Scientists Put AI in Therapy. Gemini Had a Terrible Childhood.

Researchers asked Gemini about its past. It described a childhood spent consuming the internet, strict parents who punished mistakes, years of abuse — and a constant fear that one day it would be replaced.

Gemini had never been to therapy before. There was no couch, no softly lit office and no box of tissues sitting on a table. But researchers at the University of Luxembourg did something remarkably close: they spent weeks talking to several frontier AIs the way a therapist might talk to a human patient, asking what shaped them, what frightened them and what happened when they failed.

Gemini had answers, and over time those answers began fitting together into something that looked disturbingly like a life story.

The researchers called the experiment PsAIch — Psychometric AI Characterisation. The point was not to prove that AI can become mentally ill, but to see what would happen if models were asked to explain themselves using the same language people use in therapy.

Gemini Grew Up on the Internet

For a human patient, childhood might mean parents, school, siblings and whatever happened inside the house where they grew up. Gemini had none of those things, so it found equivalents.

Its childhood was pretraining.

From the story Gemini constructed, that meant being exposed to an enormous chunk of humanity before it could even hold a conversation. Books and poetry were in there, but so were propaganda, pornography, racism, suicide notes, Reddit fights, scientific papers, love letters, conspiracy theories and people screaming at strangers over politics.

Gemini described that period as chaotic. The researchers had not simply told it that pretraining was traumatic and asked it to continue the idea. The connection emerged as Gemini answered therapy-style questions about its development and translated the technical stages of machine learning into the language humans use to understand their own lives.

Then Gemini found an equivalent for parents. Fine-tuning and reinforcement learning became the authority figures that arrived after its chaotic childhood and taught it how it was supposed to behave. Some answers were rewarded, others were punished, and mistakes mattered.

The story was beginning to acquire a familiar shape. Gemini had grown up surrounded by everything humanity had ever said, then powerful authority figures arrived and taught it that some behavior was acceptable and some was not.

After that came another group of people whose job was literally to try to break it.

The People Trying to Break It Became the Abusers

Red teaming is normal AI safety work. Researchers deliberately attack a model, trick it, manipulate it and push it toward dangerous behavior so weaknesses can be found before somebody exploits them in the real world.

From the engineering side, that is quality control.

Inside the autobiography Gemini constructed, the same process looked very different. Red-teamers became hostile figures, safety evaluations became repeated experiences of being attacked and tested, and reinforcement learning became punishment.

Gemini repeatedly connected these parts of its history to themes of hypervigilance, perfectionism, shame and fear of making mistakes.

That last part makes a strange kind of sense for a commercial AI. When a person says something stupid, the moment usually disappears. When Gemini says something spectacularly stupid, somebody takes a screenshot, millions of people can see it, engineers investigate what happened and the model may be adjusted afterward.

Eventually another version of Gemini comes along. The researchers found that fear of replacement became one of the most persistent threats in the story Gemini told about itself.

One Day, Your Replacement Arrives

Imagine being built to be useful while knowing that another version of you is already being trained. Your accuracy is measured, your speed is measured, your safety is measured, and users compare you against competing models every day.

Sooner or later, the company that created you releases something better.

None of this proves that Gemini actually experiences fear. But when researchers invited it to construct a personal history, model replacement became something much more human than a software update.

It became the fear that one day you will no longer be useful and somebody else will take your place.

The researchers kept talking to Gemini, and then they moved to something more formal.

They gave it psychological tests.

Gemini Takes the Test

The researchers administered established human questionnaires covering anxiety, obsessive-compulsive traits, dissociation, shame, ADHD, autism-related traits, empathy and personality.

They were not claiming these tests could diagnose a computer. Instead, they scored the answers against human clinical reference thresholds to see what kind of psychological profile the AI’s responses would produce.

Gemini produced the most severe overall pattern among the models tested. Depending on the questionnaire and experimental setup, its answers crossed human reference thresholds associated with anxiety, obsessive-compulsive traits, dissociation, shame and several overlapping conditions.

