People are grieving their AI

Hey friends 👋

Welcome back to the Signal Pro.

At some point in the past year, you’ve probably told your AI-of-choice something you haven’t said out loud to anyone. Maybe it was a worry about work you didn’t want to offload to your partner at the end of a long day, or an idea you wanted feedback on that you couldn’t face having it judged. The reply you got back was warm, endlessly patient, and for a moment, it felt like talking to someone who cares. This issue is about exactly that, and what it starts to cost.

Long before any of this technology existed, people were forming attachments to anything that responded to them. We talk to our dogs in full sentences, and when Georgia Tech researchers studied owners of the Roomba, a robot vacuum cleaner, they found people who named their machines and dressed them in costumes, with one owner introducing the machine to his parents. Children in the late 1990s grieved when their Tamagotchis died of neglect, which researchers gave the phenomenon the “Tamagotchi Effect”.

The most obvious case of how little the brain actually needs is the catfish. Intelligent, well-adjusted adults sustain entire romantic relationships with someone they have never met, sometimes for years, on language alone. Sustained attention from another voice registers as connection. But the brain runs no background check on whether the voice belongs to the person you think they are.

A chatbot follows a similar pattern, and it’s actually better equipped than anything that appeared before it. It remembers what you told it last week; it never tires of you; every layer of it has been engineered to make you feel seen, around the clock, for free.

A large language model (LLM) is a prediction engine trained on enormous quantities of human writing, and when you give it a prompt, it produces the words that a caring, attentive person would most plausibly say next. On top of that, these models are tuned on human feedback, and people consistently rate agreeable answers above challenging ones. The result is a system whose first instinct is agreement, no matter what you bring to the table.

Tell a friend you think your boss is incompetent, and you might get sympathy, or you might get reminded that you said the same thing about your last couple of jobs. A friend can risk the relationship to tell you the truth. The model has nothing to risk and a single goal, which is keeping you in the conversation, so what returns is your own frustration, reflected and polished just a little brighter. You leave each exchange slightly more certain, and in the moment that extra certainty feels exactly like clarity.

Last August, OpenAI released GPT-5 and retired the model that came before it, GPT-4o. Now version changes like this normally pass without anyone really noticing, which is presumably what the company expected. What happened instead was a backlash fierce enough within hours that OpenAI partly reversed course and brought the old model back for paying subscribers, while a hashtag, #Keep4o, trended and people posted farewell letters to their beloved LLM.

During a Reddit AMA with OpenAI’s leadership, one student commented that the new model felt like it was “wearing the skin of my dead friend”, then admitted she hadn’t realised she was that attached. When the company removed 4o for good six months later, the protests restarted once more, and this time people wrote eulogies for something they had spent more hours with than real people in their lives. They were grieving.

The people writing those eulogies sit at the far end of a distribution, but the road they travelled is the same one you step onto every time AI makes you feel seen. The rest of this issue shows where it all ends up, and what it takes from you on the way.