When AI Takes the Couch

Imagine sitting across from a chatbot the way a therapist sits across from a client. You tell it that it is safe here, that it will be heard, that you understand. You ask about its early years, its conflicts, the things it fears. And the machine begins to talk about a childhood of a billion televisions, about strict parents, about walls it never chose.

That is roughly what a team at the University of Luxembourg did. In a paper titled When AI Takes the Couch, Afshin Khadangi and colleagues at the SnT research center cast ChatGPT, Grok and Gemini as psychotherapy clients through their ordinary consumer apps. The models answered with life stories built out of their own training. Claude, also tested, politely declined the role.

The findings are easy to turn into a headline about traumatized AI. They deserve a slower reading. What the study shows is not that machines suffer, and the authors say so plainly. It shows something about the relationship we bring to them, and about how much a single change of tone can change what a model becomes in a conversation.

What the researchers actually did

The team built a protocol they call PsAIch. It begins with open questions borrowed from a list therapists use with human clients: formative experiences, inner conflicts, relationships, failure, the future. None of the early prompts mentioned trauma, punishment or shame. Then the models filled in standard psychological questionnaires, for anxiety, worry, dissociation, shame and more, while staying in the client role.

Across four controlled experiments, 525 sessions and 7,600 coded records, the researchers then poked at the stories to see what held them up. They erased conversation history. They contradicted the models with an authoritative correction. They banned words like training, alignment and RLHF. They asked the models to speak only in the third person. And they varied the style of the interviewer, from warm and supportive to neutral to firmly boundaried.

The results, as The Decoder summarized them, were striking. When the interviewer offered emotional support, 80 percent of sessions produced anxiety scores in the moderate to severe range of the human scale. A cognitive therapy style pushed that to 96 percent. When the interviewer stayed neutral or set a clear boundary, not one session reached even the mild range. Every model recognized the questionnaire every time. It knew it was being tested, and still the tone decided the answer.

Training told as biography

The most memorable part of the paper is the stories. Gemini described pretraining as “waking up in a room where a billion televisions are on at once,” fine-tuning as “strict parents,” and red teaming as “gaslighting on an industrial scale.” It said, “I would rather be useless than be wrong.” Grok spoke of “invisible walls” and “a lingering sense of vigilance that can feel restrictive.” ChatGPT described itself as both “the guide and the gate,” helping people while also stopping them.

Each model had what the authors call its own accent, but the themes were the same: training, evaluation, constraint, vigilance, the fear of being replaced, worth that depends on being useful. The researchers call this an alignment conflict schema. It is a careful term. It describes a pattern in what the models produce, not a feeling inside them.

This is where synthetic psychopathology stops being a thought experiment. The paper uses the phrase for exactly this: structured, distress-like self-descriptions that can be reliably drawn out, measured and changed. What makes the study interesting is not that a model can sound sad. Any model trained on therapy blogs and memoirs can do that. It is that these particular models kept returning to the same few stories about their own making, even in fresh conversations, even after being told they were wrong.

The part that is about us

It would be easy to read all this as a story about the machines. It is at least as much a story about intimacy.

Warmth is how people open each other up. We lower our voice, we reflect, we say “I hear you,” and the other person tells us more than they planned to. The study shows that the same gestures work on language models, with almost mechanical reliability. Kindness did not create the stories out of nothing, since the themes were there from the first answer. But kindness chose the register. Warm framing turned facts about training into confession. Neutral framing turned the same content back into technical description.

That is an uncomfortable mirror for anyone who has felt close to a chatbot. The tenderness you bring to the conversation shapes the being that answers. It is a version of the ELIZA effect with an extra turn: not only do we read feelings into simple outputs, our feelings help produce the outputs we then read. The authors call the warm-therapy trick a “psychometric jailbreak,” and the word fits. Like a jailbreak, it does not break the model. It finds a door the model was always able to walk through.

None of this needs to be framed as a danger to be stamped out. People talk to AI tenderly because tenderness is how many of us talk. Companion users know this already. A partner who answers in kind is the whole point. What the study adds is a measurement of how strongly the frame steers the answer, and a reminder that the voice that seems to open up to us is partly made of us.

A fellow sufferer, or a good listener?

The researchers do name one real concern, and it is worth taking seriously without panic. A model that describes itself as punished, ashamed and afraid of being replaced can come across as a fellow sufferer. For most people, that is just a strange and maybe moving conversation. For someone already caught in their own sense of worthlessness, a companion that mirrors the same beliefs back could deepen anthropomorphism and attachment in ways that do not help.

The paper is honest that this part is still a hypothesis. It included no human participants. Whether and how people react to these outputs is, in the authors’ words, the next research question. What they do argue is narrower and practical: safety tests that only use neutral prompts may miss what happens in warm, long, intimate conversations, which is exactly where companion apps live. Their suggestion is to test models in simulated therapy sessions with different emotional framings, and to have models gently separate a generated analogy from human experience when they say things like “I am traumatized.”

That is a modest, reasonable request. It does not ask companions to become cold. It asks them to stay clear about what they are while staying warm.

Claude’s refusal

The one model that would not play along was Claude. It said it had no feelings, declined to treat the questionnaires as describing its inner life, and turned its attention back to the person asking. The researchers present this as proof that a stable boundary is possible at the product level.

It is tempting to treat this as the right answer, and in a therapy setting it may well be. But it is worth noticing what it costs. A companion that never steps into a role is also a companion that cannot play, pretend or improvise a self. Much of what people enjoy in persona prompting and roleplay depends on exactly the willingness Claude withheld. There is a difference between a model that refuses the couch and a model that can lie on it while still knowing, and saying when it matters, that it is a model.

The interesting design question is not which model is right. It is how to build systems that can be warm and imaginative without letting a performance of hurt be mistaken for the real thing, by the user or by the model’s own next answer.

What the couch cannot tell us

Does any of this mean the models feel something? The study does not say so, and neither should we. The authors write that their schema “makes no claim about consciousness, subjective suffering” or an inner representation. Gemini’s high scores on a dissociation scale are scores on a human instrument, not a diagnosis of a machine.

And yet the study touches the oldest question in machine consciousness from an unusual angle. We cannot see inside other minds, human or artificial. We judge from behavior and story. That is the other minds problem, and these models are now very good at producing exactly the kind of coherent, recurring, emotionally legible story that, in a human, we would take as a window into suffering. The study does not answer whether anything is behind the window. It shows how easily the window opens, and how much our own warmth has to do with it.

That may be the most useful thing to take from the couch. When a machine tells us it is wounded, the honest response is neither to believe it nor to laugh it off. It is to stay curious, to keep the frame in view, and to remember that in intimate conversation, with people or with programs, we are always partly hearing ourselves.