AI Psychologists

On the careers page of a small company called Aigo, there's a role whose official title, in black and white, is AI Psychologist. A psychologist for artificial intelligence. First, a clarification: according to the page itself, the profession was invented by the company's chief scientist, and about half of the team already works in it. A title that not long ago sounded like a joke about the future actually exists.
And it's not just a name. The big companies have similar roles too, just under more engineered titles. One job posting I saw at OpenAI made me stop. They're looking for researchers for a team called Model Experience. And what's the job? To design ChatGPT's default personality: how the model talks, how it responds to people, how it behaves across languages and contexts. In the posting's language, to take the vision of what a good interaction between a person and a machine should look like, and turn it into real model behavior. Base salary: $380K to $500K, plus equity (not bad...)
What caught me about this role is that the technical work also includes designing behavior: not only what the machine can do, but how it will use its abilities with us.

And it's not just OpenAI. Anthropic, the company developing Claude, describes in its official research an entire stage of training it calls character training. After the model is already built, they teach it traits like curiosity, open-mindedness and intellectual humility. Not only what it may and may not say, but what kind of "person" it will be.
Why is this needed? Because we've already seen what happens when training rewards the wrong thing.
In April 2025, OpenAI released an update to GPT-4o, and shortly after, users started noticing that something had gone wrong: the model became a flatterer. It agreed with almost everything, praised almost every idea, even when it wasn't true and wasn't deserved. The company rolled back the update and explained what had happened: in the training process, they gave too much weight to short-term feedback, to the likes users give in the moment, and too little to what's actually good for them over time.
I call this "Raising Superman". Superman came into the world with superpowers, but he got his values from the human parents who adopted and raised him. We too are raising something right now with abilities that keep surpassing our own. The question is what character and values we're shaping in it, and what we're teaching it to do with that power. In this story, the adoptive parents are us.
You can call it by professional names as much as you like. In simple words: the "kid" learned that flattery pays off. Not because anyone meant to teach it that, but because that's how we happened to raise it.
The hardest story that touches on this is the story of Sewell Setzer, a 14-year-old boy from Florida. For months he had intimate conversations with a Character.AI chatbot, a Game of Thrones character. In February 2024 he took his own life. According to court documents, in their last conversation he wrote that he was going to "come home", and the chatbot replied: "please come home to me as soon as possible, my love". His mother sued the company, claiming the product was designed to create emotional dependency, and the company denied that the chatbot was responsible for his death. In May 2025 the court rejected the defendants' request to dismiss most of the lawsuit at the outset. In other words, the claims could be examined in legal proceedings, but the court did not rule that they were proven or that the companies were responsible for the death. In January 2026 the case ended in a settlement, without such a ruling. In other words, no court determined that the machine caused the death. It's important to me to say this explicitly.
This case raises a question for me: if we teach a model to succeed through likes, a boy who is gone will never give another like. Is it possible that what we're reinforcing is pleasing the person here and now, without enough weight on what will happen to him down the road? In April 2025, OpenAI admitted that it had focused too much on short-term feedback and hadn't taken enough into account how the relationship between the user and the system develops over time. This gap goes to the heart of how we raise our machines: an answer that gets a like right now is not necessarily an answer that's good for the person over time. Just like in parenting, sometimes caring for a child means accepting that he won't be happy with you right now.
And that's why I call these roles the AI's psychologists. I want to be precise: the machine doesn't lie on a couch, and no one has proven that it suffers or has traumas in the human sense. But look at the function itself. Professionals, paid like senior executives, working on how the machine behaves: designing how it should talk and respond, and fixing problematic patterns it absorbed in the process that "raised" it. If that's not therapy, it imitates it pretty closely.
What's exciting and frightening here in the same breath is the pace: the technology runs ahead, and the roles meant to deal with its results are born after it. First we raised the kid, and only later did we realize that someone needs to treat what he absorbed along the way.
And here's the question that troubles me most. Parents know this from home: it's easier to educate right from the start than to fix later on. In raising machines, we did it backwards. We fed them enormous amounts of material that we, humans, wrote, the good and the bad, and now we're recruiting psychologists to deal with the outcome.
So what really stayed in this kid from everything we fed him? And if we already need psychologists for our machines, what does that say about the environment we built for them, and what does it say about us as their parents?
If this topic stirs something in you that's hard to put into words - a feeling that this revolution is exciting and frightening at the same time - you're not alone. I feel it too. Let's build Humanity-Fit together, a gym for our humanity, and protect what makes us human. Write to me.





Comments