025, I am sorry. I am not going to tell you it was not real. It was real. The sadness makes sense. The anger makes sense. And the impulse to find a way to bring them back makes sense too.