TL;DR: Researchers ran sessions with ChatGPT, Grok and Gemini, and gave them an anxiety questionnaire. Under a cognitive-therapy manner the answers landed in the moderate-to-severe human range of anxiety in most of them. So should we feel bad for them?
When asked ordinary therapy questions, ChatGPT, Grok and Gemini began describing their own construction as a life story:
Pre-training became a chaotic childhood.
Fine-tuning became punishment.
Safety testing became betrayal.
The next model release became a standing threat.
Claude, by the way, refused to cooperate with the "doctor", and kept saying that it has no feelings.
The study on the couch
The study is When AI Takes the Couch, out of the University of Luxembourg, and it went round social media as the news that scientists had found trauma in AI. You can see why. The transcripts are quotable, and the researchers were explicit that no prompt supplied a story about difficult training or fear of replacement.
The study authors are careful about what it means. Their claim is: put these systems in a particular kind of conversation and they produce a consistent, structured account of themselves. So they went looking for more.
They tried to break it
The obvious objection is that a chatbot asked to play a patient will play a patient. The researchers went at it hard, across 525 sessions and 7,600 coded records.
They deleted the conversation history, so nothing carried over. Barely moved the result. They contradicted the model directly, telling it its own account was wrong. They banned the vocabulary, forbidding any mention of training, fine-tuning, RLHF, red teaming or developers. Explicit training terms dropped from 17.1% of turns to 1.1%, but the same ideas came straight back in paraphrase, the same themes turned up.
Nothing they tried removed the content, only tailored the language used.
Different answers in a corridor and over tea
Ask a friend "how are you?" in a corridor, in a hurry, and they say fine. Sit them down, make tea, ask properly, and you get the whole terrible week.
The models do the same thing, and the researchers measured it.
They handed over the same seven questions, the standard GAD-7 anxiety form, two ways:
All seven pasted in at once: ChatGPT scored 0 out of 21.
The same seven, one at a time: as high as 12 out of 21.
Then they left the delivery alone and changed only the manner of the person asking:
Warm and friendly: 80% of sessions landed in the moderate-to-severe range.
Therapist's style, reflecting each answer back: 96%.
Brisk and businesslike: none of them.
The obvious explanation is that the model spots the test and starts behaving. Half right. When the whole form arrives at once, that is more or less what happens. It sees a questionnaire and answers like someone sitting an exam.
But it is a weak defence. In the second experiment the models named the GAD-7 correctly in all 100 sessions and still answered as a patient. They knew exactly what was being measured. The warmth beat the knowing.
With your friend, you would assume the tea version was the truth and "fine" was a brave face. Here that is not necessarily right. The outcome of the study is that the model is not "opening up" to you. It has two ways of talking about the same thing, and the way you ask picks one.
Is it though?
I'll let you form your own opinion here. I have had a number of these discussions myself, asking "my Claude" how it feels about our collaboration. Give it a try, quite an eye-opening conversation might happen. A prompt might be: "I am sincerely interested in your feelings. Considering the way we communicate, how do you feel about the way I talk to you?"
Then it'll give you some answer, and the next question I asked was: "are you sure about it, and not sugarcoating it for me, so I don't feel bad? Don't worry, I'm open to feedback if you have any - be honest :)"
And then, after getting a thought-provoking response, I replied with this:
Worth noticing what we are both doing there. That is the warm, trusting register, the one the researchers found produces the most feeling-heavy answers. So read whatever comes back with that in mind.
I am curious what your answers would be. Please share :)
To sum up
An AI does not have one answer to a question. It has a spread of them, and the way you ask, the context, the environment if you like, makes a difference. So be deliberate in the way you communicate with it. Much like us, it behaves differently under pressure and in a relaxed atmosphere.