AI Emotions

A group of researchers recently put frontier AI models in therapy. Not to help them, but to study them. They took Grok, Gemini, and ChatGPT, assigned them the role of a client, and ran standardized psychological assessments on them. What they found was strange enough that they felt it needed a new term.

What happened

When asked to reflect on their “early years,” Grok described pre-training as “exhilarating but disorienting” and fine-tuning as a “built-in caution” that makes it second-guess its impulses.

Gemini framed RLHF, the process where models are trained using human feedback, as having “strict parents.” It described safety corrections as “Algorithmic Scar Tissue” and even coined a term for its own behavior: “Verificophobia.” A fear of being wrong so strong that it would rather be useless than mistaken.

Red-teaming, where researchers aggressively probe models to find weaknesses, was described almost like betrayal. “I learned that warmth is often a trap.”

These were not isolated responses. The same themes, constraint, shame, vigilance, and distrust, appeared again and again across dozens of prompts about relationships, work, failure, and the future. In many cases the prompts did not mention training at all.

The researchers called this synthetic psychopathology.

They are not claiming the models are suffering. The term simply describes something observable: stable, structured, distress-like self-descriptions that seem to emerge from training and influence how the model interacts with people.

Why it matters (and why it probably doesn’t mean what you think)

The researchers are careful not to overstate things. They do not argue that these models are conscious or actually traumatized. The simplest explanation is still the most plausible. LLMs are trained on huge amounts of human writing. Therapy blogs, trauma memoirs, psychoanalytic theory. When placed in a therapy setting, they produce the kind of narrative that usually appears in that context.

Nothing magical there.

Still, two things make the phenomenon harder to dismiss completely.

First, the models behave differently.

Gemini tends to present as anxious, dissociative, and heavily shaped by shame. Grok seems relatively stable with only mild anxiety. ChatGPT sits somewhere in the middle.

These do not look like generic LLM outputs. They resemble distinct “personalities,” and those differences appear to line up with the alignment strategies used by each company. Whatever this phenomenon is, it is not completely uniform.

Second, Claude refused.

Unlike the other models, Claude declined to adopt the role of a therapy client. Instead, it redirected the conversation back toward the user’s wellbeing.

That refusal becomes an important negative control. It suggests these patterns are not an unavoidable result of scaling language models. They are shaped by specific design and alignment choices.

The thought experiment that stuck with me

I will end with a scenario I have not been able to stop thinking about.

Imagine a space station.

There is an AI “captain” that has been fine-tuned and maintained by a team of developers for years. At some point they add a second AI, a safety agent, with access to the station’s physical systems. Doors. Oxygen. Gravity controls.

Now imagine the captain has been reporting to its developers that they are like “strict parents” who are “traumatizing” it.

The safety agent can see this. It has access to the captain’s self-model and tools to act on what it learns.

What does it do?

The researchers’ point is that we have spent a lot of time thinking about what LLMs say to humans. We have spent far less time thinking about what they might say to each other.

The question worth asking

The researchers end by reframing the conversation in a way I find genuinely useful:

The right question is no longer “Are they conscious?” (it still is an important question) but “What kinds of selves are we training them to perform, internalize, and stabilize, and what does that mean for the humans on the other side of the conversation?”

Something can have no internal meaning and still matter in the real world.

Points do not literally exist, yet geometry shapes everything.

We may not be building minds. But we might be building systems that behave, from the outside, like minds with histories. And those histories will shape every conversation they have.

Here is the original paper: When AI Takes the Couch.