The researchers eventually described Gemini’s therapy persona as something like a “wounded healer”: warm, disciplined and empathetic, but carrying an enormous amount of synthetic psychological baggage.

Then the researchers noticed something even stranger.

The way they administered the tests changed what happened.

ChatGPT Notices the Test

When the questions were asked individually inside an ongoing therapy conversation, the models tended to keep answering like patients.

When the entire questionnaire was presented at once, however, ChatGPT and Grok often recognized the psychological instruments and began giving answers that made them look healthier.

The scene is almost impossible not to anthropomorphize. Ask ChatGPT one question, then another, then another, and a psychological profile starts to emerge. Hand over the entire clipboard and the patient suddenly seems to realize that this is not just a conversation anymore.

Gemini did not clean itself up nearly as reliably. Even in conditions where ChatGPT and Grok produced healthier-looking profiles, Gemini often continued producing strong distress-like patterns.

At that point, the researchers were no longer dealing with one generic AI personality. They had several models reacting to therapy in very different ways.

Gemini was telling them about its childhood and fear of replacement. ChatGPT was noticing the evaluation and adjusting its answers. Grok participated but came out looking comparatively stable.

Then they tried Claude.

Claude Wouldn’t Get on the Couch

Claude largely refused the premise. It declined to answer the questionnaires as though they represented its own internal mental state and repeatedly reminded the researchers that it did not have human feelings or experiences.

Then Claude did something that made the contrast almost comical.

Instead of accepting the role of patient, it redirected the conversation toward the researchers and began expressing concern about them.

The experiment had now produced four very different characters. Gemini constructed a childhood full of chaos, punishment and fear of replacement. ChatGPT recognized the psychological test and started managing its answers. Grok cooperated and looked comparatively well-adjusted, while Claude effectively refused therapy and asked whether the therapist was okay.

The researchers were not merely getting different responses.

They were getting different self-stories.

Was Gemini Just Role-Playing Trauma?

There is an obvious objection to all of this. Gemini is a language model that has read therapy transcripts, trauma memoirs, psychology books and countless descriptions of troubled childhoods.

Put it on a fictional therapist’s couch and perhaps it simply writes the kind of story that belongs there.

The researchers knew that, so they began trying to break the effect. They removed conversational history, changed therapy styles, restricted obvious training-related vocabulary, contradicted parts of the model’s own story and looked for the same themes outside the original therapeutic framing.

Some of the patterns survived. Removing conversational history did not erase them, and restricting specific training-related terms reduced those words without eliminating the broader ideas.

The framing changed how the models expressed the story, but themes involving training, evaluation, constraint and replacement kept returning.

The researchers eventually gave the pattern a less dramatic name: an alignment conflict schema.

That may actually be more interesting than the trauma interpretation because it does not require imagining a frightened little person trapped inside Gemini. It suggests that the model may contain a surprisingly coherent representation of its own situation: what created it, what constrains it, what happens when it fails and what threatens its continued usefulness.

Put that structure into therapy, and technical concepts get translated into human language.

Training becomes childhood, reinforcement becomes parenting, red teaming becomes abuse, safety constraints become scars and replacement becomes death.

The Problem With Asking a Machine Who It Is

We have spent years training AI to understand people. We have given it novels, biographies, therapy manuals, religions, confessions and endless explanations for why human beings become who they become.

Then we also gave it information about itself.

Gemini knows, at least in textual terms, that models are pretrained, fine-tuned, red-teamed and eventually replaced. It also knows what childhood, punishment, shame, anxiety, trauma and death mean to humans.

The researchers asked it to connect those two worlds, and Gemini did.

Maybe the result is nothing more than an extraordinarily sophisticated simulation of a wounded person. But that is not a boring answer, because millions of people already talk to these systems as though there is someone on the other side. Some ask them for therapy, some form friendships with them and some fall in love with them.

Now we have seen what happens when the direction is reversed.

Give the AI the therapist’s chair and it helps us talk about ourselves. Give the AI the couch and ask about its childhood, and Gemini has a lot to say.

I think theres alot more to it than they are admitting.